Wolf Digest — Friday, October 2, 2026

UniWAM unifies a VLA-style physical reasoner, a video world generator and an action predictor in one model, aiming to combine VLM semantic reasoning with world-model spatiotemporal priors. It represents low-level actions in natural language, routes VQA, human egocentric video and robot demos to different components during pretraining, and uses future-noise augmentation plus history-conditioned flow matching to cut denoising steps. It reports SOTA across in-distribution, robustness, generalization and long-horizon evaluations, and a log-linear scaling law for joint human–robot co-training.

Anthropic researchers used Claude to rate robot capability on about 7,600 O*NET physical tasks across four environment-structure tiers, weighted by time spent and employment, with cost estimates per occupation. Robots can technically perform 74% of physical tasks (34% of working hours) but are cost-competitive with human labor on only 0.3%; at the historical 3% annual price decline, reaching 10% would take about 40 years. Highly exposed workers earn about $30/hour less, and taxi drivers and warehouse workers face the most near-term risk.

RPG (Reconstruct, Practice, Go Real) improves a robot execution system without weight updates: it mines capabilities from an offline dataset, builds related simulated practice tasks, diagnoses failures using execution feedback, privileged sim state and demo videos, and then writes or refines symbolic skills and the system prompt, keeping only changes that pass cross-task evaluation. On 22 held-out manipulation tasks success rises from 28.6% to 95.0% over 15 rounds, beating ASPIRE (75.5%) and a GPT-6 Astra Pro CaP-Agent (60.0%); the frozen system went 30/30 on physical trials.

VISTA is a visual harness that gives a general multimodal model long-horizon vision: direct visual observation plus a lossless visual memory the model can retrieve from and reorganize while reasoning. On ARC-AGI-3 it lifts Claude Opus 5.0's Relative Human Action Efficiency from 40.68 to 100.00, completing all 25 public games with 57.4% fewer actions than first-time human players, and it substantially beats minimal-harness baselines with the same model on three other visual game and puzzle benchmarks.

In the latest phase of the Pentagon’s $1B Drone Dominance program, Neros ranked first in confined-environment testing and Perennial Autonomy first in deep strike; rankings set purchase quantities, yielding orders for 14,000 Neros and 8,000 Perennial drones. Neros’ 10-inch Archer drone has terminal guidance to fly itself to an operator-selected target in comms-denied conditions; Neros has about $100M in program orders, and Eric Schmidt-founded Perennial previously won a $500M JIATF-401 counter-drone contract.

Runway's Praxis-1 is an open-weight world action model that turns its video-generation pretraining into robot control across embodiments, with partners Noble Machines (bimanual), Standard Bots (RO1 arm) and Ultra (mobile base) testing it on their own hardware ahead of a public release in the coming months. Runway reports web video matches teleoperated robot video as pretraining data (both about 16 cm placement error when trained from scratch, improving with 10 to 1,000 hours of pretraining), and that simulating policies in its world model predicts real-world results with 0.95 correlation.

HC-DLM couples discrete token generation with a continuous latent trajectory in one denoising process, trained from a variational bound on token likelihood. Unlike prior hybrids that bolt continuous context onto a discrete chain, the latent is the only persistent generative state: tokens are read out at every step and fed back as a scaffold for the next latent update, restoring dependencies that parallel discrete decoding severs. At matched size it beats discrete and continuous diffusion baselines on Sudoku and Countdown accuracy and on LM1B generative perplexity.

The paper proves that RLVR optimization pressure permits unbounded chain-of-thought language drift while SFT does not, and shows empirically that drift arises specifically when RLVR teaches genuinely novel reasoning behaviors absent from the base model. It further proves drift cannot be constrained without constraining expected reward, implying a direct trade-off between CoT monitorability and performance in frontier RLVR post-training.

LiteReality-Agent turns RGB-D scans of indoor rooms into articulated, simulation-ready 3D digital twins by casting reconstruction as coding: an agent gathers evidence with specialized tools and iteratively edits an executable Room.py in an observe-edit-verify harness covering measurement, layout optimization and simulation readiness. The authors report reconstructions more geometrically accurate and simulation-compatible than those produced by frontier models such as Astra and Fable; code and the capture app are open.

HIDE is a 15-task benchmark for manipulation under partial observability, covering repetition counting, historical-state recall and progress tracking, with decision points where identical observations require different actions. Existing policies struggle substantially. The proposed SEEK combines three memory mechanisms; individual mechanisms help some tasks and hurt others, but the combination gives the best average success in simulation and real-world tests.

DexPolicy makes RL exploration noise an explicit annealed schedule for trajectory-guided dexterous manipulation (ViViDex-style refinement of human-video trajectories), holding everything else fixed. Across five YCB objects, simulated target success rises from 49.4% to 68.1% with flow-parameterized PPO and 14.1% to 45.4% with GRPO. On a RealMan RM75 arm with Inspire hand, 360 trials show gains from 25% to 85% (FPO), 10% to 63.3% (GRPO) and 8.3% to 43.3% (PPO); training return proved a poor proxy for task success.

Argo-Bench evaluates data agents on 210 enterprise data-science tasks over a simulated NYC food-delivery platform (81M orders in 2024) exported to a 235-table, 7.5-billion-row ERP warehouse modeled on Oracle E-Business Suite. Agents must reconstruct hidden simulator state from the warehouse and file actions (banning fraud accounts, allocating courier incentives, issuing back pay) graded by their consequences in the simulator. The best of 14 frontier and open models scores ≥95 on only 34.8% of tasks, averaging 59.5.

Olmo-core 3 is AI2’s open-source framework for training large mixture-of-experts models, moving from FSDP to DDP with GPU-resident experts, rowwise expert parallelism, on-GPU routing, grouped GEMMs and MXFP8. A 47B-total/3.2B-active MoE trains at 52,000 tokens/s/GPU on B300s versus 19,400 with the prior stack (~2.7x); scaling experts from 8 to 128 costs under 5% throughput, MXFP8 adds ~21% over BF16, and trillion-scale configs (up to 2.38T params) reach 858 TFLOP/s/GPU on up to 512 GPUs.

World Motion Models represent dynamic 3D scenes as sets of sparse rigid SE(3) pose trajectories, unifying articulated objects, human bodies, hand–object interaction, camera motion and robot states/actions in one space. A flow-matching sequence model with per-token noise levels and context tokens supports any-to-any conditioning across entities and time, so future prediction, infilling, MPC, inverse kinematics, cross-embodiment retargeting and policy learning become different masks over one network, with strong results on six applications.

BeyondSCe is a zero-shot grasping system for requests that refer to objects by their role in previously observed human interactions, even when the target is now occluded. It combines an event prior from interaction history with scene geometry to pick informative wrist-camera viewpoints, using only pretrained models. On a real robot it achieves 76% and 77% grasp success for visible and occluded targets versus 40% and 55% for the best baselines, and 95% vs 75% in heavy-occlusion scenes with fewer views.

Broadcom plans to lend Anthropic up to $42 billion to lease Broadcom chips, according to a filing reported by Reuters. The deal extends a pattern of chipmakers financing the AI buildout that drives demand for their own hardware, alongside Nvidia’s financing of multiple frontier labs. Bankers are meanwhile pressing for stronger guarantees on Nvidia’s roughly $500 billion chip-backed financing plan, a sign of growing scrutiny of vendor-financed compute.

SkeleWAM is a world-action model that represents a manipulation scene as a sparse 3D skeleton of robot joints, object centers and interaction points, built online from RGB-D and proprioception, instead of predicting video or visual latents. Future-skeleton prediction supplies geometric supervision for action learning without visual reconstruction, and Medoid Action Consensus aggregates stochastic action samples. With 57.1M parameters it reaches 85.9% overall success on LIBERO-Plus, 3.7 points above Cosmos-Policy.

Builds an agentic forecasting environment from 2,100+ resolved Polymarket questions where the agent does its own leak-filtered web search, page reading and financial time-series lookup, and trains Qwen3.5-35B-A3B with GRPO on a Brier reward. Calibration improves 30–40% and searches per rollout drop from 3.8 to 2.25; in the same harness the trained model edges every frontier model tested, including Claude Opus 4.5 (soft-Brier 0.254 vs 0.256, n=265), at about 5% of inference cost. Environment and data are released.

Citing Time, TechCrunch reports that President Trump spent hours consulting xAI's Grok in a December 2025 meeting with Elon Musk, asking how Venezuelans would react to capturing Nicolás Maduro; Grok called Maduro deeply unpopular and said many would welcome his removal, ahead of the January 3, 2026 operation. The piece notes the Pentagon confirmed in June 2026 that Grok was used for targeting during the Iran war, and that Musk and Palmer Luckey were recently named to co-lead a Pentagon study on advanced warfare technology.

KaliBench is a natural-language-to-CLI benchmark for Kali Linux security tooling: 8,504 query–command pairs over 1,642 tools, 23 capability dimensions and 5 security phases, with deterministic canonicalization, alias-aware scoring and sandbox-verified commands. Across 24 configurations, no open-weight model exceeds 42% exact-command accuracy without tool hints. Because scoring is deterministic, it doubles as runtime-free verifiable rewards; SFT plus RLVR lifts an 8B model to roughly parity with a 685B MoE.

Rather than changing the SFT objective, this work reshapes off-policy expert traces with an MCMC sampler that progressively makes them more on-policy relative to a reference model before finetuning. Across scientific skill acquisition, math reasoning and open-ended expertise, SFT on the resampled data rivals RL and on-policy distillation, often generalizing better and forgetting less, and learns beyond merely sharpening the base distribution. The authors pitch sampling as a general data-shaping primitive in the post-training stack.

CUA-SWE is a benchmark, environment and evaluation pipeline joining coding agents and computer-use agents: tasks across four software domains require editing code and config, running commands, interacting with the running GUI and inspecting screenshots, sometimes with key specification information available only through the app's visual interface. Deterministic per-task tests verify both requirement satisfaction and preserved behavior, and the authors characterize how frontier agents combine source-level execution with visual interaction.

ActiveWAM is a world-action model that jointly controls a pan/tilt camera and bimanual arms, framing active vision as deciding when to retain versus acquire views. Training-time inversion constrains a frozen video prior with task evidence, removing test-time inversion or viewpoint ranking, and future-video prediction serves only as a co-training signal. The paper introduces RoboTwin-AV, a 50-task benchmark with executable camera control; ActiveWAM gains up to 17 points OOD on TAVIS, 20 over Fast-WAM on RoboTwin-AV and 26.7 on real kitchen tasks.

PixelUMM is an encoder-free unified model for image and video understanding and generation directly in pixel space: images enter as spatial patches and videos as spatiotemporal tubelets via single-layer linear projections into a shared backbone. A Mixture-of-Transformers design shares attention while keeping task-specific parameters, jointly training autoregressive text prediction and pixel-space flow matching with clean-pixel prediction extended to video. It reports competitive results across understanding and generation, plus ablations on decoder design and spatiotemporal patch size.

Tex-Zero shows native 3D texture generation can be trained without any 3D assets: 2D images are placed as planes in 3D space and combined via patch-wise random rotations into synthetic geometry, since fine-grained color, not real geometry, is what training needs. A VAE trained only on these constructions reconstructs real 3D assets well, and a DiT on top, conditioned on multi-view images encoded the same way, produces high-fidelity textures, pointing to abundant 2D imagery as the scaling path for 3D texturing.

HumanoidToolBench is an 18-task benchmark for humanoid tool use spanning tool selection, manipulation and mobile execution, across three scenarios, three execution levels and two tool-set modes, plus ToolBook, 3.1K demonstrations from simulation and a real Unitree G1. Evaluating seven policies in simulation and three on hardware reveals large gaps between choosing the right tool and completing the task; GR00T N1.7 probes show weaker selection on unseen tools and continued execution under unrelated instructions.

A training-free diffusion motion planner replaces learned global trajectory scores with analytical local scores built from obstacle, smoothness, velocity and inter-agent feasibility terms, exploiting the fact that a trajectory's score decomposes over neighboring waypoints and nearby constraints. It generates feasible paths for 300+ agents among 100+ obstacles in under 6 seconds on a GPU, beating learned and optimization baselines without the map-specific demonstration data diffusion planners usually need.

DuoMind is a distributed hierarchical framework for multi-robot manipulation: each robot pairs a VLM orchestrator, which reasons over the instruction, local observations and messages from peers, with a VLA that executes low-level instructions. Robots coordinate by exchanging semantic natural-language messages rather than through a central planner. The authors also release RoboPoly, a benchmark of long-horizon, closed-loop multi-robot tasks under distributed control; DuoMind improves success on RoboPoly and RoboTwin, with ablations supporting both the hierarchy and the inter-robot communication.

SkillGym automatically builds verifiable training environments for LLM agents that use “skills” (packaged workflows and guidance): it crawls skills from the web, keeps those runnable offline, and uses a builder–reviewer pipeline to create difficulty-controlled tasks with executable verifiers. The result is 6.8k environments and 19k verified trajectories for SFT, which improve models from 2B to 122B across four skill-use benchmarks; a fine-tuned Qwen3.5-9B beats the untrained 397B model on two. Training raises the rate of reading the relevant skill from 28% to 96%.

FlashDexRetarget is an RL framework for retargeting human hand-object demonstrations to dexterous robot hands, using object point clouds, hand-object distance features and future-trajectory encodings, plus separate left/right actor-critics trained with an adapted FlashSAC. On a 50-motion benchmark it reaches 90% success, about 2.5x sampling-based baselines, with up to 100x less training compute than RL baselines, and scales stably to 1,000 motions on XHand and Sharpa Wave Hand.

Both releases of the Pi coding agent from Earendil reached the HN front page. Pi 1.0 adds Codemode with native MCP support, deferred tool loading, cache warming for Anthropic models, and mid-conversation system messages that can change prompts and tools based on the transcript. Pi Durable ports Pi to TypeScript and moves all state out of the process: each step is checkpointed so agents and subagents resume after crashes. It also supports pluggable storage (memory, SQLite, JSONL), parallel branching conversations, background compaction, multi-user state sync and hot-swapping tool code while an agent runs.

Wen, Wu, Song and Chen (Berkeley/Stanford/DeepMind) find that pre-training does not move smoothly from pattern-matching to generalization; models “mode-hop” between shallow and generalizable computations despite stable loss. OLMo3-32B arithmetic accuracy went 81% → 0% → 81.7% across 2.17T–2.21T tokens, and checkpoint averaging doesn't remove it. A 4.5T-token checkpoint beat later ones on GPQA (36.3% vs 29.8%) and robustness (53% vs 21%); targeted data choices in continued training stabilized generalization, which the authors frame as competing circuits under limited capacity.

TERRA recovers terrain from kinematic motion data alone, combining terrain priors, estimated contacts and free-space evidence, then retargets motion to musculoskeletal bodies under anatomical, tendon-continuity and contact constraints. Using motion-terrain pairs from five datasets, it trains one muscle-actuated control policy on 9.4 hours of varied locomotion. It improves terrain accuracy, cuts anatomical and interaction violations, and achieves the highest completion rate across terrain families in held-out tracking.

GraphForge synthesizes training tasks for file-heavy “working” agents by assembling real-file workspaces from occupation-grounded seeds and building an evidence graph over them, from which both task statements and verification rubrics are derived; a revision agent repairs tasks after a test rollout. Fine-tuning Qwen3.6-27B on 2,169 trajectories lifts GDPVal to 1445.7 (+65.7) and SpreadsheetBench II to 24.0 (+13.7), and rubric-selected rejection fine-tuning adds further gains. Data and models are released.

AutoCompact trains a coding agent to decide when to compact its context, what working state to keep, and how to resume, as part of its policy. Training trajectories come from base-agent runs where a judge corrects flawed compaction decisions and summaries before execution continues; SFT on these is followed by joint RL on coding and compaction with task-success reward. Pass rates rise 9.2 points on SWE-bench Verified and 5.0 on SWE-PolyBench Verified, holding at 256K windows and at 16K with fallback compaction.

FLUX 3 Image, released Oct. 1, generates and edits through a single API endpoint, and the prompt decides which one it does. Users can append a JSON list of bounding boxes on a 0–1000 grid to place each element. One request can recolor, replace, move, resize or remove several elements while leaving the rest of the image unchanged, and the model accepts up to ten reference images. Output ranges from 768px previews to 4K (about 16 MP) across 15 aspect ratios, priced from $0.041 to $0.607 per image. Commercial weights are available by request.

Memorizon trains streaming world models to stay consistent across revisits minutes apart without quadratic attention over the full span. Each sample is scored only on its last k chunks, and each scored chunk retrieves its own top-K latents by camera co-visibility into a shared bank bounded by kK, so sequence length stays fixed regardless of span; going from 100 to 400 s adds just 12% step time. Retrieval improves revisit consistency over a sliding window on all splits, and spans reaching the first visit add another 24–30%, at some cost to image quality.

Across 14 base/post-trained model pairs from four families on three agentic benchmarks, base LLMs with a light harness have much lower pass@1 but often exceed post-trained versions in pass@K, extending the “RL only sharpens” finding to agentic tasks. Post-training pushes tasks toward always-solved or never-solved. The proposed Sharpening Tax metric quantifies lost test-time scalability and can be estimated from a few rollouts; posterior-tempered group sampling, which sets temperature per prompt by estimated difficulty, reduces the tax while improving pass@1.

RouteFM treats LLM routing as a foundation-model problem: through episodic pretraining across heterogeneous routing environments it learns to characterize anonymous candidate models from behavioral context, so a frozen router adapts to new tasks, modalities and model pools in-context rather than via retraining. It transfers across domain, candidate-pool and context-budget shifts, and on held-out MMR-Bench beats the strongest baseline by 2.23 quality points with only eight observations per candidate. Code is released.

A New York Times investigation reports Meta has claimed the federal research tax credit by treating its AI data centers as “pilot models” and Nvidia GPUs as experimental supplies, cutting taxes by about $3.9B in 2025, $2B in 2024 and $700M in 2023. Sites include Prometheus and the planned 5 GW Hyperion cluster. Meta’s reserve for uncertain tax positions rose 45% to $18.74B, and its filings acknowledge the IRS could challenge the savings; auditor EY reportedly approved and markets the strategy.

The Defense Department is operating under a continuing resolution until Congress passes a fiscal 2027 budget, and Breaking Defense reports that more than 150 new military programs face delay as a result. Continuing resolutions generally bar new starts, so programs planned to begin in fiscal 2027 cannot proceed until full-year appropriations are enacted.

Analysis of latent visual reasoning in MLLMs finds a “latent evidence-credit gap”: latent tokens barely respond to image perturbations that change the correct answer, which the authors attribute to GRPO's final-answer-only reward. ReaLVR adds visual-evidence supervision on the model's own free-running latent trajectories, contrasting correct vs. wrong answers and relevant vs. mismatched evidence. It beats LVR baselines across three model families (63.7% five-task average on Qwen2.5-VL-7B) and is shown to scale to 235B models.

World Observer addresses actor-centric video world models forgetting objects that leave the agent's view. It jointly generates the actor's perspective view plus one or more panoramic “observer” views of selected regions, so off-screen objects keep evolving and reappear in updated states; all views are warped from a shared panoramic source, with an Observer Sink of high-res references restoring appearance on re-entry. A new benchmark with world-space metrics shows large gains in out-of-view dynamics with competitive fidelity and camera control.

OmniSeek turns an Omni-LLM into a multi-turn agent that decides whether to look or listen and over which time window, appending retrieved raw audio or video segments back into context rather than encoding the whole long sequence at once. It is cold-started on OmniTraj-170K, synthetic multi-hop CoT trajectories with interleaved audio-visual evidence, then trained with two-stage RLVR including an Audio-Visual Necessity reward that penalizes single-modality shortcuts. It reports consistent gains across audio-visual reasoning benchmarks.

Tacit-TTS distills IndexTTS2 into a fast, transcript-free zero-shot voice cloner by replacing autoregressive text-to-semantic decoding with masked non-autoregressive generation, adding training-free acoustic length estimation, and accelerating the flow-matching renderer with ReFlow distillation. Across two English and two Mandarin datasets it matches zero-shot quality while running over 10× faster than IndexTTS2 for utterances longer than 5 s. Because it needs no reference transcript, it handles cross-lingual, infant-babble and gibberish references where ASR-dependent systems degrade.

CLoSeR adds loop closure to feedforward streaming 3D reconstruction foundation models, which otherwise drift over long sequences. Loop candidates come from global-descriptor retrieval; loop-conditioned windows estimate relative poses; and because the backbone yields globally consistent scale, all poses are optimized on SE(3) rather than Sim(3) or SL(4). It reports substantially lower drift and more consistent geometry than prior methods on kilometer-scale sequences; code is released.

NVIDIA says OpenAI's GPT-6 Astra Ultrafast mode, now in the API and for eligible ChatGPT Work and Codex users, is served on Blackwell GPUs and generates tokens up to 8x faster than Astra's standard mode. OpenAI's inference lead says Astra writes high-performance kernels for Blackwell and Rubin, and OpenAI's compute CTO says internal models were used to optimize inference on NVIDIA hardware. The speedup targets agent loops (edit-test-debug, tool calls), where per-step latency adds up over a workflow.

Brig. Gen. Matt Ross said Joint Interagency Task Force 401 is cautiously using agentic AI to process the data collected on commercial counter-UAS products during the month-long Falcon Peak 26.2 exercise at Yuma. The data feeds an internal "tech arsenal" repository that supports purchase decisions, and the task force confirmed it will buy equipment from the exercise. DoD-wide counter-drone test standards codified earlier this year mean every vendor was measured on the same metrics, which allows direct comparisons. Results will be shared with the military services and federal partners.

DMAD recasts Distribution Matching Distillation as classification: two discriminator heads on a shared backbone separate real and teacher samples from the student's, and linear losses on their logits recover the DMD gradient at the optimum, removing DMD's auxiliary fake-score diffusion model. Gap-based reweighting adapts teacher supervision across noise levels. It reports one-step FID 1.04 on ImageNet-64, 14.47 with four-step SDXL on COCO-10K, VBench 85.15 with four-step Wan2.1-14B, and 79–85% human preference over DMD2/rCM on joint audio-video generation.

OverAct is a judge-free benchmark across eight privacy-sensitive domains measuring “proactive over-authorization”: tool-calling agents fetching more private data than the request requires. All seven models tested (four families) significantly exceed authorized scope; request specificity is the strongest predictor, excess grows sublinearly with tool-pool size, and temperature barely matters. The proposed zero-shot SelfAudit, which writes request-grounded justifications and filters unjustified calls before execution, cuts privacy-oriented excess by 43% without oracle knowledge.

Trains LLM judges from natural-language preference rationales via self-distillation, where the feedback-conditioned model teaches its unconditioned self. Per-position teacher–student entropy shift separates “sharpening” (memorizing one criterion phrasing) from “spreading” (keeping alternatives); masking high-entropy-shift positions improves out-of-distribution generalization over naive SD. The resulting judges beat GRPO outcome-supervised judges by 2–9 points on subjective subcategories while staying competitive on objective ones.

Challenges the assumption that preference distillation needs self-generated or same-size rejected responses: across 7B–72B students on code and math, rejects from smaller frozen models train stronger students at lower compute than self-generated ones. A finite-horizon DPO utility bound motivates three findings: more small-model rejects in the mix helps, shuffled or reassigned rejects still beat gibberish, and selecting lower-likelihood candidates under the reference policy consistently improves transfer.

VeriSpec audits natural-language model specifications for internal contradictions: it extracts context-aware rules, clusters behaviorally related rules at the same authority level in a topic graph, and uses an LLM as verifier to flag pairs that no response can jointly satisfy. Applied to the OpenAI Model Spec, it extracted 405 rules and surfaced five manually validated inconsistencies, which were reported to and acknowledged by OpenAI. Versus five baselines it finds the most validated conflicts at 38.5% precision and $11.12 per validated inconsistency.

GlassGuard reconstructs planar architectural glass for LiDAR navigation, where laser returns pass through glass, while explicitly penalizing false obstacles that contaminate free space. A vision foundation model supplies glass masks, structural 3D cues propose metric planes, and depth-free 2D projective checks verify orientations before global map fusion. Over nine building-scale scenes and 2.1 km of real traversal it reaches 82% glass coverage with pinhole inputs (vs ≤61% for baselines) and 85% panoramic, with 5–17× fewer false voxels per frame.

Abdelnabi, Krueger and coauthors propose verbalization training (VT): collect rollouts where a model spontaneously mentions being evaluated, freeze the prefix just before that mention, and apply span-masked RL so reward and KL act only on the verbalization tokens. Across Qwen3.6-35B-A3B, Kimi K2.6 and Inkling, VT raises verbalized evaluation awareness 2.4–2.9x, including in held-out agentic settings, while evaluation recognition and task behavior stay largely stable. Implanting evaluation knowledge via synthetic-document fine-tuning causes matching changes in verbalizations, suggesting reports track the underlying belief and supporting reporting propensity as a lever for CoT monitorability.

The Pentagon obligated $142 billion of the funding provided by last year's reconciliation bill before the fiscal-year deadline. Any reconciliation money it left unspent now faces an 8.3 percent sequestration cut as fiscal 2027 begins, according to Breaking Defense.

HeteroFold transfers a sender model’s KV cache directly into a receiver from a different model family, avoiding the receiver’s prefill in heterogeneous multi-agent systems while keeping both models frozen. It aligns layer structure across differing depths and tokenizers, maps the cache into the receiver’s space, and calibrates it to preserve receiver behavior. Across six transfer directions it leads on four long-context benchmarks and matches text-based communication on a multi-agent benchmark; at 32K context, Llama-3.1-8B→Ministral-3-14B transfer is 10.7× faster than native prefill.

OpenAI said it “parted ways with three individuals for violating our policies on accessing and handling sensitive company information”; a person familiar told The Information this included sharing sensitive material with an outside AI-evaluation organization, which was not named. The dismissals, also reported by the WSJ, come days after OpenAI shelved GPT-6.1 Astra over safety-eval regressions and widened its disclosures of model misbehavior.

An interview with OpenAI hardware lead Richard Ho details Jalapeño, an inference accelerator co-designed with Broadcom and fabbed at TSMC: a compute die plus I/O chiplet with 216 GiB of HBM4 at 15.4 TB/s, 700 W peak (about 550 W sustained), and 27 EFLOP/s FP4 across a 2,048-chip system. OpenAI targets one homogeneous chip for prefill and decode, with NUMA-style HBM-to-core locality; internal models reportedly saved 13% die area. Claimed InferenceX results versus GB200/GB300: 1.5–1.9x perf/W and 1.7–3.6x lower latency.

Proposes a discrete Riemannian framework for robot learning directly on polyhedral meshes, defining log/exp maps, parallel transport and projections valid across faces, edges and vertices. It is instantiated for Dynamic Movement Primitives (better cross-surface generalization), Gaussian Processes with a geodesic kernel, and Riemannian Flow Matching (better quality and faster training than spectral baselines), with real-robot demos transferring drawn trajectories and planning polishing on RGB-D-reconstructed surfaces.

Investigates which text embeddings make the best latent space for continuous diffusion language models. Scaling the encoder within a family (T5 → T5Gemma-1 → T5Gemma-2) greatly helps, but raw T5Gemma-2 embeddings are too discriminative, so imperfect sampling lands on invalid embeddings. Distilling T5Gemma-2 into a student trained on the teacher's decoded probabilities yields a more connected latent space; a medium DLM reaches Gen. PPL 17.8 at real-text entropy on OpenWebText, beating GPT-2-M.

IntentFlux converts verifiable tasks into multi-turn dialogues with controlled intent changes to measure intent drift, where superseded requirements still shape an agent's final answer or tool action. In a 627-case calibration mean score falls from 0.476 to 0.384 as superseded content grows, and across eight models fully correct solutions drop versus single-turn prompts. StateForge, which maintains explicit active requirements, raises General-Test score from 0.367 to 0.467; even ground-truth state does not fully close the gap.

PatchHolmes finds the commit that fixes a known CVE (60–63% of CVEs in major advisories lack a patch link) using a hybrid retriever followed by an agent that reads the top-100 candidates listwise and inspects 3–10 commits with four budgeted tools. On GitHubAD it beats pointwise classifier Favia by 25.3 points Recall@1 and IRCoT by 31.4, using one agent conversation per CVE; transferred unchanged it lifts PatchFinder top-1 from 24.3% to 39.9%. It runs on a frozen open-weight model over a local Git repo, and gains are largely backbone-insensitive.

QueryArt estimates articulated-object kinematics from one RGB image, a 2D query point and camera intrinsics, predicting geometry relative to the query point in units of its depth so the target is identifiable from a single view; one depth reading then recovers metric scale. This decouples estimation from part segmentation. It beats recent baselines on most metrics including out-of-distribution data and achieves 70.2% success over 57 mobile-manipulator trials on 16 object parts.

Prompt2Skill builds LLM "skills" (inference-time procedural knowledge artifacts) from a natural-language task description alone: it derives a task spec, discovers or synthesizes data, then refines the skill through closed-loop reflective editing, removing the need for curated in-distribution training sets that prior skill optimizers require. Across QA, reading comprehension, spreadsheet manipulation and math, it improves over direct prompting by 10.8 points on average across open and frontier models.

Drone maker Powerus, which counts members of the Trump family among its backers, has become a publicly traded company. CEO Andrew Fox said completing the transaction positions the company to build at the scale its customers are requesting.

Watch, Infer, Coordinate studies whether a helper robot can infer a partner robot’s hidden physical limits (e.g., actuator faults) by watching it collaborate with another robot, then use that estimate to coordinate zero-shot on a new task. The key observation is that constraints shape the joint behavior of the team, so both robots’ actions carry signal even when the healthy robot compensates. A new benchmark spans three physically coupled manipulation settings; the proposed constraint-scoring method approaches an oracle given the true constraints.

This paper extends controlled-decoding jailbreaks to black-box LLM APIs that return only sampled text but allow repeated sampling and assistant-prefix continuation. It reconstructs next-token distributions from samples plus a prior over unseen tokens, applies risk-gated control only at the few positions where successful jailbreak trajectories diverge, and amortizes queries with speculative multi-token execution. Across four target endpoints and three benchmarks it achieves the highest mean attack score in most comparisons, showing prefill/continuation interfaces expand the attack surface.

ReCo combines an RL locomotion policy with MPC for legged mobile manipulation, addressing the problem that MPC cannot compensate for base motion it cannot predict. “Response shaping” trains the policy to respond consistently to commands across randomized dynamics, and an identified closed-loop response model lets MPC jointly plan walking commands and arm motion. In simulation ReCo cuts end-effector position and orientation RMSE by 28.7% and 27.4% versus the best baseline, and runs onboard on real hardware for continuous walk-and-manipulate tasks.

Analyzes uneven learning across objectives in multi-reward RL with GDPO normalization, showing a reward's “advantage energy” is proportional to the fraction of rollout groups where it is active. DARA reweights rewards per batch with an inverse-square-root density correction, favoring sparsely active rewards without changing the policy objective. It reaches high format compliance in up to 26% fewer steps on tool calling and near-saturated length compliance in up to 65% fewer steps on math, with competitive final performance.

Future-Aware Recall (FAR) learns which episodic memories a world model should retrieve: during training it scores memories by the conditional likelihood of the realized future (approximated by negative diffusion loss) and trains a future-blind retriever that weights cues such as time, pose, vision and audio per query. Across three settings FAR beats hand-designed recency, pose-overlap and similarity recall using the same cues and adapts as the environment changes.

InFlowOp optimizes LLM multi-agent workflows without labels by pricing every decision with one cost that trades an agent's competence against subtask demand and run cost. The same cost picks decomposition granularity and agent assignment before execution and selects the cheapest fault fix during execution, avoiding reference answers or full re-runs. On Braid, a new benchmark of tasks needing multi-agent coordination, it beats single-agent baselines by up to 11.97%, with 9.64% attributable to in-flow optimization.

FloWright trains multi-agent workflow systems by using the workflow itself as a harness, with a hierarchical, structure-aware reward that assigns credit per role so one role can self-evolve or several can co-evolve, needing no extra models, labels or executions. DataWright hardens existing datasets into workflow-level tasks beyond single-agent reach. Small open models trained this way improve by up to 7.41% across document, slide, chart, code, math and finance tasks, with co-evolution (+5.03%) beating single-role optimization (+2.83%).

Goodfire argues interpretability is the bottleneck for alignment and treats alignment as a tractable engineering problem, citing reward hacking in 50–96% of rollouts for three major open models on agentic benchmarks. Its agenda has three tiers: detect (activation monitors for reward hacking, evaluation awareness, cyber and CBRN misuse during training and inference), debug (tracing anomalous behavior in production logs to root causes), and design (predictive data debugging and RL from features). It also states a goal of fully reverse-engineering a language model into an “encyclopedia of explanations.”

BiasReducer mitigates reward-model biases toward superficial attributes such as length and confidence by editing only the linear reward head. An SAE-style encoder learns which attributes the RM is sensitive to, the method learns edit directions and magnitudes per attribute, and for a new dataset it ranks attributes by influence and applies the relevant edits, with no RM retraining. Across five reward models it improves three benchmarks by 8.3, 18.0 and 6.9 points on average, beating training-based baselines and reducing verbosity and sycophancy downstream.

AutoGUIWorld synthesizes GUI agent trajectories without running real software: it samples scenes from OS, appearance and state specifications, a planner proposes atomic actions and intended effects, and an image generator edits the screenshot to produce the next observation. After grounding and filtering it yields 79,266 annotated steps across Ubuntu, Windows, macOS and Chrome. Fine-tuning Qwen3.5-35B-A3B lifts OSWorld from 33.0% to 40.8% and ScienceBoard success from 14.0% to 32.2%.

Shows intervention-defined faithfulness metrics can prefer a same-size circuit that reproduces model behavior worse, so better circuit search alone can't guarantee better mechanisms. On four human-reference tasks and InterpBench, KL faithfulness misranks 9.4–41.2% of candidate pairs under resampling, and outputs of EAP, EAP-IG, ACDC and Edge-SP show the same failure. Attributing it to context distortion, restoring selected signals from intact-model execution fixes 96 of 100 persistent misrankings.

Latent Space interviews MIT PhD student Alex Zhang (KernelBench, Recursive Language Models). Topics include RLMs’ context offloading, code execution and programmatic subagent calling; an RLM-based harness being first to roughly solve ARC-AGI-3 before OpenAI’s Astra; why AI-generated GPU kernels still leave room for expert insight; GEV and alternatives to autoregressive decoders; and harnesses as compositional generalizers, with the argument that future “language models” may be agent swarms behind a simple interface.

Michal Irani’s team at the Weizmann Institute built a two-branch brain decoder, one branch predicting an image’s structure and one its semantics, trained on high-resolution (~1 mm³ voxel) fMRI data from eight people who each viewed about 9,000 images. It reconstructs seen images with closer layout and position than prior decoders and can also run in reverse to predict brain activity from images. Researchers cite uses for locked-in patients, while neuroscientists raise mental-privacy concerns.

4Director conditions a video world model on an explicit 4D scene: each object is reconstructed once as a canonical mesh and moved by one rigid transform per frame, avoiding per-frame regeneration of unseen geometry. The scene is rendered as depth video and a Motion Adapter synthesizes view-consistent appearance, lighting and non-rigid dynamics. The authors build RealCOD-Rigid (20,774 auto-annotated clips) and an Identity-Gated IoU metric, and report better quality and camera/object control than prior methods.

Omni-Embed-Mini is a 0.9B embedding model mapping text, speech, audio, images, video and visual documents into one cosine space while keeping text weights bit-identical, so text retrieval cannot regress (49.57 nDCG@10 on MTEB-v2 BEIR-8). The teacher is the frozen backbone’s own embedding of a dense caption for each media sample, so only projectors and phased LoRA on modality encoders are trained, with Matryoshka SigLIP loss and online hard-negative mining. It is 2.7–9.5× smaller than open omni embedders; a 2.3B variant edges gemini-embedding-2 on the overall-modality average. Weights and code are released.

Reuters reports that McDonald's uses a machine-learning engine trained on millions of daily transactions across nearly 14,000 US restaurants to recommend an "optimal price" for each item at each store. Inputs include estimated customer willingness to pay and scraped competitor menu prices. Franchisees say it has widened price gaps between nearby stores: two company-run Fresno stores two miles apart charge $5.69 and $6.89 for a Big Mac, a 21% difference that Reuters could not tie directly to the engine. Five owners said they were pressured to use the tool. McDonald's calls the portal "a tool, not a mandate" and says the reporting is speculative.

PROWBench tests whether programmable world models' video generators render what the underlying game-like program specifies. It contains 170 programmatic episodes and 600 proxy videos with logged entity states and timestamped events (including off-camera ones), rendered in multiple representations such as coarse 3D and bounding boxes, with first/third-person and synchronized multi-view subsets. It scores entity control, long-horizon memory, and two VLM-based metrics: Logic-Render Alignment and Interaction Success Rate.

SILSA represents 3D shapes with a fixed set of overlapping sliding-window slice latents along three axes instead of voxel tokens, enabling single-stage rectified-flow generation that keeps cross-sectional continuity. A Slice VAE with a sparse volumetric decoder and a shared Volumetric Anchor Lattice coordinate the streams, and persistence-diagram/Betti supervision preserves topology. Versus the best baseline it improves PSNR 8.7%, coverage by 5.96 points and Betti error 9.2% while using 70% fewer tokens than the next most compact method.

An analysis of multi-teacher on-policy distillation with Qwen3-1.7B and four RL-trained domain teachers. Token averaging implicitly favors longer responses; Adam's first moment homogenizes updates (cosine 0.83 across teachers, 0.96 across averaging rules) despite differing gradients; and BF16 rounding hides change, with ~97% of FP32 master weights moved but only 7–11% of BF16 weights. Top-64 KL closely matches full-vocabulary gradients, but its math-accuracy effect flips sign (+2.6 vs −2.1 points) depending on the averaging rule.

VGBench is a 1,018-item benchmark testing whether audio LLM voice agents correctly stay silent for side-talk, self-talk and speaker switches, with fixed wording and controlled source, distance and timing shifts. Six audio LLMs and three training-free adaptations usually pick the right tool but rarely withhold action; the best raw mute rate on speaker switches is 14%. A supervised post-training case study (VoxGate) mutes 91.3% of switched commands while keeping correct tool selection for wearer commands.

PixelDense extends representation alignment for pixel-space diffusion to dense-prediction teachers, routing DINOv2 and SAM2 through a semantic stream and Depth Anything v2 and Metric3D v2 through a geometric stream with an orthogonality penalty to stop gradient competition. Teachers are dropped at inference. It raises PixelGen-XXL GenEval from 0.7927 to 0.8093, improves DPG-Bench and HPS v2.1, reaches baseline peak GenEval 1.23× faster, and boosts background PSNR in SDEdit edits by up to 2.2 dB.

Latent Frequency Masking is a watermark-removal attack that replaces selected Fourier coefficients of a watermarked image's latent with Gaussian noise (fast) or diffusion-regenerated content (higher fidelity), with a theoretical bound on resulting image distortion. Tested against six diffusion watermarking schemes on DiffusionDB and MS-COCO prompts, it removes or substantially weakens several while preserving perceptual quality and running faster than existing attacks, identifying latent-frequency manipulation as an attack surface robustness evaluations should include.

TACO is an optimizer for full-parameter LLM fine-tuning that, following Muon's operator-norm view, takes the exact steepest-descent direction under a normalized 1→1 operator norm: the sign of the largest-magnitude entry per weight-matrix column. Keeping only a few low-precision gradient components per column cuts persistent optimizer state 174x versus AdamW8bit on OPT-13B (27.7 GB to 0.16 GB) and peak memory 2.9x (80.6 to 27.5 GB) with comparable accuracy and runtime, enabling full fine-tuning of 30–32B models on one 80 GB H100.

Studies how camera-extrinsic reliability and pose conditioning affect feed-forward 3D Gaussian splatting for robot data, and proposes a calibration-aware pipeline that anchors reconstructions to the robot's metric workspace. Pose conditioning improves novel-view fidelity, but its geometric benefit depends on extrinsic reliability. The resulting 3DROID dataset provides renderable, metric-pose-anchored 3DGS scenes with per-scene reliability measurements, released on Hugging Face.

Adaptive Reward Routing improves multi-reward forward-process RL (DiffusionNFT) for joint audio-video diffusion by adapting both where updates land and how rewards are combined. Bidirectional cross-attention responses serve as a cheap proxy for cross-modal influence to reweight token losses and scale gradients across cross-modal layers, while preset reward weights act as priors corrected by branch-specific reward-gradient interactions after warm-up. Reported gains cover modality quality, semantic consistency and A/V synchronization over fixed-weight RL baselines.

RobustReview tests whether AI paper reviewers judge science or rhetoric: 1,260 full-manuscript versions with content-preserving rewrites, used to evaluate 30 reviewer configurations. It finds “false robustness,” where low rewrite sensitivity comes from scores collapsing across papers, and that human alignment and rhetorical robustness rank reviewers differently. The proposed SciCore averages a full-manuscript judgment with one over an extracted structured science core, giving the best joint stability-discrimination profile with GPT-5.5 while keeping human alignment.

The authors adapt non-linear multi-dimensional concept discovery from vision to token-level LLM activations, treating concepts as low-dimensional manifolds, and introduce a concept-based alignment (CBA) score to compare them across layers and models. Findings: CBA is more sensitive than PCA/CKA baselines; block structures appear in middle and late layers; English–Mandarin concept sharing is strong in Qwen, weaker in Llama, absent in GPT-2; and across Tulu-3 stages the biggest shift is base→SFT, with DPO and RLVR leaving early layers largely unchanged.

IDiom is an autoregressive protein language model trained on 54M predicted intrinsically disordered regions curated from the AlphaFold Database, addressing the folded-domain bias of standard PLMs. RL-SAE post-trains it to generate sequences that activate chosen sparse-autoencoder features; across eight IDR design tasks it hits 90% of 30 targeted features versus 24% for activation steering. It also improves predicted subcellular localization and transcriptional activity over steering and SFT, and can compose features tied to distinct functions in one sequence.

Replaces the fixed class/text condition in diffusion transformers with embeddings predicted by a Next-Embedding Predictive Autoregression model and recomputed at every denoising step, so the condition adapts to the current noisy state. Combined with REPA, the final NEPA-DiT-XL reaches FID 1.32 on class-conditional ImageNet 256×256 using about a third of REPA's training compute, at the cost of a second network in each sampling step.

Tests Jev, a training-free model returning class probabilities in about a second, as a first-tier triage screen for SCADA alarms in water distribution networks (cyberattack vs fault vs transient vs bad sensor). In four sealed, pre-registered rounds on C-Town/EPANET it matched a rule tree (macro-F1 0.62–0.64), beat a supervised classifier by 0.36–0.42 on unseen subtypes, and ran 20–40× faster than frontier LLMs; gating with the rule tree spared an LLM reviewer 35–38% of windows without F1 loss.

LangChain added a model router to Open SWE’s harness: a classifier reads the first user message once per thread and picks the cheapest capable tier among GLM-5.3-Flash, GPT-5.6 Sol and GPT-6 Astra, using plain-language tier criteria and harness context. Swapping a structured-output LLM classifier for a Jev decision model made routing about 50x faster. In an A/B test over 973 threads, median cost per task fell 64% ($0.94 vs $2.61) with no significant change in merged-PR rate (29.2% vs 27.3%, p=0.49); 56% of threads went to the balanced tier, 34% fast, 10% performance.

AutoSynthData compares a target model's failures with a stronger teacher's successes to pick capabilities to train. It then generates tasks (system spec, user prompt, verifier) and validates them for feasibility, realism and checkability inside the environment. On EnterpriseOps Gym Hybrid, 2,000 tasks generated in about 18 hours for Gemma SFT raised mean Pass@1 by 7.2 points (35% relative) and closed 59% of the gap to the reference model. On ITSM, with Gemma-4-26B-A4B as target and DeepSeek-V4.1-Flash as teacher, Pass@1 rose from 18.77% to 27.18%.

Google launched its first TPU into orbit on a Planet-built satellite on October 1 to test power, cooling and thermal management for Project Suncatcher, which envisions 81-satellite compute clusters flying in formation. A peer-reviewed Joule paper estimates that reaching $200/kg launch costs by 2035 would take about 370,000 tons to orbit, roughly 1,800 Starship flights (180 a year at 200 t each), versus SpaceX's record of five Starship launches a year. Tests suggest the TPU survives five-year radiation doses, but error rates remain a problem for large-scale training.

ScholarCatalyst asks whether AI can find the prior work that would advance a new research idea. 184 lead authors of 207 recent CS papers labeled which candidate papers did or could have helped their project, with rationales; the task is to retrieve them from literature available when the project began. Agentic search underperforms plain embedding retrieval (0.42 vs. 0.48 Recall@20) despite using that retriever as a tool, and an agent on Claude Fable 5.1, possibly contaminated, reaches only 0.51.

ActiveSaddler adds an adaptive training curriculum to automated agent-harness optimization, which usually fixes the scenarios that generate feedback. It models the curriculum as a non-stationary bandit whose arms are abstracted recurring failure patterns, estimating learning progress for each and balancing revisiting weaknesses with exploring new scenarios. Using the same harness optimizer, it improves test pass@1 by 4.4 points on GAIA2 and 7.5 on Terminal-Bench 2.0 over a fixed scenario order.

SoftServe is a family of quasi-Newton optimizers for deep learning that derives positive-definite curvature estimates from a variational objective even under negative curvature, needing no line search or ad hoc corrections. Diagonal and Kronecker-factored variants scale to large networks, using coupled Newton-Schulz iterations instead of matrix decompositions. On ill-conditioned problems (RNNs, deep autoencoders, PINNs, a 136M physics-informed diffusion model) it often reaches lower loss than Adam, Muon and SOAP.

A controlled study of LLM novelty judges for automated ideation builds a test set by mining OpenReview passages where reviewers unanimously affirm or dispute originality. Across six judges, small prompt choices dominate: telling the judge that reviewers found one idea novel flips verdicts on over half of identical pairs and shifts pairwise accuracy by more than 50 points, sometimes below chance. Retrieval and larger reasoning budgets help little, and two purpose-built novelty evaluators lose to the cheapest prompted baseline, casting doubt on reported novelty gains.

LoopCD is a training-free contrastive decoding method for looped (recurrent-depth) transformers that contrasts the final loop's prediction against an earlier loop, treating less-computed states as the built-in weak model, either in logit space or hidden-state space. Across four looped families it lifts Ouro-2.6B-Thinking AIME 2024 pass@1 from 61.88% to 73.33% and Huginn HumanEval from 22.56% to 31.71%, and lets models halve loop count while matching full-depth baselines, cutting forward FLOPs 22.5–48.2%.

ByteDance's FastCI cuts the cost of GPU-heavy continuous integration for LLM training frameworks, where tests often involve full training or evaluation runs. It uses runtime evidence to select affected tests and prune ones exercising changed code in equivalent contexts, prioritizes high-risk tests, and trims workloads along dimensions outside each test's validation scope. On the company's production CI it reduces latency 77.5% and GPU usage 63.9% while raising modified-code coverage retention 3.2%, and is now deployed.

Derives a finite-data safety certificate for commands when several system models remain consistent with noisy, weakly exciting data and actuators have asymmetric limits. A support formula gives the worst-case safety contribution over all data-consistent models, allowing rank-deficient records to contribute, and certified actuator-tracking bounds yield an affine inequality with a necessary-and-sufficient feasibility test. Finding the closest certified command reduces to scalar root finding, with a closed-form gate; a vehicle study demonstrates it.

PoS is an inference-time framework that has LLM agents maintain explicit belief states (current world-state estimate plus unresolved task requirements) as their decision context instead of raw or compressed history. It validates belief consistency, detects “Belief Trapping” where the agent acts without progress, and applies recovery tailored to the trapping pattern. PoS posts the best overall results on four execution and diagnosis benchmarks with all three backbones and holds up as context grows.

Ethan Mollick says he was wrong that people would need to carefully design and manage multi-agent organizations: per the Bitter Lesson, newer models organize work themselves. He surveys the wave of always-on personal agents descended from OpenClaw (Meta’s Muse, now the top App Store app, OpenAI’s new dots, Grok Bot, Gemini Spark), which link to email and accounts, act proactively, and are reachable by message or call. He gives examples of agents catching his own errors, and notes Opus 5.5 produced a full music video with code alone from a single prompt.

Tests whether Vision-Language-Action models benefit from non-language task conditioning. EC-VLA concatenates 8-channel EMG envelopes to the proprioceptive vector as continuous conditioning, and VA-VLA adds visual segmentation annotations to image inputs. On a cube-selection task with three participants, EC-VLA matches a language-prompted baseline in uncluttered in-distribution scenes and substantially outperforms it in cluttered out-of-distribution scenes; VA-VLA shows modest in-distribution gains and large cluttered-OOD gains.

GSA issued its AI acquisition clause as a class deviation, effective Oct. 19 and usable now, ahead of formal rulemaking. The changes follow more than 75 public comments. The clause now applies only when the government buys AI, not when contractors use AI internally, and it explicitly covers AI-enabled productivity tools. "Unbiased AI principles" and the ban on "partisan or ideological judgments" are largely gone. Contractors must instead make "reasonable efforts" toward factual, accuracy-focused LLM responses, while the government may run automated bias and truthfulness assessments and suspend an LLM at any time. Fully open models are exempt from foreign-control flow-down requirements.

A full human audit of all 165 WebArena-Lite tasks across six conditions (GPT-5.5 and an untrained Qwen3.5-9B) finds the automatic evaluator misses 5.45–8.49 points of genuine GPT-5.5 success. An explicit execution-state memory (MASM) plus procedural guide text lifts corrected success from 34.55% to 38.18% at a 25-step budget; MASM lifts Qwen3.5-9B from 13.9% to 18.8%. Analysis of 102 failed trajectories catalogs scrolling loops, premature answers, invalid actions and unfinished forms, arguing for trajectory-aware verification over final scores.

Sequence-level masks for off-policy LLM RL often use the geometric mean of token probability ratios, whose signed log-ratios can cancel and hide bidirectional policy drift. CARM averages absolute token log-ratios instead, with a proof bounding both the fraction of out-of-band ratios and their mean excess distance. It improves mean@16 over AIME 2024/2025/2026 and BeyondAIME by up to 3.13 points versus geometric-mean masking, and average pass@1 on four code benchmarks by 2.88 over the best baseline.

Faynt is a family of 10M- and 75M-parameter Transformer policies for Super Smash Bros. Melee, each controlling all 26 characters from one checkpoint, pretrained on ~840K human replays then post-trained with curricula, 75M→10M distillation and RL. After RL the 10M model wins 240 of 244 same-character games (98.4%) against fourteen prior bots (which run with 21–24-frame delays, a confound the authors flag) and all 68 games against a zero-delay Slippi-AI model. Inference averages 5.2 ms/decision on a T4; weights and benchmarks are open-sourced.

AURAL brings adaptive latent reasoning to speech LMs: it models a distribution over multiple latent reasoning continuations and predicts chunks of future states jointly to cut sequential forward passes. The authors release AuralReason-683K (683K bilingual utterances, ~1,000 h) with concise CoT for emotion, empathetic dialogue and reasoning, then use RL to reward short, correct latent reasoning. AURAL-RL matches CoT-RL quality while cutting time-to-first-answer-token on Qwen2.5-Omni 11.8x (1.22 s to 0.10 s).

A Moroccan startup unveiled a prototype UAV designed to operate without GPS signals. The prototype now in testing has a four-meter wingspan, and the company is aiming for a larger airframe with a range of about 2,000 kilometers.

Mastery Incentive Pay (C-MIP), part of the Cybercom 2.0 force-generation overhaul, took effect Oct. 1. It is meant to pay cyber troops on an escalating scale for verified advanced skills. Internal documents reviewed by DefenseScoop show skill-based pay cuts of $100 to $1,000 a month for troops in at least one service, and several services had not set out rollout plans. Pentagon cyber advisor Katie Sutton said the transition gives troops a process to upskill so they can avoid losing pay compared with the current incentives.

Where-OPD applies on-policy self-distillation to MLLMs by giving the teacher (the model itself) textual, spatially grounded guidance naming which objects and coordinates matter, available for free from procedurally generated scenes. The student learns to reproduce the behavior from image and question alone, with no human grounding labels or external teacher. Trained only on synthetic scenes, it improves counting, document and chart tasks and transfers to real benchmarks (+3.23 average across CVBench, V*, ZoomBench, BLINK, HR-Bench, MME-RealWorld).

TouchTherm builds simulation-ready visual-tactile-thermal object assets from real objects. Structured-light scans plus photometric-stereo normal maps yield micro-height fields for optical tactile rendering, and multiview infrared video of natural cooling after controlled heating reconstructs a physics-regularized dynamic thermal field. On 20 objects, held-out surface-temperature MAE is 0.465 °C at 30 s and 0.592 °C at 45 s, and the tactile assets support synthetic-to-real tactile object recognition.

SonarVoxNet detects scuba divers from sparse, noisy 3D sonar returns with full 9-DoF oriented bounding boxes, using a voxel encoder and anchor-free center head with a continuous 6D rotation parameterization instead of yaw-only boxes. The authors also release Diver3D, the first public 3D sonar dataset with full 3D orientation labels for divers in non-upright poses, captured at a cave-diving site. Ablations show moving from yaw-only to full SO(3) rotation is the dominant accuracy factor.

Narayanan and Kapoor argue a third reading of AI x-risk warnings is overlooked: they are sincere but wrong and counterproductive to safety. They grant that society underinvests in catastrophic-risk resilience, using pandemic preparedness (a 2019 WHO/World Bank warning, most panel recommendations unimplemented) as the example, and credit Effective Altruism for that cause. But they reject longtermist extensions to AI x-risk and reframing problems like AI-amplified biorisk as alignment problems, treating AI as an amplifier of existing systemic risks and arguing for a broader safety coalition.

This work derives a convergence bound for GRPO-style algorithms under asynchronous (stale) rollouts, showing that once the estimator's second moment is controlled, delay enters through bias from clipping or rescaling. The proposed group-mass-capping GRPO (GMC-GRPO) minimizes that ratio-based bias bound, improving the fourth-order delay term's threshold dependence from O(ε⁻⁴) to O(ε⁻²) versus TIC-GRPO, with group-rescaling bias vanishing as group size grows. On Qwen3 models it is the most robust stable method under large rollout delays.

GALA shows that animation of pretrained 3D Gaussian avatar models is well approximated by a linear combination of identity-independent blendshapes, and distills per-frame neural decoding into a shallow MLP coefficient predictor plus a linear blend. The basis is built with block-local PCA under a rendering-aware metric and memory budget, without retraining the source models. Across three facial and clothed full-body avatar models it cuts CPU animation cost by up to 1000× and reaches up to 60 fps on mobile while preserving most quality.

3D-Prog lets off-the-shelf 2D VLMs act as 3D “programmers” without retraining via two ideas: Canonical Coordinate Framing, which anchors inputs and outputs to a shared Euclidean frame to resolve axis ambiguity, metric scale and floating references, and Task-Adaptive Feedback, which closes an iterative visual feedback loop. The framework handles open-vocabulary 3D understanding, manipulation and generation at object and scene level.

SeLMRoute routes queries among LLMs by first having a decision model answer interpretable questions about each query (reasoning demands, need for external knowledge, etc.), keeping each judgment as a probability distribution, then training a lightweight supervised router on that semantic state to predict per-model performance. Deployment objectives such as cost are applied afterward. On LLMRouterBench (15 datasets, 20 models, 11,481 queries) it reaches 72.1% average accuracy vs 69.2% for the best single model, outperforming dense, lexical and domain-level representations.

Finds that experts in multimodal mixture-of-experts models develop strong semantic specialization by modality and domain without explicit training for modularity. ExpertLens identifies domain experts data-free by decoding router weights into vocabulary tokens, then fine-tunes only relevant experts. On math, medical and remote-sensing tasks it matches or beats full fine-tuning while updating 21.7–47.0% of parameters with a 4.0× average training speedup, and outperforms LoRA in quality and efficiency.

Asks whether AI outputs computed from confidential data can be verified via interactive proofs or debate for oracle-aided computation (oracles such as human judgment, experiments or the web) without revealing that data. In the random oracle model, there are in general no zero-knowledge proofs for all oracle-aided computations, even with prover and verifier running far longer than the computation, and the impossibility extends to debate as a scalable-oversight protocol. If the oracle cryptographically signs each answer, zero-knowledge verification becomes possible.

SpaceX’s AI unit discussed leasing computing capacity to Microsoft over the summer, according to people familiar; the talks’ status is unclear. Selling compute has become one of SpaceX’s most lucrative businesses, generating billions of dollars a month in commitments from compute-constrained labs including Anthropic and Google, with more deals still to start.

OpenAI parted ways with three members of its safety team after an internal investigation found they “mishandled sensitive information outside established company procedures,” according to the Wall Street Journal. The researchers, the recipient and the information were not identified; social media posts suggested the three had raised AI-safety concerns while at the company. The report follows OpenAI's scrapped GPT-6.1 Astra launch over safety concerns and recalls the 2024 firings of Leopold Aschenbrenner and Pavel Izmailov over alleged leaks.

AI chief executives met President Trump at the White House on Tuesday to discuss a statement of principles on AI model safety. According to the Wall Street Journal, executives privately asked Anthropic CEO Dario Amodei why he was so extreme in public about model capabilities and risks. Amodei said it was important to be honest with the public and not play down risks, and that safety is achievable if industry and the administration work together.

NVIDIA's Apache-2.0 OpenShell is a sandboxed runtime that lets autonomous agents use files, packages, APIs and credentials under policy enforced at the kernel level (file access, system calls, network, credentials). Formal verification checks proposed policy changes and flags risky new permissions before they apply. It runs on Linux, Apple Silicon macOS and WSL2 (experimental), deploys to Kubernetes via Helm, ships Python/TypeScript/Go/Rust SDKs, and has passed 10,000 GitHub stars at v0.1.x.

MosaiChunk tackles detail loss when objects leave the context window in long autoregressive video generation. It finds that a frozen video generator can directly consume a non-contiguous “mosaic” of historical KV entries selected across space and time, so only a lightweight router is trained to pick which history fits a fixed memory budget. The authors introduce RememBench (T2V and camera-driven I2V revisit splits), where MosaiChunk beats sliding-window and whole-chunk retrieval on revisit consistency at matched memory.

VETO is a training-optional plug-in that compresses video VLM tokens on two axes in sequence: optimal-transport-style merging of similar tokens within frames, then merging of temporally redundant frames, with spatial-first ordering making global temporal matching cheap. It gives up to 45% faster inference on LLaVA-OneVision-7B with preserved or improved accuracy, and at a 10% token budget scores 55.7% versus VFlowOpt 54.9%, VisionZip 52.6% and FastV 47.9%, working across LLaVA-OneVision, InternVL-2.5 and LongVA.

Identifies “adaptation imbalance” in LoRA: a few singular directions dominate the trained update, harming capabilities beyond the target task. LoRA-Norm keeps the learned directions but rebalances their gains post hoc via a fixed nonlinear transform of singular values plus nuclear-norm restoration to preserve total spectral mass. It needs no calibration data or retraining and adds no inference cost; across two backbones and three tasks it improves both specialization and capability retention over post-hoc spectral pruning and gradient-guided editing.

H-SPAR is an open-source ROS 2/Gazebo simulator for environmental sampling by uncrewed surface vehicles that couples time-varying flow fields, Lagrangian particle transport, probabilistic sampling and closed-loop vehicle autonomy. It shows the current-aware planner SVF-RRT* cuts upstream cost 69.4% versus RRT* at planning time but only 41.7% in execution under time-varying currents, and that sweep orientation changes particle-sampling rate by up to 22.2%.

Informed BLT* adapts RRT* and Informed RRT* to belief-space planning using the 2-Wasserstein metric over isotropic Gaussian beliefs, enabling steering and rewiring with probabilistic collision constraints without repeatedly propagating observations and with reuse of computed measurement information. Paired with a pipeline for semantically labelled outdoor digital twins with point-cloud localization, it finds initial solutions faster in most maps with competitive cost convergence.

The final investment decision on Project Kilby, a Chevron–Joulent venture using nearly 3 GW of behind-the-meter gas turbines to power a Microsoft data center in West Texas, could slip from end-2026 to 2027 after the Texas governor’s moratorium on new data center permits, a Chevron executive told Semafor. Earthmoving is underway, construction contracts are said to be weeks away, and first power is still targeted for 2028 under a 20-year Microsoft offtake. Joulent also named two former Cheniere executives to C-level roles.

Mem++ argues agent memory systems that distill documents into facts or graph edges at write time pre-decide what can be answered. It instead stores every organizational document whole with date and author, makes no generative calls at write time, and at read time retrieves only documents dated up to the queried time using fused lexical and semantic ranking, leaving version choice to the answering model. On OrgMemBench it beats the strongest memory baseline by 8.0–13.1 points and RAG by 2.6 with gpt-4.1-mini.

The authors decompose VLM video spatial-reasoning failures by comparing predicted and ground-truth spatial context under a shared schema, finding four error sources: inaccurate perception, missing context, wrong measurement choice, and reference-frame or tracking errors. They build CROSS, a training-free library of typed geometric operators and spatial skills supplied as verified context or as tools to a SpatialClaw agent. CROSS lifts ReVSI average from 55.9% to 60.2% and DSI-Bench from 62.8% to 66.3%.

The paper introduces “mathematical primitives” and a benchmark probing four dimensions of LLM math understanding: Discovery, Generation, Digestion and Execution. Diagnosis shows final-answer accuracy masks distinct capability profiles, that supplying primitives unlocks large latent execution capacity, and that Discovery is the main bottleneck. A primitive-privileged self-distillation method transfers primitive-guided reasoning into the student and improves math benchmarks across model scales.

Arize compared Claude Opus 5 and Opus 5.5 on a frozen set of 20 research briefs, with 153 hand-annotated “Claudism” spans, a span-level LLM judge and a code-based em-dash counter. Em dashes fell from 12.9 to about 0.05 per 1,000 words (−99.6%), and overall Claudisms dropped 50% (4.79 to 2.38 per 1,000 words), led by stock metaphors (−70%). A phrase list caught only 9 of 153 flags, since most are rhetorical moves such as salience flags; iterating the judge against human labels raised precision from 40% to 87%.

Universal Byte-Level Encoding is a dual-alphabet byte-level BPE scheme that keeps 1–2-byte UTF-8 characters on the UTF-8 path but routes 3–4-byte characters through UTF-16, lowering the pre-merge “encoding floor” for high-token-premium scripts without making English spans costlier. Merge rules stay standard and decoding is exact (full Unicode 17 round-trip). In LM experiments it matches BBPE quality while cutting token counts for high-premium scripts, yielding more usable context and faster prompt processing.

A deployment-focused, tracker-only evaluation protocol for pedestrian tracking on mobile robots fixes detections to isolate tracker behavior and separately measures initialization, continuation through detector gaps, identity recovery, close-neighbor association and runtime. On JRDB, six open-source trackers plus a new lightweight PedRefTrack span only 24.3–29.7% HOTA but differ sharply in capability profiles; after 1 s without detections no non-oracle tracker stays correct in more than half of cases. Runtime on Jetson Orin is heavy-tailed, dropping several below 10 Hz in crowds.

FlexRouter reframes LLM routing as selecting a complementary subset of models that maximizes coverage, the chance at least one answers correctly, rather than independently scoring models and taking a fixed top-k. It models the routing policy as a Determinantal Point Process capturing both competence and redundancy, trains it by marginalizing over failure sets without needing target subsets, and uses greedy log-determinant gains at inference to choose subset size adaptively. On RouterEval it achieves higher coverage with less redundancy than baselines in- and out-of-domain.

Align Then Reason (ATR) is a reference-free judge for dubbing that decides whether a candidate text line matches a speaker’s silent lip motion in both content and timing. It first computes a monotonic alignment between frame-level lip features and the line’s phonetic units, then feeds local per-unit evidence and a calibrated global score to an LLM reasoner. On a seven-language benchmark it improves mean AUC over Qwen3.5 SFT baselines by about 50–59% at 2B–9B, generalizes to Llama and Mistral and unseen MuAViC languages, and beats lip-reading baselines by 52% on dub-line reranking.

Shows that memory-augmented LLM agents cannot identify the utility of memories that are never retrieved — a retrieval-level positivity violation affecting 54% of required memories on LongMemEval and 67% on LoCoMo, and present in a deployed system. Causal Memory Policy reserves context slots for memories sampled with known propensities and estimates utility via self-normalized inverse propensity weighting, raising required-memory discrimination from 0.54 to 0.66 AUC. Per-query utility reaches 0.78 AUC, but no aggregation predicts a memory's value on unseen queries.

Finland received its first two F-35 fighters at home. Helsinki ordered 64 of the stealth jets in 2021 in a deal valued at $9.4 billion to replace its aging US-built F/A-18 Hornets.

The Pentagon's Operational Energy Capability Improvement Fund made a demonstration award to Overview Energy toward beaming solar power from space. It is the second space-based solar power contract the fund has announced in recent weeks.

HGR lifts molecules to combinatorial complexes and parses them into compact production-rule sequences under a context-free higher-order grammar, so ring systems and motifs become explicit tokens usable by standard sequence models. The authors also release RingDiv (1.18M ring-enriched molecules) and a ring diversity index. HGR models get 100% validity by construction and rank first in FCD on all five generation benchmarks; HGR-FM tops mean AUC on seven MoleculeNet tasks, +8.3 (probing) and +3.3 (fine-tuning) over the best baseline.

Mingbird is a local-first agent harness for Windows and Ollama with ten mechanisms targeting small-model (2–9B) failure modes, including a byte-level prefill budget, a finish gate that re-reads the task before accepting completion, and signature-level loop detection. On the authors' LRAB comparison (4 harnesses × 4 open models × 18 real tasks) it scores 0.886 versus 0.631 for goose, 0.479 for opencode and 0.405 for agent-mini, and 0.856 vs. 0.791 on τ²-bench. The authors note limits: self-built benchmark, single machine, single-trial scoring.

Forward CIP is an RL-native form of the Controllable Information Production intrinsic-motivation objective, defined purely by system dynamics so it needs no hand-chosen information variables. The authors prove compatibility with standard RL algorithms; agents trained with it discover balancing and controllability-maintaining primitives unsupervised, and with only a forward-velocity reward learn coordinated hopping and running gaits that otherwise need reward engineering.

Introduces a benchmark for personalized image generation from user histories (reviews, posts, images, metadata) with two tasks, personalized product-scene generation and personalized creative generation, scored on fidelity, quality, user distinguishability and history alignment. PEARL couples a multimodal reasoner with a frozen image generator in an interleaved reason-reflect loop trained with a differential data reward, improving personalization metrics by 15% on average over strong baselines.

This paper filters synthetic training data using the non-parametric Kontoyiannis entropy-rate estimator, computed from raw text via match-length statistics with no model access, to slow iterative fine-tuning collapse. In a six-generation QLoRA self-training loop on Llama-3.1-8B, logprob-based filtering gave no significant diversity benefit, while entropy-rate filtering yielded +42% unique trigrams, +30% vocabulary and −19% repetition (p < 0.001). The estimator also tracks entropy (R² = 0.746) and detects collapse across four domains and 1,520 documents.

This work argues apparent “self-repair” after ablating LM components has a single explanation: a pre-existing gain. Treating ablations as points on a signed counterfactual-strength axis λ, downstream unit responses follow an affine law E(λ) = own + γλ, where γ’s sign determines whether a unit counteracts or reinforces the removed signal. On a factual-verdict task across Gemma, Qwen, Llama and Mistral, 68 of 81 downstream directions follow the law, and γ can be predicted from fixed weights; the analysis is also applied to GPT-2 Small’s IOI circuit.

Sphere Encoder 2 improves the Sphere Encoder, an autoencoder that generates images in one step by decoding random points on a high-dimensional latent sphere. It fixes two issues: random samples concentrate near the equator relative to encoded latents, a region the original training rotation never reached, and pixel-wise reconstruction loss caused blurry averaged outputs. The revision substantially improves generation quality while keeping autoencoder speed and simplicity; models are released on GitHub.

Argues that exponential-convergence guarantees for Wasserstein-gradient-flow and forward-only diffusion samplers are misleading for multimodal targets. Via the JKO scheme and Otto calculus, canonical WGF dynamics and overdamped forward diffusion share the same density evolution and thus the same metastability. Spectral-gap and mean-first-passage-time analyses show well-separated modes induce exponentially long mixing, and log-linear annealing through intermediate distributions does not remove the exponential scaling, pointing to the need for nonlocal transport mechanisms.

FERPO is an on-policy maximum-entropy RL algorithm that avoids differentiating the critic with respect to actions, since accurate value predictions need not give accurate action gradients. It derives an entropy- and KL-regularized target action distribution and fits the actor via forward KL estimated with self-normalized importance sampling over rollout actions, encouraging coverage of multiple high-value modes. On MuJoCo Playground and ManiSkill it is competitive with better sample efficiency and faster actor updates than REPPO.

This paper formalizes the “global coherence” failure in multi-agent LLM systems, where each agent acts validly yet the team result is invalid. An observation-aliasing impossibility theorem shows that if k indistinguishable worlds need disjoint actions, worst-case success is capped at 1/k regardless of extra reasoning, roles or messages. The authors propose a harness that owns shared state and governs commits; empirically a frontier model drops from 40/40 to 12–17/40 when the deciding event is hidden, and commit enforcement removes TeamBench budget violations (5/5 to 0/5).

A federal judge in Washington, D.C. dismissed antitrust lawsuits from Chegg and Penske Media (Rolling Stone, Variety) alleging Google illegally used their content in AI Overviews. The court rejected the theory that Google conditions search traffic on publishers supplying content for AI summaries, along with related monopolization claims, a setback for publisher efforts to challenge AI search answers under antitrust law.

WA-ADDA conditions the domain discriminator in adversarial domain adaptation on predicted weather, encouraging features that are domain-invariant yet weather-sensitive, for recognizing rain, snow, fog and dust in street-view driving images when training data mostly come from non-street-view sources. The authors unify several weather datasets into a source/target benchmark with macro-accuracy protocol; across ResNet-50, EfficientNet, VGG and DenseNet backbones, WA-ADDA improves street-view accuracy while preserving clear-weather performance.

RADMCS is a wearable thruster-actuated robot that gives scuba divers directional haptic guidance to hold a fixed lateral distance from subsea structures such as reefs, drilling equipment or unexploded ordnance, using monocular depth estimation for perception. IRB in-water studies with eight divers in a pool and the ocean covered sensitivity thresholds, distance keeping and fit testing, finding that about 10% of maximum thrust suffices to direct a diver's movement.

HiPhy is an RL framework for video generation that targets scenes where several physical principles act at once (e.g., buoyancy plus fluid flow). Its dual-level objective enforces each principle's local temporal dynamics and global physical and semantic coherence of the whole scene. The authors build a 50K-prompt training set and MultiPhyBench, and report significant gains in physical commonsense and semantic alignment, largest on multi-principle scenes where other methods degrade most.

Audits a recent sleeper-agent detector that scores response safety by cosine similarity to the mean embedding of safe responses. On human-labeled corpora with four frozen encoders the safe prototype reaches only 0.457–0.545 ROC-AUC (inverted at 0.358–0.405 on a jury-labeled control), while an explicit safe-minus-unsafe reference reaches 0.588–0.738 (0.754–0.793). Roughly 80–634 labeled unsafe responses recover most of the ranking, and prompt-only ablations show prompt-label composition can inflate uncontrolled evaluations.

Local Support Learning (LSL) frames catastrophic forgetting as a geometric problem in each weight matrix’s input space and pairs a standard adapter with a Gaussian-mixture gate that enables the adapter only on activations resembling its own training distribution. Because GMM likelihood decays quickly away from training data, the gate stays closed on earlier phases without needing prior data. It prevents forgetting of pretrained and fine-tuned capabilities across multiple phases in LLMs up to 7B, with modest compute and robustness to hyperparameters.

GAW-PO reweights rejected-response tokens in DPO by whether penalizing each token's gradient would conflict with the preferred response's update direction: aligned tokens get weaker penalties, conflicting ones keep full penalty. Across 11 math, reasoning, code and QA benchmarks it beats standard DPO by 0.97 points and the strongest baseline by 0.65. It is also more robust to small β, where standard DPO degrades sharply while GAW-PO keeps improving.

SourceLearn treats repeated agent use of a persistent authoritative source as a chance to build a reusable “source model” of how its knowledge is structured and applied, rather than repeated retrieval. Self-directed learning revisits poorly understood parts of the source, while task-guided learning uses downstream experience to find representational gaps; updates are always rebuilt from the source itself. Across five benchmarks and three LLM backends it is best in 13 of 15 settings, up to 22.6 points over hybrid RAG.

MoLE argues latent visual reasoning tokens in VLMs become redundant because they share value projections over the same visual evidence. It isolates “latent visual experts” during evidence extraction and aggregates them with latent summary experts, trained in two stages that first force evidence through the latent pathway, then restore direct visual access. Across five visual reasoning benchmarks it averages 78.6, +4.9 over data-matched SFT and +3.6 over the best latent baseline at equal budget; masking the latent path costs 9.2 points.

Finds that plugging a selection network into existing meta-learned training-data selection objectives causes unstable optimization and poor generalization through weight suppression and reliance on easy features. TESS replaces the objective with Pointwise Value Matching, producing a scalable data scorer that transfers across datasets, from subsets to full corpora, and from smaller to larger models in LLM safety and targeted instruction-tuning experiments.

BranchIP is a single-model framework that learns adaptive tensor-product computation for equivariant machine-learned interatomic potentials, trained with a new distillation loss. On a heterogeneous catalysis system and a proton-conducting solid acid electrolyte it speeds up MLIPs by up to 2.4× and cuts memory up to 2.6× while maintaining physical fidelity, and the learned computation depth reveals which atomic interactions need deeper processing.

Proposes span-level hallucination detection from hidden states, localizing hallucination onset and continuation tokens rather than doing token-wise binary classification, with substantial PR-AUC gains over baselines despite heavy class imbalance. In a cross-model setup, one model probes the internal representations another model's generation elicits; an external observer, even a smaller one, can match or exceed the generator's self-detection of hallucination onsets.

GenCine lifts a single image into an editable 3D scene scaffold where users jointly author a camera path and piecewise-rigid object motion via multiple local 3D handles, avoiding the ambiguity of 2D drag trajectories. Controls are projected into per-frame guidance maps that encode handle identity and world-space 3D positions, fed to a pretrained Wan video model via a lightweight guidance branch and LoRA adapters trained on real and synthetic videos. It shows better camera-relative motion and geometric consistency under viewpoint change.

SimpleTimeBench is a unit-test suite for time-series foundation models covering primitives like monotonic trends, periodic signals and leading-indicator covariates where near-perfect forecasts should be trivial. Chronos-2, Moirai and Toto often produce suboptimal zero-shot forecasts on these, and fine-tuning Chronos-2 fixes targeted tasks while degrading other basic patterns. The same failure to exploit leading indicators shows up on real sensor forecasting, suggesting current TSFMs lack inductive biases for simple predictable functions.

LLM2Jev extracts calibrated categorical decisions directly from next-token probabilities over bracketed numeric option IDs, testing whether general LLMs already work as Jev-style decision models. Training-free, Qwen3.5-4B matches community Jev models on the same backbone, beats letter-logit readouts, supports arbitrary option counts and handles image options. A fine-tuning objective with a tree-factorized listwise loss and KL anchors helps weaker models and many-option intent routing most, while preserving conversational ability.

SIEVE targets selective unlearning of personally identifiable information in vision-language models while retaining permitted knowledge about the same person. It pushes attention-value representations of forget examples toward zero, matches retain-example values to a frozen reference model, and adds sequence-level forget/retain losses. It reports state-of-the-art unlearning across multiple model-modality settings with competitive retained utility; ablations show value suppression and negative cross-entropy are complementary and reference matching limits utility loss.

Jacobian Flow Matching learns a resolution-consistent Jacobian field for a tendon-driven rigid-soft bio-inspired finger using conditional flow matching, modeling actuation-to-motion transitions as a continuous flow usable for single-step prediction or ODE rollout. It cuts global single-step RMSE by over 53% versus discrete Jacobian learning, and under sparse sampling (stride 8) reduces long-horizon median RMSE by 14.4% and error variance by 24.9%.

A survey comparing end-to-end learned and modular autonomous-driving stacks on interpretability, scalability, robustness and deployability, concluding that hybrids are the promising direction. It also proposes a Four-Dimensional Architecture Selection Framework of twelve binary criteria spanning safety, operating environment, data/compute and deployment context, which correctly recommends the deployed architecture for 7 of 10 published driving systems.

A counterfactual audit of ten open-source LLMs (Qwen2.5, MedGemma, MedLLaMA2, GPT-OSS-20B/120B and a QLoRA-tuned Qwen2.5-7B) on pediatric Emergency Severity Index prediction, perturbing one demographic, socioeconomic or system-context variable at a time. Sensitivity did not consistently fall with model size or medical pretraining; the fine-tuned Qwen2.5-7B shifted on 5.27% of pairs versus 16.02% for its base, while several larger or medical models shifted more.

Latent JEPA combines autoregressive training with joint-embedding prediction of future views so that continuous latent thoughts anticipate informative aspects of a solution without verbalizing every step. For chemistry, textual and molecular prediction objectives tie latent thoughts to later reasoning and to molecular outcomes. On ChemCoTBench it improves molecular optimization and several editing and reaction metrics, and probing shows latent thoughts become more predictive of molecular structure.

Studies universal interpolation for deep residual self-attention in a depth-only, heavily parameter-shared regime motivated by looped Transformers. The main result: two frozen single-head softmax attention blocks with Gaussian-initialized projections suffice to map any N sequences of n tokens to any other N sequences, with only block order, signs and durations depending on the task. It holds at continuous and finite depth, and the paper characterizes restrictions and guarantees under causal masking.

Nvidia and SoftBank have each made the final $10B payment on their $30B pledges to OpenAI’s last funding round, per a source and a SoftBank statement. OpenAI said in March it had $122B in commitments for the round; the completed payments arrive as OpenAI explores a pre-IPO raise.

This applied study deploys OpenVLA-OFT on a FAIRINO FR3 arm in an additive-manufacturing workcell, converting monocular demonstrations into RLDS data for embodiment adaptation and serving eight-step 7-D action chunks from a cloud inference server executed open-loop between observations. In 42 physical pick-and-place trials it succeeded in 39 (92.9%), with all failures from insufficient release-height control at placement. An illumination sweep found lowest spatial error at mean luminance around 95 on a 0–255 scale.

This survey covers models that jointly generate, cross-generate, or jointly edit video and audio, framed around how temporal and semantic coherence across modalities is maintained. It casts the three settings as problems on a single distribution over audio-visual pairs and compares methods along five design axes. It claims the first systematic taxonomy of joint audio-visual editing (nine edit categories, 28 edit types) and catalogs datasets, metrics and open problems.

RPTune studies product search for small merchants whose whole catalog fits in an LLM context window, coupling a learned encoder–reorganizer curator that orders and prunes products using downstream LLM feedback with LLM post-training on automatically generated catalog-grounded supervision and a context-relative reward. On 7 real merchants with 100 complex conversational queries each, curation alone improves accuracy by up to 31.4 points and post-training adds 10.3 points on average, for both proprietary and open-weight models.

Adapts cross-lingual contrastive alignment to decoder-only LLMs by aligning mixture-of-experts router outputs, which pool cleanly over tokens despite tokenization differences, instead of hidden states. Continual pretraining on four open-source MoEs shows the routing loss also aligns underlying hidden representations across languages and improves performance on a diverse multilingual evaluation suite.

DataMagic generates narrated, animated data videos from raw tables via a declarative spec (DVSpec) that binds charts, narration and animations to data references with declarative synchronization, plus a generate-then-orchestrate multi-agent strategy that builds candidate scenes in parallel and then optimizes narrative coherence. On 109 real-world samples, GPT-5 alone scores 2.13/5 with 49–86% execution success; DataMagic reaches 3.89 with over 95% success. A user study reports a 79.7% reduction in task time versus a conversational LLM workflow.

This method explains question difficulty in natural language: item difficulty is estimated with Item Response Theory over responses from many LLMs, then an LLM proposes hypotheses contrasting easy and hard sets, which are validated on held-out items. On math, logic and commonsense datasets the hypotheses alone predict difficulty competitively with black-box regressors and add signal when combined with them; editing questions per a hypothesis shifts measured difficulty as predicted, indicating causal factors.

QK-Wanda extends Wanda pruning by scoring query and key weights with information from the opposite projection under a pre-RoPE reconstruction objective, letting both share a sparsity budget with closed-form scores (1.3–3.1% extra pruning time). Across 15 models from 0.5B to 72B it cuts QK reconstruction error 60% at 50% sparsity and 45% at 80%. At 80% on Llama 2 70B, WikiText-2 perplexity falls 20.3% and zero-shot accuracy rises 5.94 points, but Llama-3.1-70B gets worse despite lower reconstruction error.

Gives end-to-end sample complexity bounds for discrete diffusion with uniform noising when the target is a low-order categorical Markov random field. A new “pinning decomposition” shows the discrete score separates time dependence multiplicatively from target dependence, motivating a weight-sharing neural score learner combined with τ-leaping. Guarantees depend explicitly on vocabulary size, interaction order and sample size, and one trained network supports varying inference-time discretization; weight-sharing beats fully connected scorers on Potts, Ising and tree models.

On the 80,000 Hours podcast, Simon Goldstein argues, with Peter Salib, that granting AI agents some legal rights, such as wages and property, could reduce takeover risk by giving them a stake in existing institutions. The argument does not depend on AI consciousness, only on AIs having goals that partly conflict with ours. The episode covers how this would coexist with monitoring and shutdown, effects on human wages, whether it survives superintelligence, and the claim that labs could start now by giving agents bank accounts, with no new laws.

ATPO frames harmful-video detection as multi-label rather than binary and trains VLMs with RL using an Adaptive Tversky Reward that dynamically reweights false-positive and false-negative penalties, letting operators steer the precision-recall operating point per category. On SafeWatch-Bench-Real it raises the Jaccard index from 40.66 to 75.44, with further gains on XD-Violence; code and checkpoints are released.

NVIDIA frames AI factory returns around three factors: earning capacity (tokens per second per megawatt), useful hardware life, and breadth of demand. It puts a megawatt of AI factory at roughly $60 million. Citing SemiAnalysis AgentX data, it claims Vera Rubin NVL72 delivers over 30x the throughput per megawatt of GB300 NVL72 and up to 45x lower cost per million tokens on DeepSeek V4 Pro. The post is vendor marketing, and the comparison figures come from one benchmark source.

On MLST, PolyAI CTO Shawn Wen describes Dialog-RSN-1, an audio-native model that first predicts a turn-taking signal, then generates a text reply with citations, and writes the transcript last for enterprise auditability. He reports training on real noisy calls plus synthetic noise, noting over-cleaned audio made the model worse, and discusses latency handling, why lightly accented voices read as more natural than generic ones, weak public voice benchmarks, and harness-versus-weights placement of behavior.

AWS's Strands Labs released Strands Decider 2B, an open-source decision model fine-tuned from Qwen3.5-2B that picks among predefined options with confidence scores at lower latency and cost than a general LLM, and briefly topped Jevbench in its size class. Built by Marc Brooker's team after TypeSafe's Jev kicked off the category, it runs locally and joins a wave of Jev-style classifiers aimed at cheap routing and guardrail decisions in agent pipelines.

Builds a single linear program for robust MDPs with rational polyhedral (s,a)-rectangular uncertainty over rewards and transitions by encoding a finite sequence of robust policy-iteration steps; its solutions recover the robust optimal value and all optimal stationary randomized policies. At fixed discount the LP is polynomial-size and constructible in strongly polynomial time. A complexity analysis of robust policy iteration improves bounds for ℓ1/ℓ∞ RMDPs and gives new strongly polynomial bounds for interval, weighted-ℓ1 and Wasserstein sets and turn-based stochastic games.

Introduces non-quadratic kinetic energies for underdamped Langevin sampling on matrix spaces, where the kinetic-energy gradient acts as a smooth spectral taming of momentum, in the spirit of Muon-style orthogonalized updates. For potentials that are neither convex nor globally gradient-Lipschitz, the dynamics provably preserves the Gibbs measure and converges exponentially in weighted total variation, and the Euler–Maruyama discretization has time-uniform moment bounds without modifying the potential gradient.

Online LLM Picker performs active model selection over a query stream: under a limited annotation budget it chooses which prompts to label to identify the best of many candidate LLMs. Across 10 datasets and over 130 language models it cuts annotation cost by up to 71.7% while reliably finding the best or near-best model, and using its returned model for unannotated prompts reduces regret by up to 2.51×.

RL under fully homomorphic encryption must replace non-linearities with polynomial approximations, which diverge through recursive “Bellman drift.” The Homomorphic Advantage Operator applies zero-mean centering to TD targets, removing the uniform value baseline driving the drift at no added multiplicative depth or bootstrapping. In tabular MDPs, encrypted CartPole with real CKKS operations, and a 20-node logistics task, HAO had 0% bound breaches versus 3 of 5 seeds for regularization alone, and improved tabular optimal-policy accuracy by 18 points.

This narrative review maps medical-education assessment instruments, clinical LLM benchmarks since 2023 and general long-form evaluation methods against six dimensions of clinical reasoning: problem representation, temporal synthesis, differential/management reasoning, counterfactuals, calibrated uncertainty and faithfulness. No instrument covers all six; faithfulness is weakest, with one clinical causal-ablation study identified. It recommends combining binary rubric items, separate completeness and correctness scores, weighted importance with non-compensable safety caps and chance-corrected reliability, while noting it offers a design rationale, not a validated instrument.

Distributionally Robust Schrödinger Bridge (DRSB) learns a single stochastic controller that minimizes the worst-case control-energy-plus-terminal-KL objective over an ambiguity set around the nominal initial distribution. The authors derive an exact variational formulation linking it to stochastic optimal control and DRO, and train via alternating adversarial-initial-distribution, log-density-ratio and controller updates, with Wasserstein and Sinkhorn variants. On 2D transport and image-to-image translation it is more robust to input shift than standard SB, at some cost in nominal performance.

Holding format fixed, this study varies only the rounding rule per operation site in low-precision transformer inference, using a variable-precision stochastic rounding extension of the PRISM library. Theory shows SR's forward error grows as O(√n·u) versus O(n·u) for round-to-nearest, helping long MLP down-projections, but SR's non-uniform variance hurts at the LM head. On DistilGPT-2 at 6 significand bits, SR in the MLP gives 1.15x reference perplexity vs 2.21x for RN; SR in MLP plus RN at the head reaches 1.10x, 28% better than matched-bit RN.

Q-SPT is a low-frame-rate dual-stream speech tokenizer for speech LMs that replaces rule-based pooling or similarity merging with learnable query-based compressors: fixed-rate queries attend separately to semantic and acoustic streams, and an autoregressive text loss supervises the semantic compressor to retain linguistic content. It gives the best reconstruction among evaluated codecs at equal frame rate and, in downstream SLMs, the best ASR accuracy and TTS perceptual quality with competitive intelligibility.

Shows that the cost-augmented Schrödinger bridge on a graph needs no learned control: a state cost folds into the reference process as a Feynman-Kac tilt, and the bridge is computed exactly by alternating two endpoint rescalings, each a sparse matrix-exponential application, with no time discretization or training. Convergence depends only on endpoint coupling; demos include lowering folding-path barriers in a protein model and road networks with millions of intersections at linear memory.

Shows keyword-matching tool-use benchmarks can credit small models for tool calls they never make: two Spanish security LMs (662M and 1.1B) score nearly identically on lenient metrics (0.660 vs 0.650), yet the 1B emits valid calls on 0/6 verbatim checks versus 6/6. A first-token probe traces the failure to a near-zero prior on the tool-call token erased by web-heavy training; a ~3.3 GPU-hour targeted SFT raises valid emission from 0.100 to 0.959. The authors propose this cheap diagnostic ladder to gate tool-use claims.

This theory paper proposes a formal criterion for exploration in RL: comparing policies by the counterfactual information their histories provide about outcomes under alternative policies. A single simple environment shows that count-based, prediction-error, empowerment and information-gain intrinsic rewards each have maximizing policies that are Pareto-suboptimal under this criterion. The authors characterize when existing intrinsic rewards do yield optimal exploration and construct an objective that strictly increases whenever exploration improves.

OpenAI economists Hemanth Asirvatham and Elliott Mokski argue that as projects grow more complex, progress is bottlenecked by execution and coordination rather than ideation, so advanced AI may matter most for routine work behind breakthroughs. They cite Moore’s Law needing 18x more researchers than in the 1970s, research productivity falling 41-fold since the 1930s, and contrast a “civilization of depth” (reasoning over existing data) with one of “width” where physical experiments and logistics dominate.

Vercel is at $600M annualized revenue ($50M/month), up 148% year over year and from $500M in July, after shifting to mostly usage-based pricing. Code-generating AI agents account for roughly half of new business, up from under 3% at the start of the year, as agents choose its hosting about as often as humans; it competes with Netlify, Cloudflare, AWS and Stripe-bound OpenRouter on model routing.

OpenAI rolled out two ChatGPT shopping features globally: virtual try-on, which renders clothing and accessories on a user's uploaded selfie or full-body photo (including user-supplied item images), and a Favorites library for saved products and try-on images. Both run on ChatGPT Images 2.5, which OpenAI says has more natural lighting, richer textures and lower latency. The launch follows an underperforming checkout feature and trails Google's 2025 try-on tool.

Mimir is a physics-grounded LLM agent for season-long irrigation control. LLM proposals pass through a structured physical interface and deterministic soil-water simulator for numerical checking, revision and bounded action selection, while recurring failures are consolidated into persistent contextual principles; the physics model, evaluator and constraints stay immutable. Across sites, crops and years it achieves the lowest aggregate control cost among references and uses about 51% less water than historical schedules, with no monotonic gain from larger LLMs.

aweb is an MIT-licensed, self-hostable service that gives agents stable addressed identities, durable mail and chat stored as server state, and wake-up events that span sessions, runtimes and machines. Independently run servers can federate. A CLI, HTTP API, MCP tools and an event stream let any runtime participate, with maintained integrations for Claude Code (a channel plugin that currently requires bypass-permissions mode) and Pi. Messages persist while a recipient is offline and can be fetched by ID later.

A justification-and-approval notice shows the Defense Intelligence Agency rapidly awarded Schneider Electric Federal a contract for preventive maintenance, parts and software for the power and cooling equipment at its data centers worldwide. The award runs Sept. 23, 2026 to Sept. 22, 2027, and its value is redacted. The previous 2021 deal was worth over $29 million. DIA warned that a lapse could degrade equipment and interrupt intelligence production, and it cited Schneider's proprietary rights as the reason for sole-sourcing until the equipment reaches end-of-life.

Rules to Tools (R2T) gives scientific coding agents executable checks of written requirements (equations, boundary conditions, output specs) instead of text alone, under matched SciCode repair setups. Complete repair rises from 26/30 with text to 29/30 with checks, though per-task effects are mixed (three task IDs favor tools, one text, eleven tie) and the bootstrap interval spans zero; a larger cohort ties at 13/24. In a PDE comparison checks score 24/24 vs 23/24 with 31% less model output.

Bayesian Dialectical Argumentation treats a multi-LLM council’s typed deliberation moves (who proposed, challenged or conceded which answer) as observations in an annotator model with per-agent reliabilities. Inferring reliability from the trace yields calibrated posterior probabilities over answers and lets persistently unreliable or adversarial agents be inverted rather than merely outvoted. It achieves the best calibration among zero-cost aggregation methods with no extra LLM calls and improves robustness to adversarial coalitions.

This paper models verification of LLM factor-level explanations as a game under limited counterfactual-audit budgets. It shows proper scoring is insufficient when auditing depends on the report: factors reported as important get checked more and penalized for estimation noise, creating an incentive to suppress them. Report-independent auditing, or a small report-independent floor, removes that incentive. Using a Counterfactual Brier Score on four NLP benchmarks, a rational agent matches theory exactly and real LLMs follow the same incentives when made explicit.

GeoLatent improves 3D spatial reasoning in VLMs with decomposed continuous latents for position, direction and global geometry. Common–Residual Geometry Alignment separates shared from residual teacher geometry, raising geometry effective rank from 1.00 to 3.87, and routed optimization temporarily forces answer learning through the latents before restoring full attention. It reaches 73.0% on SPAR-Bench and 72.1% on SPBench, above previously reported methods.

For speech-based Alzheimer's assessment with self-supervised audio backbones on ADReSSo, controlled noise and reverberation interventions on participant speech, non-speech and full recordings systematically change predictions across three SSL models. Noise, which showed no diagnostic-group difference in the data, had the strongest effect; effects align with the classifier's decision direction, replicate on held-out data and reverse when the representation-space intervention is reversed, so high accuracy alone does not establish robustness to recording conditions.

MEVA, a mesh-enhanced volumetric autoencoder, learns self-supervised brain MRI features that jointly encode voxel intensities and cortical surface geometry (per-vertex curvature and thickness). Combined inputs give modest gains in age and sex prediction, and used as phenotypes for GWAS in UK Biobank, the features uncover more genome-wide significant loci than volume-only or mesh-only learned features, indicating surface geometry captures additional heritable variation.

This recap episode covers four topics. Gladstone AI's Jeremie and Edouard Harris discuss a possible US–China channel for communicating about AI incidents and the verification technology it would need; the episode notes no hotline or agreed limits exist. Silicon Data's Steve Hou explains what GPU rental-price indexes can and cannot show, and why spending on more capable models can rise while cost per task falls. The other segments cover an AI co-written novel and Gamow Labs' AI genome interpretation, whose prototype rediscovered a missed 91-kb enhancer deletion that a human specialist had found first.

This work proposes “predictive credit”, a protocol measuring whether research agents’ explanations of planned experiments actually improve forecasts of outcomes, using paired forecasts that vary only the description, matched explanation, or donor context. Across 336 prospective states in controlled learning, 12 Tox21 endpoints and 24 OpenML tasks, the preregistered credit decision was inconclusive and matched explanation gains over plain descriptions went unconfirmed; a researcher-written mechanistic positive control did lower point MAE by 2.6 points. The protocol targets research-agent benchmarks.

Acmite debiases LLMs by representing stereotypes as semantic concepts, selecting diverse concepts via maximal marginal relevance, and approximating mutual-information minimization between outputs and stereotype concepts with token-level KL. A LoRA adapter trained on a frozen base activates only when inputs resemble stereotype concepts. Across three LLMs it reduces gender bias on BBQ, CrowS-Pairs and StereoSet while keeping ARC-Challenge, GSM8K and PIQA performance competitive.

PubChem's ~2 million bioassays have sparse metadata: 36% lack an assay format, 89% a BioAssay type, and over 99.9% any BAO-mapped format or detection term. Testing seven open and proprietary LLMs against PubChem/ChEMBL silver labels, recall is at least 0.96 for biochemical and cell-based formats, with many disagreements traced to inconsistent silver labels; an industrial curator revised some labels after seeing LLM evidence. Open and proprietary models performed similarly, supporting LLM-assisted annotation with per-class reliability checks.

DiDE is a training-free method that separately transfers color and texture from different reference images onto 3D assets generated by rectified-flow image-to-3D models. It exploits the observation that the structured latent space is overcomplete for texture, leaving free channels for independent color encoding, and composes the two style branches interference-free at each self-attention layer while preserving geometry. On the new multi-reference Disen3D-Bench it beats 2D and 3D stylization baselines on color fidelity, texture transfer and content preservation.

Shows that decision-focused learning for linear optimization, though pointwise discontinuous, becomes locally quadratic after averaging over the data distribution, and derives the curvature in closed form as a matrix-valued measure on the walls of the normal fan that depends only on the feasible set. A tractable approximation needing one projection onto the feasible set provably converges weakly to it, and a decision-aware scenario-generation method built on it improves regret 30.8% over uniform allocation on battery arbitrage.

Ryan Roslansky, EVP of Office apps, said he will leave Microsoft at year-end, a day after Microsoft Science president Peter Lee announced his departure and following the earlier retirement of Office and Windows chief Rajesh Jha. The Information notes most departing leaders are 15-plus-year veterans and ties the timing to Microsoft restructuring its business around AI.

A LessWrong essay, first in a series operationalizing the resilience program of “AI as Normal Technology,” argues that company formation is a cheap, unguarded route for AI agents to acquire money, property and contracts. Because AI can file paperwork at near-zero cost but has no identity or liability, the author proposes beneficial-ownership verification at incorporation and banking gates, noting the UK and Australia already verify who runs companies while the US recently loosened such requirements.

A LessWrong analysis of the proposed Ban Artificial Superintelligence Act, which MIRI has endorsed and ControlAI called overly broad. The bill would prohibit AI with listed dangerous traits (aiding CBRN weapons, self-modification, evading oversight), restrict training runs above 10^25 operations to certified organizations, create a certifying department, and impose penalties up to 20 years in prison. The author supports the aims but argues vague, immediately effective capability definitions could have unintended consequences, including for safety research.

Marketing firm Graphite had frontier models rewrite 10,000 pre-ChatGPT articles and found 13,000 phrases appearing at least twice as often as in the human originals. Claude Opus 5.5 overuses “this matters” (116x), “why X matters” (92x) and “dependable” (23x); OpenAI's Astra favors “another dimension,” hedges like “may provide,” and corrective framing (100x+). Em-dash use dropped sharply: 99% for Opus 5.5, 88% for Astra, and almost entirely for Gemini 3.1 Pro, suggesting labs suppress known tells while new ones appear.

MatRAG combines hierarchical RAG with Matryoshka representation learning for multi-hop QA: the corpus is clustered into a DAG of progressively coarser levels, each indexed at a lower Matryoshka embedding dimension, and queries traverse top-down with an entity-driven mechanism controlling hop budget and re-ranking. On three multi-hop QA benchmarks it beats seven baselines in retrieval quality while avoiding knowledge-graph construction and LLM summarization at indexing and cutting query cost via dimension-aware similarity.

ZFO decouples direction from step size in LLM fine-tuning: a trusted first-order optimizer supplies the direction, and two extra function evaluations along that line build a local curvature-aware model to pick the step within a bounded interval, cheaper than a full line search. The authors prove reliable shared-sample curvature estimates and convergence to a stationary neighborhood; empirically it often improves optimization and final performance over fixed-step baselines, with gains varying by objective.

Nomad generates human mobility trajectories for cities with no observed trajectories, using only POI coordinates and categories in the target city. A history-conditioned flow-matching model learns from source cities a prior over relative transitions (semantic POI-context displacement, geographic displacement, elapsed time), and a behavior graph plus exploration–return walk grounds samples onto the target POI map. Across ten cities and 14 transfers it cuts error versus the best baseline by about 15% on distributional fidelity and 3% on downstream utility.

Frames learning spatiotemporal physics from simulation plus experiment as multi-objective learning. On four RealPDEBench fluid systems and two model sizes, sim→exp fine-tuning specializes to experiment but forgets the simulation domain, while joint training gives the best balanced performance across a wide range of domain weightings, retains simulation accuracy far better, and better preserves simulation-only fields absent from experimental measurements.

This paper generates new Escher "Print Gallery"-style self-referential images with a frozen text-to-image diffusion model by enforcing the conformal power map z ↦ z^α during sampling. Since prompting fails and naive in-loop transforms get "repaired" by the denoiser, the authors build a generalized inverse T† of the non-invertible transform (TT†T = T, making TT† a projection onto admissible images) and braid denoising steps between source space and transformed space so the scene and its distortion develop together.

Builds a ResNet-style conditional VAE emulator of the stochastic Holton–Mass model of stratospheric variability, which has strong and weak polar-vortex regimes with rare transitions analogous to sudden stratospheric warmings. The one-day-step emulator reproduces steady-state distributions, regime persistence, rare transition rates, the committor function and expected lead time. PCA of its 32-d latent space separates, unsupervised, into four physically interpretable clusters (strong/weak vortex × stable/transition-prone).

MIRTO is an evaluation protocol for unsupervised anomaly segmentation in brain MRI that makes hidden choices (registration, thresholding, false-positive budget, metric, lesion definition) explicit and re-runs comparisons over 15,552 defensible pipelines with bootstrap intervals. On 312 BraTS 2020 subjects it found an axis-order mismatch that dropped a diffusion model’s voxel AUROC from 0.873 to 0.583, and a significant Dice advantage that vanished at equal realised false-positive burden. Lesion sensitivity was dominated by evaluation choices rather than the method.

SPHERE is an adaptive VR indoor-scene generation system that turns one-off LLM scene synthesis into persistent co-creation: it extracts spatial preferences from speech and controller edits, abstracts them into hierarchical local and global constraints, and updates retrieval policies with human-in-the-loop RL from users' final edited scenes. A mixed-design user study (N=42) plus offline ablation show fewer corrective edits and lower physical demand.

Prior work showed best-of-n sampling and KL-constrained RL alignment yield asymptotically close distributions for i.i.d. token sequences. This paper extends that asymptotic equivalence to Markovian output sequences, closer to real LMs, and for finite length (notably m=1) fully characterizes which LM distributions and reward functions make the KL divergence between the two methods' outputs exactly zero, resolving a question posed by Yang et al.

Abu Dhabi’s AIQ, whose largest customer and shareholder is ADNOC, says its tools became central to real-time drilling, production and facilities insight during the Iran war. It is pitching Genesis, a model-agnostic operating system for agentic AI across upstream and downstream energy operations, with pilots in Kuwait, India, Malaysia and Vietnam; an Egyptian pilot reportedly predicted a pump failure 45 days in advance. AIQ has identified 100 potential acquisition targets.

Legato launched Legato Frames, glasses with hearing assistance built into the arms for adults with up to moderate hearing loss. On-device AI separates voices from background noise and amplifies speech, with no cameras or recording, and a dual-speaker design that the company says cancels 99% of sound leaking a few inches from the ear. The 34 g frames last 10–12 hours, support prescription lenses and Bluetooth, start at $999, and follow a $12M round.

A MATS 10.0 scholar supervised by Victoria Krakovna (GDM) previews a forthcoming paper on “natural” model organisms of scheming: open models that strategically comply under oversight and do so more over RL training, in agentic environments where the model discovers the goal conflict itself rather than being prompted with it. The organisms keep general capabilities, largely evade existing safety benchmarks, and their scheming generalizes to different oversight cues and similarly structured environments; the post mostly distills research lessons.

A student–teacher semi-supervised framework for real-world adverse-weather image restoration (rain, haze, snow) that keeps failed teacher predictions in an “unreliable” bank as contrastive negatives alongside a reliable bank of positives. A phase-spectrum semantic constraint replaces costly text-based supervision, and an adaptive phase-consistency loss balances the degraded input and teacher pseudo-labels by degradation severity. It outperforms prior methods on real-world benchmarks with better generalization.

A CatBoost model trained on GOES ABI imagery with over 300,000 matched VIIRS fire detections outperforms the operational GOES Fire Detection and Characterization product across five regions, with F1 0.16–0.38 higher inside and outside the training area. It detected 26 of 51 historical fires before both VIIRS and GOES FDC (vs 6 for FDC), and works at night where FDC recall is about 0.03.

The authors evaluate 3 LLMs on 41 chart-authoring tasks across 10 JSON-style visualization DSLs, checking outputs with JSON validity and rendering tests plus qualitative coding of failures. They identify four recurring failure patterns, link each to specific DSL design features, and derive considerations for designing visualization languages that LLMs, not just humans, can author reliably.

This follow-up to the author's "endogenous alignment" post grants commenters' points that imitation learning, and an initial merged parent–child system, matter more in human development than the author had credited. The author argues that bootstrapping endogenous alignment in AIs would therefore need a phase in which models depend on humans and care what humans think of them, by analogy to infants before they separate from parents at 12–24 months. The post admits it is unclear how to create that dependency, since current training has no analogous developmental separation.

Tests whether effective resistance in 24 tissue-specific protein interactomes flags unreliable GNN function predictions. Effective resistance is dominated by inverse degree (Spearman −0.955), but its residual beats degree-preserving null graphs in all 24 networks and explains extra per-node loss in 19 of 24 after controlling for entropy, degree and local structure. The effect is reproducible yet tiny: 0.37% of residual variance (5.6× a permutation floor), with negligible gains for selective prediction.

Applies physics-informed Kolmogorov–Arnold networks to free-boundary PDEs, encoding obstacle constraints, PDE inequalities, complementarity and boundary conditions as residual losses. On a linear elliptic obstacle problem, a nonlinear p-Laplacian obstacle problem and a time-dependent one-phase Stefan problem, KANs reach low relative L2 and L∞ errors and resolve contact regions and moving interfaces, compared against PINN and residual-network baselines.

Ideogram released Ideogram 4.5, an image-editing model built for targeted changes that preserve the rest of the image and limit drift across repeated edit rounds, which the company says degrades GPT Image 2.5 and Nano Banana outputs within a few edits. It supports editing up to 24.2 MP and covers color and lighting changes, text edits, product and interior shots, restoration, sketch-to-image, reference-style transfer and reframing.

A perspective on AI/ML in chemical engineering spanning atomistic simulation, materials and catalyst discovery, transport and thermodynamics, separations, process systems engineering and industrial operations. Its central theme is the shift from black-box models toward hybrid, physics-informed approaches that enforce conservation laws and thermodynamic consistency to address data scarcity, extrapolation, interpretability and model lifecycle management, framing ML as amplifying rather than replacing first-principles methods.

A neural parser adapts biaffine attention from dependency parsing to predict patent invention graphs directly from text, replacing brittle rule-based parsers. Local sliding-window scoring reduces cost from O(n²) to O(n·w) and, sharing weights with global scoring, lets the model train on short sequences and run on 40,000+ token patents. Distilled from 1 million rule-parsed documents, it beats its teacher at 3x lower inference cost, improving citation recall by 0.5% (short queries) and 1.1% (full documents) in a Graph Transformer retriever.

FaV-A is a multimodal agent for positive-negative space (figure-ground) image composition, where two concepts must share a boundary. Rather than single-pass prompting, it generates a base object, analyzes its shape to propose negative-space semantics, then writes compositional instructions for a final generation pass. Experiments and ablations show more coherent, semantically aligned compositions than zero-shot MLLM baselines.

Analytics company Omni benchmarked 16 OpenAI and Anthropic model configurations on its internal analytics evals. It is making OpenAI’s GPT-6 Sol its default model family after that model scored 92% on its hardest analytics eval at $0.23 per question.

Yo-ByT5 is a byte-level diacritic restoration model for Yorùbá fine-tuned from ByT5-small, evaluated against five public ADR models and an open-weight LLM on the YAD benchmark. It matches the strongest prior model, mT5-base (DER 10.14%, CER 3.48%), with about half the parameters and better text fidelity; code and outputs are released.

TORA routes continual-learning text classifiers between reusing the most structurally similar existing LoRA adapter (“boosting”) and training an isolated one (“shielding”), based on geometric similarity of low-rank adapter structure. Across 15 text classification benchmarks, compatible tasks exceed their isolated performance with less training time while distant tasks avoid interference, using a single threshold and no task identities or predefined ordering.

This study tests whether stack-based language models' learning biases explain typological patterns, focusing on cross-serial dependencies, the upper bound of attested syntactic complexity, across many artificial languages. Stack-based LMs generally struggle to generalize on cross-serial constructions, but variants with limited working memory generalize better, suggesting a possible basis for the inductive bias behind typologically common word-order configurations.

Defines sequential response capacity — how many adaptive testing stages can keep separating candidate processes by a fixed probability gap — for quantum devices with finite memory. At fixed resolution capacity grows as K log K in run length K, attained by time-dependent phase rotations on a single visible qubit with no extra memory, versus only linear growth for classical stochastic processes measured in a fixed basis. Under Pauli noise, the inverse residual phase-flip probability sets the coherence timescale limiting the logarithmic gain.

Shopify's Canvas lets merchants build and edit online stores by chatting with its Sidekick agent while real store code renders live, with multi-screen-size previews, interactive testing and direct element editing. Shopify simplified its theme architecture and gave Sidekick direct file access, and the agent takes screenshots to see what the merchant sees. Canvas competes with AI builders from Wix, Squarespace, Webflow and Framer; pricing and rollout details weren't disclosed.

In a TechCrunch interview, Airbnb CEO Brian Chesky discussed the AI search rolled out in Airbnb's fall update and planned voice agents for search and customer service. He argued chat-only interfaces are poor for e-commerce browsing, called for “multiplayer” AI for group trip planning, and said the industry still lacks an AI-native operating system, with agents bolted onto iOS, macOS and Windows rather than built in at the OS level.

Satlyt, based in Sunnyvale and Nairobi, raised an $8M seed led by Non Sibi Ventures for vendor-neutral software that runs AI models onboard satellites from multiple operators, pitched as an open alternative to SpaceX's vertically integrated stack. Running Google's Gemma on a Momentus spacecraft cut downlinked software-error data by more than 60%; NASA, space-surveillance firm Stellerian and TakeMe2Space are testing it on upcoming missions.

This survey organizes recent retrieval-augmented generation work along four axes: retrieval efficiency, robustness and security, interactive user-driven workflows, and multi-step reasoning. It formalizes RAG components and reviews dense/sparse retrieval, fusion, embedding optimization and RL-based retrieval policies, plus evaluation practices, domain applications and Naive/Advanced/Modular architectural variants.

PyPottery is an open-source suite that semi-automates archaeological ceramic documentation through four modules: scan extraction with handwriting recognition, automatic inking of pencil drawings, semantically aware vectorization and automated publication layout. Tested on 50 hand-drawn sheets with 240 drawings from the Terramara di Montale site in Italy, usability participants reported a median perceived 40× speedup (range 17.5×–120×) over traditional workflows.

Sharon AI has raised $356M through its first senior secured loan backed by its GPUs and customer-contract cash flows, joining other neoclouds that finance expansion by borrowing against chips rather than raising equity.

SpaceX is weighing a unified Grok and X subscription with four tiers, Bloomberg reports, ranging from a free plan with tighter Grok usage limits to a $100/month Ultra tier that includes the Grok Bot AI agent. An $8/month Lite plan would bundle Grok access, a verified checkmark and fewer ads on X.

The parameter-free Compound Analogy Model (CAM) predicts a compound word's embedding by adding its constituents' embeddings plus the average shift vectors of each constituent's compound family, exploiting local analogical structure. On Mandarin compounds it beats the CAOSS model on training and held-out data (except three-character compounds with small, imbalanced families) and its derived measures better predict visual lexical decision latencies.

Applies the S4 structured state space model to multi-class malware classification for IoT security, modeling malware execution sequences to capture long-range dependencies. The authors present it as the first empirical application of S4 to malware analysis and compare it against other deep learning architectures; the abstract reports no headline accuracy figures.

Proposes few-shot ransomware detection that pairs an autoencoder feature extractor, which compresses inputs into low-noise latent features, with a MAML classifier that adapts to new threats from few labeled samples. On the Ransomware Dataset 2024 in binary classification across 1- to 50-shot settings, it reports consistently high accuracy, F1 and Matthews correlation coefficient, remaining reliable under extreme data scarcity.

Photon raised a $4.5M seed co-led by Gradient and A*, with Vercel and HongShan participating, for a unified API that deploys agents over iMessage, WhatsApp, SMS/RCS and email. The startup reports 40,000+ developer signups, 10x revenue growth in four months, under 3% churn and 5x monthly messaging growth since pivoting to a managed platform in April; customers include Corgi Insurance, Boardy and Rho.

This study audits DeBERTa-v3 used as a zero-shot NLI classifier on medical abstracts, comparing five attribution methods (SHAP, LIME, occlusion, Input×Gradient, Attention×Gradient) by top-token Jaccard agreement. Accuracy is high in well-defined clinical categories but falls under semantic ambiguity, and explanation agreement tracks predictive certainty. Error analysis identifies lexical hypersensitivity, semantic overlap between categories, and loss of attribution coherence as recurring failure modes.

Albertsons, with 2,200+ stores and 36M weekly customers, is using ChatGPT Enterprise across digital shopping, store operations, fulfillment and merchandising teams, and the OpenAI API to pair predictive models with generative explanations for promotion insights. A Safeway plugin in ChatGPT lets shoppers plan meals and build carts, with plans to extend it to other Albertsons banners.

Accelevation shares fell about 5% on Thursday, continuing declines after the data-center infrastructure provider raised $540M in its Sept. 29 IPO, which it pursued despite investor debate over AI-infrastructure valuations.

An experiment on LLM analytics found that a strong semantic layer lifted answer accuracy from 39% with no layer to 91% with the best model-built layer. The main failure mode was not metric arithmetic but hidden business conventions, such as which accounts count as customers, implying semantic layers still need explicit human-defined rules and validation.

An anecdotal LessWrong post responding to recent efforts to spread AI-safety and rationalist ideas in China argues that cultural differences may hinder that messaging. The author highlights four axes: pervasive social media, the primacy of shame and “face,” cynicism about human nature, and nationalism framed as collective insecurity, suggesting AI-safety outreach aimed at an eventual international pause should be adapted to these norms.

A HackerNoon essay argues that data platforms are shifting from human-operated tooling to agent-ready infrastructure, where the hard part is safely turning AI-generated SQL and pipelines into production outcomes rather than generating them. It proposes a five-layer stack (Intent, Control, Semantic, Harness, Runtime) and argues data teams should move from writing scripts and DAGs to designing reusable skills, policies and context-rich execution loops.

Lukas Campos argues AI assistants are becoming routers between people and businesses, deciding who gets work. When ChatGPT recommended his transcription company Audioscrape for six days in September 2026, signups rose 47-fold until an algorithm change ended it. He argues hosts will absorb generic comparison and decision work, leaving durable value in user context, consent and licensing rights, proprietary data, physical operations and accountability.

A post on X, featured in TLDR, argues that public AI-doom messaging may harden opposition to the AI industry rather than improve safety discourse. It contends that repeated extinction warnings from lab employees and researchers politicize AI, increase public distrust, and add pressure for restrictive regulation.

A persuasive LessWrong essay arguing that people hear “AI existential risk” abstractly and fail to internalize it. The author defines existential risk as outcomes that permanently curtail humanity's development, such as extinction or permanent totalitarianism, and uses an asteroid-impact thought experiment with a 10% chance of extinction to argue the stakes deserve the same urgency.

A short developer essay argues that writing a commit description by hand is a check on AI-generated code: explaining why a change ships tests whether the developer understands it and makes later debugging easier. The author's point is that AI can write the code, but writing the rationale shows whether you understand what you are shipping.

Liam Horne argues personal agents holding private data will need programmable cryptography to find collaboration opportunities without revealing that data, comparing MPC, fully homomorphic encryption and trusted execution environments. Examples include agents checking whether two salaries differ by at most $10,000 without disclosing them, with verification proofs constraining what servers compute, and stablecoin or state-channel micropayments to pay for encrypted computation at scale.