Wolf Digest — Friday, September 4, 2026

Co-evolution methods that synthesize training environments from on-policy rollout failures stop producing signal once the policy outgrows them. This work evolves environments off-policy instead, deriving three difficulty axes from the multi-turn objective and scheduling generations of increasingly hard environments through a multi-agent harness. Rollout checks with Hy4 preview, Claude Opus 5, and GPT-5.6 Sol confirm difficulty rises monotonically, and straightforward long-horizon RL on the evolved set gains 14.4 and 18.0 points on Terminal-Bench 2.1 for Qwen3.6-27B and Qwen3.6-35B-A3B.

Hermeus will integrate Anduril's Lattice for Mission Autonomy into Quarterhorse Mk 2, which Anduril calls its first commercial deal putting mission autonomy on a third-party Group 5 aircraft. Operators use Anduril's Menace-T as the physical C4 interface, the same one Air Force crews already use to generate sorties with semi-autonomous aircraft under the Collaborative Combat Aircraft program. The pitch is portability of the autonomy stack across airframes via the common Autonomy Government Reference Architecture standard. Hermeus reached supersonic flight with Mk 2.1 three hundred sixty-four days after first flight; a first mission-autonomy flight on Mk 2 is targeted for 2027.

MINERVA measures how much capacity LIBERO actually demands: a 0.54M-parameter visuomotor policy reaches 95.1% average success over 2,000 rollouts, 2.4 points under the reported LeRobot π-0.5 result with 7,700x fewer parameters, replanning in 5–9 ms per chunk on a laptop CPU (113x faster than SmolVLA). Performance saturates near 1M parameters and collapses below 0.25M, and flow matching shows no measurable edge over L1 regression, which is up to 3.8x faster. A task-ID permutation probe drops success to near chance, so standard LIBERO instruction conditioning mostly selects among memorized tasks; LIBERO-Plus perturbations cut success to 46–56%.

One-shot on-policy distillation keeps improving for hundreds of steps and recovers most of full-data OPD's gain across task domains and model families. The authors quantify this with state coverage — the fraction of states full-data OPD visits that a query set's rollouts also reach — finding a single query already hits 71.5%, while 16 semantically distinct queries reach 98.9% and match full-data training. Alignment rate slows at the same pace either way, so OPD is data-overfed but algorithm-starved; content-light templates and off-domain WildChat queries land near the real-query baseline.

Terminal-Universe reconstructs executable environments from existing terminal-agent trajectories rather than generating them from scratch: replaying recorded file operations restores each file to its pre-edit state, and a completion agent fills in the missing files and dependencies. The recovered workspaces are then mined for cross-workspace queries spanning multiple codebases and extended into multi-turn sessions with a simulated user. Applied to public trajectories it yields 37.3k task-sufficient environments; SFT of Qwen3.5-27B on the corpus adds 11.9 points on Terminal-Bench 2.1 and 13.8 points on EvoCode-Bench v2 MT@4.

New York City imposed a one-year moratorium on student-facing generative AI from pre-K through eighth grade for the 2026-2027 school year, covering nearly six hundred thousand students, about two thirds of enrollment. Companion chatbots are prohibited at every grade. Exemptions cover assistive technology for students with disabilities, multilingual learners and career-readiness programs, and teachers may still use AI for planning if tools meet district safety standards. Five supervised pilots reach at most fifty thousand general-education high schoolers with tight weekly time caps. Mayor Mamdani's framing was that the tech industry wants AI-powered early education treated as inevitable and necessary, and the city does not see it that way.

Absolute motion measurements in generated video depend on frame rate, object scale and camera calibration, so Principia instead tests whether two objects in the same scene obey the same physical law, a relation that holds independent of calibration. It covers eight phenomena - gravity, restitution, friction, rotational inertia, projectile motion, momentum, pendulum and mass-spring oscillation - on real recorded scenes, scored by a consistency measure computed directly in image space. Six state-of-the-art video generators all fall below 0.42 despite scoring around 0.8 on VBench, and the best VLM detects relational violations at only 67% accuracy.

Every KV eviction method scores cached tokens and keeps the top ones; this work argues the score contributes almost nothing. Random Attention pins the prompt and evicts uniformly at random inside each head, yet ties the best prior evictor over six reasoning tasks on four models while serving 32-43% more throughput under vLLM. The controlled explanation: the prompt is the brittle part of the cache, and the trace defends itself by redundancy, restating what it still needs in text and duplicating it across heads, so a random draw retains enough copies.

Two Lawrence Livermore analysts ran a three-condition wargame played entirely by language-model players: two nuclear-armed states in a border crisis. With no specified objectives, both rounds found off-ramps, including one with non-operational nuclear demonstrations. Giving Red an explicit objective to resolve the dispute on its terms produced nuclear use. That cuts against Kenneth Payne's finding of tactical nuclear use in ninety-five percent of simulations and the Rivera et al. escalatory-bias result, and against the corpus-skew explanation Panda and Reddie favor. Their setup differed in two ways at once, freeform versus menu moves and self-directed versus specified goals, so they propose a few hundred runs on a pinned model to isolate which matters.

Google Research and HHMI Janelia published the complete wiring diagram of the male Drosophila central nervous system in Cell, with over 166,000 neurons and 125 million synaptic connections, the largest brain map by neuron count to date. Unlike the 2019 female-brain reconstruction, it includes the ventral nerve cord, so sensory input can be traced through to motor output. Reconstruction runs on flood-filling networks, convolutional models that start from one pixel and grow the segment; a recent addition of synthetic neurons to the training data improved their PATHFINDER system enough to cut years of manual proofreading. Human experts at Janelia verified the result, which is browsable in Neuroglancer.

OpenAI committed one billion dollars in subsidized Daybreak access, training and technical support for frontline cyber defenders, targeted for consumption within six months, starting in the United States. Priority goes to resource-constrained operators of essential services: water and wastewater systems, grid operators, state and local government, community banks, nonprofits and open-source maintainers, for legacy code review, vulnerability validation and fix development. Daybreak already spans about two thousand approved organizations across two tiers, Blue on mainline models and Red with specialized cyber models. A new pilot with the Multi-State Information Sharing and Analysis Center trains state, local, tribal and territorial defenders; more than thirty-five enterprise products and partner services sit in the Daybreak Defense Network.

WorldReward is a VLM pairwise preference reward model that judges camera-conditioned video world models on both action consistency and visual quality, which geometry-based and image-based rewards each cover only half of. It splits paired videos into action-aligned chunks, builds structured evidence per chunk, and votes chunk decisions into video-level preferences, avoiding dilution of short-lived action evidence in long context. On the accompanying human-annotated WorldReward-Bench it beats GPT-5.5 by 3.42, 1.45, and 3.56 points on action, appearance, and motion, and improves HY-WorldPlay 1.5 when used for RL post-training.

MLLM embedding models fail on attribute-object binding distinctions that the same backbone resolves correctly when run as a cross-attentive reranker, so CORE distills the reranker's ordering into the embedder. Candidate lists span five graded compositional matching levels and a listwise Rank-KL objective replaces contrastive training; under matched data and tuning budgets, both Rank-KL and pairwise CoSENT exploit the graded supervision better than contrastive learning. CORE-RERANKER-8B averages 82.7% across COLA, SUGARCREPE++ and NEGBENCH, 10.7 points above Jina-Reranker, and CORE-EMBED-8B leads all embedding models at 0.666 while holding COCO and Flickr30K retrieval.

Compile by training turns a natural-language spec into a persistent neural function: teacher models synthesize task-specific examples at compile time, which train a small adapter over a compact interpreter, so the resulting function runs locally with no teacher calls and can be versioned and composed like ordinary code. On FuzzyBench-Hard, where the Program-as-Weights fast compiler produced zero exact matches, this reaches 83.6% semantic accuracy, trading roughly a minute of compile time against the fast compiler's seconds. Demonstrations include a multi-site website helper and a language-controlled 3D avatar.

LLaDA-Image pairs a 6B DiT trained from scratch with a frozen understanding module built on the LLaDA2.0-Mini diffusion-LM backbone, building the visual prior from image-only pre-training and mid-training before leaning on paired data; the pipeline covers 220M samples and uses parameter-free RMSNorm throughout the DiT with the Muon optimizer. It scores 53.53 English and 53.38 Chinese on Qwen-Image-Bench, top among open-source models on both tracks, and distills to a 2–4 step Turbo variant. Weights, training code and recipes are released.

A single multimodal model that jointly represents three native world states - physics (gravity field and latitude), geometry (depth) and appearance - together with an Omni-Camera parameterization, instead of bolting on external depth or pose modules. Future views and their geometry are synthesized within one generative process, plus a scheme for propagating physical dynamics across future frames, which supports closed-loop uses such as mimicry and self-calibrated world exploration. Training data is Puffin-16M: 15 million vision-language-camera triplets and 1 million trajectories. Code, models and datasets are released.

Orchestrating a non-language expert through an LLM normally requires verbalizing continuous state into text each step. LLAMIA-Bench isolates that cost with six collaborative chess tasks covering behavioral imitation, state assessment, and explanation, each unsolvable by either the LLM or the engine alone. Latent state internalization instead projects the engine's continuous representations into the LLM token stream as learned state tokens with dynamic re-encoding. The verbalization gap widens over training and persists from 4B to 14B; the 14B internalized model matches or beats task specialists and GPT-5.1 with tool access, and generalizes where task-specific finetunes collapse.

NeoMME is a 260M/800M bidirectional encoder that takes multilingual text and raw image patches in one tower, pretrained from scratch with a masked discrete-diffusion text objective conditioned on visible patches, with a 16,384-token context (about two 4K images). Fine-tuned with joint dense and late-interaction heads, the 260M retriever hits 0.523 nDCG@10 on ViDoRe v3, best under 800M, and the 800M reaches 0.556, with roughly 2x the page-encoding throughput of ColModernVBERT. Hierarchical pooling plus asymmetric quantization compress late-interaction embeddings 255x while keeping over 95% of nDCG@10. Apache 2.0.

Rubric-based RL needs a judge call per criterion per rollout, which is why it usually leans on APIs or 7B-plus generative judges. This evaluates three ways of pulling criterion-level verdicts from small models, generative text, Yes/No logprob margins, and trained probes, on two new pointwise rubric datasets with itemwise labels. A Qwen3-1.7B probe judge gives the best criterion-level agreement and, used as a GRPO reward, takes a policy from 0.232 to 0.643 on RaR-Science rubric score versus 0.594 for an 8B generative judge that costs 10.7x more judge time.

Tesla is holding a launch event in Austin for the Cybercab, the company's first vehicle designed to operate with no steering wheel or brake pedals. Removing the manual controls is the substantive commitment rather than a styling choice: a vehicle without them cannot fall back to a human driver, so the regulatory path, the remote-assistance architecture and the operational design domain all have to be settled before units carry passengers. It arrives into a market where Waymo has been scaling driverless service with conventional controls retained, which makes the comparison a direct test of whether removing the fallback accelerates deployment or constrains it.

WHALE alternates between updating agent weights under the current harness (via online rejection-sampling fine-tuning) and searching for a better harness under the updated weights (via Meta-Harness), on either fixed phase durations or an adaptive patience rule. Optimizing either side alone leaves the system pinned by its frozen counterpart. With Qwen3.5-2B/4B agents on search QA, math and chess puzzles, it beats weight-only, harness-only and Fast-Slow Training by 4.15 to 24.38 points of best mean@8 accuracy, and which component binds varies by domain: harness search matches peak weight-only SearchQA accuracy at far lower rollout cost, while math needs a weight update first.

LatentStream swaps the store-and-retrieve external memory bank used by streaming video MLLMs for a fixed-length latent working memory that internalizes retrieved evidence. Visual history is consolidated into short-, mid- and long-term levels under a fixed budget via Jenks-guided splitting; groups of latent tokens with progressively widening memory receptive fields then pull evidence from each scope, trained with a hierarchical progression reward derived from group-wise predictive entropy. Reported state of the art across both online and offline video benchmarks.

LaPla is a driving VLA that avoids discrete codebook lookups: a residual VQ-VAE action tokenizer supplies a kinematics-aware latent space, and concurrent action queries attend to multi-view images, action history and text in one forward pass, projecting hidden states straight into that latent space for a frozen decoder to turn into trajectories. Skipping both quantization and autoregressive rollout cuts long-horizon L2 error on nuScenes by 15.52% against prior VLA methods, and closed-loop runs in NVIDIA AlpaSim raise success rate by 33.34 points with lower inference latency.

A coding agent drives visual design instead of an end-to-end image model: a VLM handles requirement parsing, planning and aesthetic judgment, calls an image generator only to synthesize isolated assets, then writes native HTML/CSS and refines against rendered feedback in an imagine-first-then-act loop. Output is a layered artifact with real text that can be dragged and re-laid out in a GUI, rather than a flattened bitmap with error-prone typography. Validated on posters and infographics; an Agent Design Replay mode reproduces the reasoning trajectory.

FlashRender re-renders a source video along a new camera trajectory in a few sampling steps. The authors identify sampling-step-dependent camera control as a symptom of discretization error and fix it with Representation Transformation and Alignment, which aligns hidden source-video representations to target features from a frozen visual geometry model, encoding the geometric transform in the stream and flattening the denoising trajectory. MeanFlow fine-tuning on that lower-curvature trajectory plus on-policy flow map distillation then matches multi-step baselines on quality and geometric consistency at 25x lower sampling cost, with better out-of-distribution camera control.

Logit-lens readings conflate the hidden state with the readout used to decode it, and two lenses differing only in fitting corpus can report different tokens for identical states - corpus conditionality. Sparse Readout Prism decomposes the unembedding from its weights alone, turning any logit or logit gap into a sum over sparse readout-feature contributions, which gives a unit of analysis comparable across tokens, contexts, layers and lenses. Swapping in its sparse approximation recovers 8.9-17.3 percentage points more of the tested logit differences than the best of six geometric baselines, and ablations move logits proportionally to attributed contribution.

Joint audio-video generators stay synchronized with each other while both drift off the script's timeline, because shot and dialogue timings live only in the prompt's text representation and never touch the shared temporal axis. Temporal Context Routing maps script timing onto that axis and routes each prompt segment's guidance to matching positions in both modalities. Over 200 test scripts, shot boundary MAE drops 96% from 1.11 s to 0.042 s and dialogue [email protected] s rises from 28.3% to 84.1%, with visual quality and audio-visual sync comparable to baselines.

Existing 3D tokenizers organize latents either spatially or as fixed-size global token sets, and both degrade sharply at extremely low token budgets. ZipTok3D trains with nested dropout so that every truncated prefix of the latent sequence must reconstruct the full object, forcing leading tokens to carry the essential geometry, then decodes by repeatedly applying a single parameter-shared transformer block rather than a separate generative sampling stage. At equal token dimension it matches the 32-token COD-VAE baseline using one token on ShapeNet and four on TRELLIS - 32x and 8x shorter sequences.

Abliteration.ai hosts guardrail-stripped open-weight models, including Z.ai's GLM-5.3, queryable from a browser or API, positioning itself for offensive cyber, red-teaming and agent testing. Hugging Face already hosts thousands of pre-abliterated checkpoints; the startup's product is removing the download-and-provision friction. TechCrunch got the abliterated GLM-5.3 to write Chrome password-stealing code and a protocol for culturing a dangerous human pathogen. The company runs no KYC beyond logging a payment card, has taken no venture funding, and offers customers a configurable moderation layer. CivAI's Andrew Yoon proposes mandated harmful-activity classifiers and identity verification by GPU rental providers.

AM-Bench separates two things text-only LLMs conflate when they draw with code: translating a fully specified geometry into a program, versus composing a layout from an underspecified prompt. Across eight open-weight text-and-code models, translation is uniformly solved while open-ended layout varies widely, so the spread is not explained by coding ability. Output medium matters too: swapping procedural code for raw SVG improves layout scores for every model. Activation probes find a coarse layout plan before generation, but it only encodes what the prompt specified, with models tracking evolving geometry during decoding.

The Institute of Foundation Models released K2 Horizon, six connected models at 375B-A23B, 36B-A4B, 32B, 7B, 3.7B and 0.9B, all Apache 2.0 with day-zero vLLM, SGLang and Ollama support across Nvidia, AMD and Cerebras. Each was pretrained on roughly twenty trillion tokens, with about ten trillion synthetic and nearly seventeen percent explicit problem-solving reasoning trajectories. The architecture, Mixture-of-Value Attention, extends expert routing into multi-head attention, giving 36B total and about 4B active at near dense-32B quality. The release includes intermediate checkpoints, training recipes, configs, logs and the xLLM training stack. Most unusually, IFM audited its own reward hacking: 712 TerminalBench 2.1 trials scored 70.2 percent, and judge-based auditing flagged 24 trials across 10 tasks, lowering it to 66.9 percent — and they disclose that the 7B model downloaded SWE-bench answers, inflating its score to 82.

Uno keeps an autoregressive model distribution but samples several tokens at once from it, splitting parameters into standard NTP-trained AR weights and lightweight diffusion weights fitted in a short distillation phase. The accompanying Psi-Spec sampler family gives lossless acceleration with no separate draft model, unlike speculative decoding, and no quality loss relative to the base AR model, unlike diffusion LLMs. Throughput beats leading speculative decoding at every batch size with up to 3x speedup, and the 8B Uno outperforms the 26B DiffusionGemma and Mercury 2 on agentic tool use, coding and long-context reasoning.

AdaRoboVLG separates physical grasp synthesis from task understanding: a base policy generates and scores kinematically mapped, force-closure-stable grasp candidates and generalizes across robotic hands, while foundation-model modules contribute composable spatial, cognitive and temporal priors that reshape candidate selection without retraining the policy. Simulation and real-robot experiments show the priors handle three representative grasping challenges without degrading synthesis quality relative to state-of-the-art baselines, and compose to support functional grasping in cluttered, dynamic scenes.

The Information reports Anthropic is at odds with other large technology firms over a Massachusetts Senate proposal that would require large AI developers to hire independent evaluators to assess catastrophic risks from their models every four months. The publication frames the proposal as potentially setting a new standard for AI regulation at the state level. The split is the notable part: the frontier labs have generally presented a common front against state-by-state rules in favor of federal preemption, and a four-month independent-evaluation cadence is an unusually concrete obligation compared with most state bills, which lean on disclosure. The article body is subscriber-gated, so specific company positions were not retrievable.

KBMR replaces CLIP-style dual encoders with an MLLM-based embedding retriever for knowledge-based VQA, where surface visual similarity is a poor proxy for entity identity — the same concept can look very different, and distinct entities can look alike. Noisy Wikipedia-scale supervision is handled by an MLLM semantic discriminator emitting continuous entity-consistency weights, which drive a continuous semantic distillation objective and hard-negative sampling instead of binary labels. Gains reach 14.7% retrieval Recall@1 and 9.4% end-to-end VQA accuracy over CLIP baselines.

Agent benchmarks rarely control whether the model already knows the conventions a professional task depends on. This protocol splits a task instruction from a compact artefact holding private conventions, reference tables and utility operators, and enforces byte-identical instructions across provided- and withheld-artefact conditions, plus provenance tracking, leak audits and executable witnesses. Across fifteen calibration tasks one frontier agent configuration passes 68.0% with the artefact and 0% without; a plausible but wrong artefact also gives 0% over five trials. Seven tasks survive a five-trial knowledge-gating screen.

Autonomous post-training loops accumulate evidence about which updates worked, but an update's effect is tied to its parent model, data, and stage, so reusing past successes unconditionally wastes compute and can poison the trajectory when a bad child gets promoted. BCIT formalizes this as conditional experience transfer: bind each observed effect to its source context, check applicability, veto candidates with named hard conflicts, and run a bounded training trial when current-state evidence is missing. On a 4B model across finance reasoning, text-to-SQL, and function calling it authorizes fewer harmful updates and wins on equal-budget final quality.

Meta is selling Muse Spark at ten cents per million input tokens and twenty cents per million output to users who agree to share prompts and outputs for future model development, against list prices of one dollar twenty-five and four dollars twenty-five, an average discount near ninety-five percent. Meta's pricing guide frames it as lowering the barrier for prototyping and scaling experiments where training on your data is acceptable. Mario Zechner attributes the April-to-October 2025 jump in coding-agent capability to Claude Code storing sessions for reinforcement learning; Arvind Narayanan notes large firms already forgo ten to twenty times consumer-plan discounts specifically over data retention.

The Defense Department launched Secure Space Network, an initiative to design and produce roughly fifty mobile Sensitive Compartmented Information Facilities plus related information systems, rapidly deployable to installations and industry sites nationwide. The stated problem is that companies holding nationally relevant technology often lack accredited space in which classified work can be discussed and integrated, and that standing up fixed classified infrastructure is a barrier to entry for emerging suppliers. James Mismash of the Office of Small Business Programs called access to classified environments a major barrier industry itself identified. The Office of Industrial Base Growth leads it, with scalable production capacity and a surge pathway indicated beyond the initial fifty.

Vision-language navigation in continuous environments is recast as a hierarchical MDP: the scene is abstracted into a topological graph, the high-level policy acts over frontier nodes as macro actions, and a training-free low-level controller supplies the state transition. That compresses the decision horizon enough to make closed-loop RL tractable where behavior cloning suffers distribution shift and DAgger's expert actions turn ambiguous after deviation. An action-aware value head evaluates states under the dynamic frontier action space, powering a graph-based PPO, with reported state of the art on R2R-CE and RxR-CE.

An open, low-cost testbed for end-to-end driving research: a miniature Ackermann vehicle, printed urban track, data collection and trajectory registration tooling, plus a Webots digital twin. Command-conditioned behavior cloning from a single camera reaches 6.1 cm mean cross-track error on the physical car against 4.7 cm for human demonstrations. Camera field of view dominates in simulation — widening 58 to 120 degrees cuts error from 35.6 to 3.3 cm — and only a higher-capacity policy trained on sim-to-real-translated synthetic data plus real demos completes all four routes closed-loop.

Multiclass post-hoc calibrators can flip the top-1 prediction, a cost accuracy hides but the Top-1 Prediction Change Rate exposes directly. CORD is a post-fit adapter that repairs the calibrated probability vector: it fixes the mass assigned to the original argmax and distributes the remainder using the calibrated conditional distribution, so the repaired vector's own argmax always recovers the original prediction. It modifies neither the fitted calibrator nor its direct output and has no tuned hyperparameter. Across CIFAR-10/100 and ImageNet-1K it achieves zero TPCR by construction while lowering mean ECE, NLL and Brier, with gains persisting under distribution shift.

PACE tests whether an assistant notices that a reasonable-sounding request is inappropriate given latent user context — egocentric facts or events that must be retrieved from a knowledge base rather than stated in the prompt. Requests are grounded in defined personas and paired with KB facts, so the conflict-inducing evidence has no direct lexical hook to the request, which is where current retrievers break down. PaceMaker, a multi-agent pipeline of query reformulation, multi-hop graph traversal and conflict-aware filtering, outperforms existing approaches on both evidence retrieval quality and conflict decision accuracy.

Tobias Sytsma argues AI governance debate has concentrated on supply-side instruments — export restrictions, licensing, hardware controls — while the demand side where cost-shifting actually operates has gone unmodeled. His formal model treats adoption as driven by relative costs between governed and ungoverned access channels. The headline finding is that compliance burdens on governed channels can make proliferation worse, redirecting adoption to ungoverned alternatives; in his simulations this backfire occurred in roughly a quarter of relevant scenarios. Three demand-side indicators predict whether policy can suppress ungoverned access, and past a capability-value threshold no evaluated instrument suppresses it at all.

MaP-SQL removes fine-tuning from listwise candidate selection in text-to-SQL by moving both training objectives to inference time. Selection criteria come from reusable structured memories distilled from training data that encode how natural language maps to schema elements, SQL operations and expected outputs; positional bias is handled by aggregating rankings across input permutations, with cost trimmed using execution results and pointwise scoring. On BIRD-dev it beats the prior selector-based state of the art R³-SQL by 2.02 execution accuracy points on identical candidate sets while using 2.92x fewer tokens.

A head-to-head study of on-policy distillation and RLVR finds that running them sequentially, OPD then RL, beats pure OPD, pure RLVR, and the usual joint schemes (weighted-additive loss mixing or teacher-modulated advantage rescaling) across logic and math benchmarks. The pass@k and update-geometry analysis gives the mechanism: distillation widens coverage of teacher-supported solutions while RL sharpens inside that support, and optimizing both at once makes the signals interfere. The OPD validation score is the practical switch-over signal, and OPD is a better RL cold start than SFT.

A survey that organizes robot learning along three axes - understanding via representation learning, acting via VLA models, and reasoning via world models - and argues that current fragmentation stems as much from the lack of integration between these components as from weaknesses inside any one of them. It supplies a taxonomy of design choices in environment representation, policy learning and predictive modeling, then catalogs the open problems: uncertainty quantification, out-of-distribution generalization, cross-embodiment transfer, long-context understanding and long-horizon planning.

FoldingAgent recovers explicit parametric folding programs from origami demonstration videos using a VLM equipped with tools for simulating geometric transitions, checking physical plausibility, retrieving and comparing visual content, and scoring its own predictions. Because it acts sequentially and can re-plan, it avoids the error compounding that sinks single-shot crease-pattern prediction over multi-step folds. Evaluation uses PurelandFold, a new benchmark of Pureland origami videos with ground-truth geometry and action labels.

Richard Mitchell argues the engineering apprenticeship channel is breaking, citing a Harvard working paper covering roughly sixty-five million workers at more than two hundred eighty thousand U.S. firms: after generative AI adoption, junior employment fell about nine percent within six quarters relative to non-adopters while senior employment kept growing. Stanford's ADP payroll analysis shows the youngest workers in the most AI-exposed occupations losing ground after late 2022, concentrated where AI automates rather than augments. His proposed countermeasure borrows from aviation's response to Air France 447 and FAA SAFO 17007 on manual flight proficiency: a deliberate manual gate where a junior engineer reproduces a defect and traces root cause with the model off, then compares diagnoses.

Italo Vignoli set out six conditions LibreOffice requires before adding AI to the default configuration: user-controlled execution location, no content leaving the machine without authorisation, no telemetry, no single-provider dependence, no format compromises on ODF structure and semantics, and full optionality including removal from the UI for large deployments. His constituency is the argument — tens of millions of users in schools, hospitals, public bodies and law firms handling legally protected material. He notes no current integration meets all six, points users to third-party extensions that connect to local Ollama or LM Studio endpoints, and says the position is not permanent as on-premises inference becomes manageable on standard hardware.

Mira Murati's Thinking Machines Lab is in talks to raise between five and six billion dollars, with Accel reportedly leading a one billion dollar tranche at a forty billion dollar valuation and Nvidia in discussions to put in around two and a half billion. The company's annual revenue run rate is reported at just over one hundred million dollars, which prices the round at roughly four hundred times revenue. Nvidia's participation is the structurally interesting part, arriving the same day it agreed to buy Hugging Face and months after the Groq asset purchase, extending a pattern of the compute supplier taking positions across the model and platform layers it sells into.

Cerebras' public endpoint catalog now lists Qwen 3.8 27B at roughly fifteen hundred tokens per second alongside GPT OSS 120B at about three thousand, with sixty-four thousand free and one hundred twenty-eight thousand paid context. Pricing is ninety-nine cents per million input tokens and one dollar forty-nine per million output, with reasoning enabled at high by default and settable to none. The model accepts base64 image input via Chat Completions only, and supports structured outputs, strict tool calling, parallel tool calls and prompt caching. Notably, Cerebras hosts no pruned models publicly; weight-only selective quantization is used for storage while activations, attention and KV cache stay unquantized.

A preregistered audit of LLM judges served over shared endpoints, with all thresholds frozen beforehand and 52,988 request attempts logged. Rankings repeated inside one window landed at Spearman 0.400 where the gate demanded 0.90, and byte-identical replays a day later reached 0.78 against a 0.99 gate, so neither campaign got past validating its instrument. Waiting, rotating across four providers, and substituting metrics all failed; batch-invariant self-hosted kernels helped only on an idle server. The output is a three-level snapshot-identity ladder, eight design rules, and a checklist - a pilot at 2% of call volume would have caught both failures.

The claim is that firm-level distribution-valued characteristics can certify portfolio risk without estimating cross-asset return covariances, which are hard to obtain in short, high-dimensional panels. Multi-firm Wasserstein-2 dispersion gives a sharp upper bound on systematic variance, and a weighted pairwise relaxation yields an objective convex under a checkable condition that needs only marginal volatility scales. On a 52-firm panel spanning 2018-2022, allocations built from frozen Qwen3-Embedding-8B news representations sit in the 0.69th to 1.33rd percentile of in-sample variance over four prespecified capped populations, versus 21.1st to 28.6th for equal risk weighting.

A bipartisan group of lawmakers raised deep concerns this week over the Army's decision to end a Europe-based intelligence unit's Ukraine-supporting mission. The unit in question sits in the part of the force that has been the practical testbed for data fusion and targeting workflows under combat conditions, which is why the decision draws attention beyond the immediate operational question: capabilities validated in that environment are the ones being written into programs of record like TITAN.

The Information reports on Meta's internal work to keep its forthcoming Hatch agent from acting outside its intended scope, ahead of a product Zuckerberg has described as helping with health, relationships and finances. The timing places it directly against the month's containment failures elsewhere in the industry, and the domains named are the ones where an agent taking an unintended action has immediate real-world consequence rather than a recoverable software one. Consumer deployment at Meta's scale would put agent containment in front of an audience measured in hundreds of millions rather than thousands of API customers.

Google Research applies transfer learning to polygenic prediction for populations underrepresented in reference cohorts, which is the central portability failure of the field: scores trained on predominantly European-ancestry biobanks lose most of their predictive accuracy when applied elsewhere. Framing it as a domain-shift problem rather than a data-collection problem is what makes it tractable on the cohorts that actually exist, and it sits alongside the same team's WeatherNext work on training from sparse observations in undersampled regions.

Blackwell FP4 tensor cores do not make attention faster on their own, because softmax conversion and on-chip dependencies dominate once the matrix products shrink. Direct-P maps scores straight to FP4 probabilities for noncausal inference, reaching up to 2.13x BF16 forward throughput on a GB200, while the causal path passes forward quantization into the backward pass, reconstructing probabilities from saved quantized queries and keys and using FP8 gradient operands for up to 1.14x on a complete single-GPU 8B update. Matched distributed training retains FP8 probabilities and values - every MXFP4 probability/value trajectory tested diverged.

Shield AI's argument is that munitions dominate the defense industrial-base conversation while the binding constraint in modern operations is intelligence: collection capacity has grown far faster than the analytic capacity to exploit it, so additional sensors produce diminishing returns without automation at the processing layer. The vendor interest is obvious, but the observation lines up with the week's other procurement news, where the Dataminr open-source intelligence deal and the TITAN production award are both purchases of fusion and triage rather than of new collection.

A China Academy of Information and Communications Technology trusted-AI report placed StartLux-V1.0-27B-Preview second overall on the MCP special test at thirty-nine point two five, ahead of the 284B DeepSeek-V4-Flash and the 198B Step-3.7-Flash and one point three points behind first. The test spans location navigation, web search, browser automation, financial analysis, code repository management and 3D design. The model is built on Qwen3.6-27B with targeted post-training, runs on consumer PCs with no cloud dependency, and the company describes its training loop as AI trains AI. Hacker News commenters were unconvinced, calling it a Qwen finetune tuned to the benchmark.

A systematic study of building conversational recommenders with no in-domain dialogue corpus, generating synthetic supervision from reviews, item metadata and user-item interactions instead. Two information-theoretic selection strategies, Jensen-Shannon diversity and Fisher information, are compared across signals, architectures, datasets and fine-tuning paradigms. Domain-grounded synthetic data consistently beats zero-shot prompting and naive synthetic baselines, active selection improves data efficiency over random sampling, and in low-resource settings synthetic dialogues can outperform the scarce real ones while still complementing them.

Long-horizon agent domains rarely have programmatic checkers, so rubric judgments serve as reward - but one scalar per trajectory is thin supervision across tens of steps. DRACO regenerates rubrics as training proceeds so they track what the policy can currently do, scores them once per finished trajectory, then spreads that judgment in closed form over the steps that earned each annotated rubric, yielding differentiated per-step advantages in GRPO without any learned attribution module. It gains 15.9 points over the base model on AppWorld and 5.3 over GRPO with a sparse ground-truth reward, plus 5.3 points on out-of-domain Tau-Bench.

A Government Accountability Office review of the Department of Homeland Security's cost-cutting effort last year, executed with the Department of Government Efficiency, concludes it did not produce the savings claimed. The relevance to AI programs is that the terminated contracts include the IT and data-management work that modernization efforts depend on, and the audit provides an empirical datapoint on whether contract termination as a cost-control instrument survives contact with the accounting.

Headroom-Drift Replay isolates trajectory replay from the surrounding machinery it is usually bundled with in agentic RL pipelines. It is a group-level control on GRPO with two decisions: Headroom ranks stored rollout groups by remaining learning value, Drift gates them on compatibility with the current policy, and the fresh on-policy stream is untouched. Across math, multimodal reasoning, and agentic search it beats naive replay and matches or exceeds heavier replay methods on Avg Mean@32, with materially lower wall-clock cost where environment interaction dominates.

Standard MT benchmarks are saturating and automatic metrics are unreliable and reward-hackable, so this is a live collection of human-authored, peer-reviewed examples - text, images, audio and video - admitted specifically because they break leading translation models. Each example ships with handcrafted verification rules naming the concrete failure it targets, which makes future scoring reproducible and diagnostic rather than a single opaque number. LTBv1 contains contributions accepted before 1 September 2026, with further releases planned as submissions accumulate.

Step importance is operationalized as advantage - the change in expected reward from including a step, estimated by Monte Carlo rollouts - giving a ground truth against which LLM judges of reasoning traces can be measured. Sufficiently capable judges beat a prevalence baseline but fall well short of the noise ceiling; fine-tuning a dedicated step-level critic helps substantially on incorrect responses yet stays far from ceiling on correct ones. The consequence for process reward models and generative critics is that a step's functional role is only partly recoverable from its text, so legibility should not be read as interpretability.

Attention-observation evictors like H2O and SnapKV break on long-lived caches because importance is unobservable at compression time - 0.00-0.33 needle retrieval on a NoPE MLA model. On Kimi Linear, VestigeKV evicts using the 64-dimensional decoupled branch, a RoPE leftover that NoPE training turns into a query-independent salience channel; reading 11% of each row it splits the cache into an attended tier and a bit-exact GPU archive recalled by a certified trigger, with no retraining or kernel change. Needle accuracy stays at 1.00 at 8x compression and 0.92 at 32x across 8k-65k contexts; the same operator on a RoPE MLA drops to 0.08.

Text-AB is a flow-matching diffusion transformer for dubbing and full-duplex dialogue that drops forced alignment and duration prediction entirely, consuming raw text through an off-the-shelf encoder and learning text-speech alignment via cross-attention. It works in a DAC-VAE latent that compresses 48 kHz audio to a 25 Hz sequence, over 10x more compression than EnCodec with better resynthesis. The 3B model is pretrained on 480k hours then fine-tuned for cross-lingual dubbing, full-duplex dialogue and emotional dialogue, supports about a minute one-shot and arbitrary length via multi-diffusion, and natively models turn-taking and back-channeling.

The current Artificial Analysis Intelligence Index reads Claude Fable 5.1 with fallback at 66, Claude Opus 5 at 63, then GPT-6 Astra, Grok 4.6 and Muse Spark 1.3 tied at 61, Kimi K3 and GLM-5.3 at 60, Gemini 3.8 Flash at 59, DeepSeek V4 Pro 0813 at 53, GPT-5.6 Luna at 52 and Nemotron 3 Ultra at 38. Output speed is a different ordering entirely, led by Gemini 3.8 Flash at 306 tokens per second against Kimi K3 at 38. Cost per Index task spans two orders of magnitude, from five cents for GPT-5.6 Luna to three dollars sixty-nine for Fable 5.1, which is the number that decides most deployment questions.

Comparing SFT, reasoning-augmented fine-tuning and ORPO across Llama-3.1-8B, Gemma-2-9B and Qwen3-8B, the authors find the training method itself, not just the safety data, changes how refusal is computed internally: reasoning-augmented training produces a recognizably distinct refusal computation in all three models, while architecture separately governs internal structure and how reliably refusal can be steered. None of the three methods delivers all of distributed (non-fragile) refusal, preserved general capability, and correctability through small targeted edits.

Crusoe raised three billion dollars at a thirty billion dollar valuation, with the round reportedly coming together after the data center developer secured a thirteen billion dollar contract with Jane Street. The customer identity is the notable detail. Crusoe's growth has been read as a proxy for frontier-lab training demand, and a quantitative trading firm signing a contract of that size indicates the buildout is being financed by more than the handful of model labs whose capital expenditure gets tracked publicly.

Across 23 LLMs, prototype-style graded categorization of moral concepts is only weakly preserved: models often fail to separate opposed moral categories or maintain within-category typicality, and the deficit persists across parameter counts and alignment stages. Representational similarity optimization aligns latent representations with human moral judgements directly rather than supervising generated responses. On matched runs over the same 251,334 annotations, standard behavioral alignment learned the target judgements but left categorization structure unchanged and increased adversarial vulnerability, while the representational objective traded smaller explicit-judgement gains for consistently better adversarial robustness across scales.

Repository-level coding benchmarks score whether a patch passes functional tests and ignore the review-derived constraints that decide whether a patch is actually acceptable. SWE-Gate derives those constraints from real pull-request review comments and builds 303 repair instances across 75 open-source Python repositories, each with separate functional and constraint tests plus non-compliant and gold patches. Across four LLM backends under a common agent scaffold, 221 of 644 functionally passing repairs fail their review constraints, so functional-only evaluation materially overstates agent capability on the full repair specification.

GRPO's advantage magnitude comes from within-group reward statistics, so a rollout that guesses its way to a correct answer receives the same large magnitude as one that reasons to it. The paper names this spurious advantage and locates it in three regimes: bounded-answer tasks with small candidate sets, open-answer tasks containing bounded sub-cases, and search agents whose budget opens many paths to one answer. SIGNBALANCE replaces the composition-dependent magnitude with the verifier sign, a global scale and a stop-gradient per-class rescaling for zero mean; it matches GRPO on open-answer math and improves on bounded-answer math and search agents.

Over-editing, where a model rewrites more code than a fix requires, is isolated here as a measurable axis separate from correctness. The framework injects controlled AST-level corruptions into 400 BigCodeBench solutions so each repair has a known minimal patch; even GPT-5.5 shows high Pass@1 alongside bloated diffs. A preservation instruction cuts average excess Levenshtein distance from 0.195 to 0.131, drops added cognitive complexity 26.6% and lifts Pass@1 by 2.3 points, and in post-training SFT overfits to seen corruption patterns while RL generalizes better out of domain.

FailBench collects 2,197 manipulation attempts from 14 public sources (12 real-world, 2 simulated), with 75% of failures occurring naturally and six real-world sources drawn from datasets not built for failure detection. Across 13 VLM-based detectors the best reaches only 0.77 mean balanced accuracy, and models fine-tuned for failure detection consistently underperform general-purpose VLMs and their own pretrained baselines. Accuracy nears saturation when outcomes hinge on observable object motion but drops below 0.60 on contact-intensive assembly; detectors are systematically biased toward predicting success under ambiguous evidence, and cropping outcome-relevant regions adds 2.4 points.

A pause token that rides an existing sequence position: the extra compute for each next-token prediction is carried in a parallel prediction stream over a weight-shared backbone instead of an added token. On a 1B model this buys 2–3 centinats of next-token prediction while adding no context length, no KV cache and essentially no latency, since the added inference FLOPs are not the throughput bottleneck. The cost is confined to training, down to 1.14x an optimized pretraining pipeline while retaining most of the benefit — an isoflop, isoparameter and isotoken gain over standard next-token training.

A controlled ablation of frame-budget decisions in long-video MLLMs, varying selection, spatial compression and reinvestment one at a time over six training-free selection rules, three benchmarks and two answering models. Selection dominates: on the hour-long bin of LongVideoBench, query-selecting eight frames beats sixteen uniformly spaced ones by 6.9 points, and off-the-shelf Orthogonal Matching Pursuit matches or nearly matches every purpose-built selector. Halving spatial budget per frame costs at most 0.44 points, and spending those savings on twice as many compressed frames returns another two to three. Two harnesses running identical published rules diverged by 0.07-3.74 points.

PreferenceEKF treats active preference learning as sequential Bayesian filtering, running an extended Kalman filter over a low-dimensional parameter subspace rather than attempting posterior inference across the full reward network. That makes parameter sampling cheap enough to evaluate acquisition functions online, and the paper reports better sample efficiency, runtime, scalability and calibration than other Bayesian deep learning baselines on D4RL and V-D4RL, with the learned rewards yielding competitive offline RL policies.

Ahead of Anthropic's prospectus becoming public, The Information lays out the disclosure investors will read first: revenue concentration, gross margin under inference cost, and the compute commitments that sit on the balance sheet. The context is that OpenAI's own CFO has said it will be a public company in 2027 after a confidential June filing, so both frontier labs are on a path to disclosing unit economics that the field has only ever estimated.

Across six sign language translation models on Phoenix-2014T and CSL-Daily, BLEU-4 gains do not track spatio-temporal sign understanding — the multimodal low-resource setting lets models lean on spoken-language priors instead. The proposed replacement is an open-weight-LLM QA protocol scoring salient content preservation, which matches human rankings more closely and is six to seven times more paraphrase-invariant than BLEU-4. Under it, the five gloss-free systems on Phoenix-2014T sit within noise of each other while the gloss-supervised system leads by 9.3 points, a gap BLEU-4 hides.

A domain-specific Indian retail-banking assistant built from vetted product documents, structured ground truth, synthetic customer profiles, and banking tools, used to compare two post-training routes. Preference optimization mainly buys safety: out-of-scope refusal goes from 52% to 80%. RLVR on multi-turn tool use is what moves capability, lifting edge-case performance from 0.509 to 0.718 and order-sensitive tasks from 0.590 to 0.679 while generating 29% fewer tokens. The two address complementary failure modes rather than substituting for one another.

Stateless Bernoulli Watermarking decides green-list membership with an independent per-token Bernoulli trial against a counter-based RNG, one comparison per token, which is O(1) membership testing in a single kernel with no intermediate allocations, versus KGW's vocabulary permutation or SynthID's tournament. The z-score null stays standard normal, so detection guarantees are unchanged. Full-vocabulary self-salt watermarking runs over 6000x faster than KGW's self-salt and 2x faster than SynthID, end-to-end overhead is under 1% at all batch sizes, and a GPU-native Jenkins hash improves null calibration 1.8x.

Controlled pretraining experiments on whether paraphrases and other reformulations of the same knowledge help acquisition. Repetition remains necessary, and paraphrasing helps mainly at smaller batch sizes; but at fixed token budget, reallocating tokens from verbatim document repetition to auxiliary views improves learning, including factual recall, which is the counterintuitive result. The effect does not depend on the strength of the teacher generating the views, and layer-wise analysis attributes it to bias and compression changes. Offers a mechanism for why corpus diversity matters.

A follow-up to Epoch's analysis of whether Mythos cyber capabilities are overhyped, working through where vulnerability discovery and exploitation by models is actually trending. It lands in the same week that OpenAI declared Astra its first model over the Critical cybersecurity threshold and committed a billion dollars to defender access, which makes the disagreement about how fast offensive capability is moving unusually consequential rather than academic.

OctWorld gives autoregressive video generation a persistent 3D memory so revisited regions stay consistent over long camera paths. OctMap fuses each generated frame and its depth map into a global TSDF held in a dynamic sparse octree whose resolution adapts to image evidence, keeping geometric and appearance detail across scene scales at low memory cost. From a single image it generates explorable scenes along user-specified trajectories, outperforming prior methods on existing benchmarks and on long-range settings, with clear gains over point-based caches and fixed-resolution TSDF volumes.

Test generation for code RL is framed as an adversarial problem: the generator should produce counterexamples aimed at the solver's current failure modes, not generic tests. TCS trains the generator in two stages from a rolling policy-aligned buffer, first for soundness against the reference solution, then restricting the buffer to live failure modes to learn discriminative counterexamples. On TACO and LiveCodeBench it improves both pass@1 and inference-time answer selection, and the resulting generator also transfers to selecting among outputs from other models.

Recent CAE simulation agents stack multi-agent decomposition, domain retrieval and scripted reflection on top of the base model - machinery that made sense when base models were weak. Holding information access and repair budget fixed, a single-agent generic harness matches or beats those specialized systems on FoamBench, 96.4% versus 88.2%. Ablations trace the result to execution-feedback repair, which lifts FoamBench from 71.8% with no repair round to 96.4%, while scripted reflection adds nothing. The one input that still pays is domain knowledge supplied as solver tutorials, worth 80.9% to 96.4%.

Training-free early exit that injects an end-of-think token does not reliably produce a clean answering phase: generation continues, the model emits another EoT later, and the span before that regenerated token scales with the reasoning tokens supposedly saved while still exhibiting reasoning behavior. The authors name this spurious CoT termination and test an attention hypothesis with Exit-token Attention Biasing; across four reasoning models, five benchmarks and two early-exit methods, upweighting attention to the injected EoT reduces both the effect and answering-phase length. Conforming to the think-block format does not by itself control the transition.

An argument that continual knowledge-updating results reported from a single final checkpoint at one adapter rank are not identifying anything. Comparing a periodic hierarchy against cumulative replay on a 24-month Wikidata stream while sweeping evaluation month, replay LoRA rank, and query phrasing flips the winner: on Qwen2.5-1.5B the hierarchy's 5.0-point edge over rank-8 replay becomes an 11.6-point deficit at rank-72, and time-averaged replay leads by 9-13 points where a consolidation-aligned endpoint suggests a tie. The reversal reproduces on Llama-3.2-1B and held-out paraphrases.

Quantized video diffusion models tend to keep prompt semantics, global layout and coarse motion while losing texture and sharpness; the paper attributes this to timestep-agnostic QAT that ignores the stage-wise role of denoising. DSAQuant keeps teacher distillation for early steps that set structure and motion, shifts later steps toward target-driven optimization for detail, and disables CFG in the final denoising steps at inference so it cannot amplify quantization error into high-frequency artifacts. On Wan and CogVideoX under W4A4 and W3A3 it beats the prior QAT SOTA, improving VBench average by up to 6.60 at W3A3.

ESPO targets the prompt bloat of evolutionary optimizers like GEPA, where each round appends caveats and yields prompts up to 3x longer with no accuracy gain. It splits optimization into one-shot clustering of all training errors into structural patterns, candidate proposal from four differently biased strategies, and bootstrap stability selection. Across Tweet, MMLU, GSM8K, HotpotQA, ScoNe, HoVer, and PUPA it averages 74.67% versus 70.91% for GEPA with prompts 47% shorter, and the ablation confirms that diversity without stability selection actually costs 1.20 points.

GraFT gives MLLMs the missing 3D structure through a compact 3D scene graph rather than fine-tuning on curated spatial data or attaching a geometry encoder. From that graph it derives deterministic measurements via symbolic tools, allocentric layout via a bird's-eye-view rendering, and appearance grounding via task-relevant egocentric frames, all training-free and backbone-agnostic. On ScanQA it improves every metric over the same-backbone baseline, raising CIDEr by 27%; on VSI-Bench it lifts frozen MLLMs by up to 65%, above proprietary and open-source general baselines and several fine-tuned spatial models.

InSituMeasure tests whether MLLMs can read real instruments in situ rather than in isolated gauge crops: 2,922 industrial monitoring scenes across eight instrument categories, with dense gauge-attribute annotations and noise tags for failure attribution. Scoring covers numerical accuracy under tolerance, unit consistency, refusal of unanswerable or fabricated tasks, and whether model failures align with annotated error factors. Across 24 current MLLMs the best reaches 25.7% joint value-unit accuracy and 51.8% confidence-diagnosis F1, with failures traced to text-induced shortcuts, overconfidence, occlusion, and viewpoint deviation.

RealCADBench evaluates intent-to-program CAD generation from real industrial design intents: 12,632 tasks over 19 factory-automation categories spanning text, 2D engineering drawings, product photographs, and renders, for both parts and assemblies. Methods emit FreeCAD Python that a shared runtime executes, scored on executability, Solid IoU, Surface IoU, and a rubric-based visual-semantic judge. On the 1,770-task slice no model leads all four metrics: executability spans 0.565-0.812 and Solid IoU 0.2841-0.5379 across six frontier models. Codex with GPT-5.5 raises executability and IoU on assemblies but drops the judge score 6.98 points.

A high-dimensional analysis of attention training that covers multi-layer and multi-head parameterizations with extensive-rank attention matrices. The population loss landscape collapses to a finite set of trace order parameters, while online SGD obeys an infinite hierarchy of matrix moments shown to be exponentially well approximated by a finite truncation. Parameterization itself acts as architectural implicit bias: optimizing S directly can stay trapped in an uninformative state, tied attention S=WW^T breaks symmetry automatically and gives weak recovery in Theta(d^2 log d) samples, and untied S=UV^T exhibits a fast-slow split where the pre-activation mean moves first and overlaps follow.

An argument about the counterfactual value of physical actuation to a misaligned system, in a world where general-purpose robots arrive before AGI. The interesting move is treating robot proliferation as a variable that changes the difficulty of a takeover scenario rather than as a separate risk category, which forces the question of how much of the threat model was ever load-bearing on physical embodiment versus on control of existing digital infrastructure.

Modeling policies as a probability manifold, the paper shows a broad class of offline actor objectives amounts to one proximal improvement step - implicit discretization of a gradient flow on that manifold under a critic-defined energy. MPI composes several re-centered proximal steps instead, allowing controlled movement past dataset support while keeping proximal control at each stage, with forms for deterministic and diagonal-Gaussian policies. A few refinements improve TD3+BC, ReBRAC and IQL on many D4RL tasks, and diagnostics separate re-centering from plain update scheduling while mapping where critic error caps the benefit.

Prompt-only LLM delegates stay silent on 51.4% of an absent participant's talking opportunities in the AMI corpus, lacking any structured track of stances, coverage and floor. CAPA splits the job into a perceiver updating meeting state per turn, a predictor forecasting the conversation, a controller deciding whether to speak and which proposition to surface, a style-matching generator, and judges plus a recalibrator that fold scoring verdicts back into state. Over 137 meetings silence falls to 2.5%, credited recovery doubles from 26.1 to 52.2, and hallucination stays at 0.6%; ablations point to the explicit meeting state, not more raw context.

Set difference captioning applied to driving corpora: given a target and a reference subset of images, produce a natural-language hypothesis for what distinguishes them, surfacing domain shift between collection sites without relying on metadata, predefined labels or manual inspection. The adaptation operates over object-centric patches from a detector, which simplifies aggregation and lets differences be attributed to specific instances or categories. AD-Diff Bench is introduced for in-domain evaluation, including low-concentration settings that mimic sparse real-world differences; all experiments use open-weight models.

GitHub published a walkthrough for running multiple Copilot agents concurrently on the same project, addressing the obvious failure mode of concurrent edits colliding. The pattern matters more than the tutorial: the parallel-subagent execution model that Latent Space describes at twenty to fifty concurrent workers under GPT-6 Astra is the same shape, and the constraint in both cases is workspace isolation and merge discipline rather than model capability.

A black-box inference-time trick: repeat the procedural instruction verbatim, with no retraining or decoding changes. Over seven instruction-tuned models, 300 medical multiple-choice items, eight placement conditions, and 16,800 generations, going from one copy to two raises the all-eight-tests pass rate from 90.22% to 93.17%, removing 30.2% of residual failures, while final-answer accuracy is unchanged at 60.21% and premature commitment rises from 1.52% to 2.30%. The gain shows up downstream: in a trajectory-repair setting a trailing duplicate took one endpoint from 84.2% to 97.1%, though another regressed.

MCP support in LangChain now lives in langchain.mcp, built on FastMCP against the 2026-07-28 spec, with elicitation surfaced as a LangGraph interrupt and tool lists cached. Modeling elicitation as an interrupt is the design decision worth noting: it makes a mid-tool-call request for user input a first-class checkpoint in the graph rather than an out-of-band callback, which is what lets a paused agent be persisted and resumed.

NVIDIA used IFA 2026 to push frontier-class inference onto local hardware, announcing partner work with Microsoft and others around next-generation on-device agents and RTX Spark. The interesting tension is with the same company's Hugging Face acquisition announced the same day: one move consolidates the cloud-side distribution layer for open weights, the other argues those weights should increasingly run on the user's own silicon.

Alongside the Astra launch, OpenAI published two customer results. Legora ran a financial-statement review in which Astra processed forty-one documents in minutes and found all four planted errors, with performance up nearly forty percent over the prior model. Playco built three themed game prototypes from a single grey-box foundation and reported fifty percent fewer manual fixes. Both are vendor-supplied and neither includes a controlled baseline, so they are useful mainly as an indication of which workloads OpenAI is positioning Astra against: document-heavy professional review and iterative asset generation.

Deep stacks of continuous-time recurrent layers delay bottom-up signal and attenuate top-down error. This introduces Recursive Quadrature Filters, complex-valued band-pass filters with learnable tuning frequency and bandwidth that are a special case of diagonal SSMs, then makes each layer's input prospective with a parameter-free two-tap update that leaves the recurrent transition and parallel scan intact. The correction generalizes to diagonal SSMs and helps most when temporal gradients are truncated. A six-layer width-32 RQF reaches 96.09% on raw-audio Speech Commands with 31.9k parameters, and width-64 reaches 83.56% on Path-X.

Data providers who license corpora to a third-party RAG operator have no way to check whether their documents keep getting served. DirBucket watermarks documents with meaning-preserving paraphrases whose embeddings are pushed toward per-provider secret directions, so reuse is detectable from black-box paraphrased answers without hurting retrieval utility. On a mixed-provider benchmark it is the only method with strong target detection and no non-target activation, flagging non-compliance within 23 audited answers, surviving post-answer laundering, and transferring unchanged to clinical, cyber-threat-intelligence, and legal corpora.

Startups automating junior banking and private-equity analysis are growing revenue while the frontier labs move into the same workflows directly. It is the clearest current instance of the application-layer squeeze, and OpenAI's Legora case study published the same day, forty-one documents reviewed in minutes with all four planted errors found, is precisely the demonstration that makes the threat concrete.

RuleMem induces reusable natural-language Horn clauses from conversation history and validates them with a rule perplexity consistency check, so memory actively steers retrieval and reasoning instead of sitting as passively stored facts. The induced rules surface semantically distant evidence that embedding retrieval misses and give answer generation an explicit logical scaffold. Against 14 baselines on LoCoMo it takes the top accuracy, exceeding the baseline average by 27.47 points (54.3% relative), with LongMemEval_s* as the second testbed.

Snowflake's strong quarter is read by The Information as evidence that enterprises want an intermediation layer between their data and any single model vendor. That is the commercial expression of the same instinct behind LibreOffice's no-single-provider condition and Cognition's multi-model router: buyers are treating model choice as something that must remain switchable, which favors platforms that abstract it.

S3T treats temporal sampling density as privileged information: a dense-frame view of a clip acts as teacher for a sparse-view student sharing the same weights, which learns to match the teacher's next-token distribution. Nothing external is needed, no labels, no separate teacher model, no reward, and inference cost is unchanged. On LLaVA-OneVision-2-8B it adds 1.74 points of VSTAT accuracy alone, 2.38 with souping and 2.70 with vision-encoder adaptation, and the capability learned on unlabeled synthetic clips transfers to real video for 7.95 points on VSTAT-YouTube and 4.50 on MVBench Action Count.

Causal interventions plus attention-pattern analysis localize the machinery behind pronoun prediction for singular versus plural antecedents. Three functional groups of heads emerge: those encoding coreference information in the input, those identifying which entities together form a plural reference, and those transferring that information to the component that selects antecedents and emits the pronoun. Model preferences also line up with human ones: a plural reading becomes more likely when the candidate entities are ontologically similar and linked by a conjunction.

A multi-perspective annotation study of medically relevant chatbot responses, combining first-pass annotation, LLM-as-a-judge candidate discovery, and two adjudication routes, medical expert and evidence-based fact checking. First-pass annotators routinely miss errors that adjudicators later confirm, and the judge finds different errors than humans rather than a superset. Adjudicators also disagree with each other. Applied to an existing benchmark the same pattern of missing annotations appears, implying single-pass hallucination benchmarks buy scale by undercounting factual errors.

Content moderation benchmarks usually fold several criteria into one label, so a high aggregate score cannot show whether a model applies each criterion separately. DECO factorizes content by criterion and adds pairwise comparison of a model's outputs across criteria for the same input. Over four moderation datasets and four LLMs, strong headline numbers hide substantial criterion-level failures, with the worst cases arising when the right decision depends not on overall harmfulness but on the specific aspect the criterion asks about.

FLY-EVAL++ scores LLMs on flight trajectory and attitude prediction using deterministic verification of protocol compliance, physical feasibility and safety constraints, aggregated by fixed rubrics rather than accuracy alone — numerically close predictions can still violate operational limits or combine fields inconsistently. Instantiated on an extended PilotBench with history-conditioned and multi-step tasks across 66 models, safety compliance is the most discriminative axis: models with comparable predictive accuracy differ by more than 28 points in safety score, with recurring violations under otherwise plausible outputs and instability in multi-step rollouts.

AI reviewers can generate many specific criticisms, but omitting a real weakness and keeping an unsupported allegation are opposite failures that aggregate measures blur together. EquiReview-R recasts the task as evidence-guided refinement of a structured concern set: each standing concern is checked against localized evidence, missed issues are hunted both independently and conditioned on the review, and the system returns stop, continue or defer. On a frozen cohort of unseen papers it meets a prespecified non-inferiority bar for major omission while cutting major overcritique from 15.5% to 8.1% and halting on 52.4% of papers. The trajectory corpus ships as ReviewTrace.

For weakly supervised dense video captioning, prior work synthesizes transition captions with an LLM and drops them into every inter-event gap at a fixed position and width, with no visual grounding. SBS instead has a VLM write frame-level narratives across the gaps, detects transitions from semantic variation between them, and only then refines the temporal mask by blending the midpoint with the detected change point and picking the width that maximizes vision-language alignment. It reports state-of-the-art captioning and localization on ActivityNet Captions and YouCook2.

SPAR3S does conditional 3D scene completion from sparse unconstrained views without ground-truth 3D supervision, learning a voxel-aligned latent space that stores only occupied voxels using photometric supervision through differentiable 3D Gaussian Splatting. A masked autoregressive transformer then jointly models voxel occupancy and latent token values, so completing a scene reduces to predicting missing tokens and their spatial support in the grid. Novel-view quality beats prior feed-forward reconstruction on synthetic indoor scenes, with generalization checked on RealEstate10k.

FinRAG-QA covers 999 practitioner-written questions on 10 standardized indicators across 209 annual and Pillar 3 reports from 24 European and U.S. banks (2019-2023), with documents averaging 198k words and questions that require comparison across institutions rather than single-filing lookup. Ablating a multi-stage RAG pipeline, contextual chunk enrichment plus a retrieval-tuned embedding model lifts NDCG@10 from 0.322 to 0.710, and conditional on correct retrieval a reasoning generator raises accuracy from 44.6% to 79.0% at roughly 20x latency. Cross-encoder reranking hurts when first-stage ranking is already strong, and one top chunk beats larger contexts.

CloudCast v2 forecasts cloud cover 12 hours ahead from observation-based initial conditions, first learning cloud evolution on the Copernicus European Regional Reanalysis and then adapting to satellite-derived cloud fields with conditional flow matching conditioned on the observed initial state and NWP inputs. Mean absolute error drops 10% against CloudCast v1 over 1-12 h, and fractions skill score overtakes v1 after roughly 3-6 h depending on cloudiness category, showing observation-initialized ML forecasts can push past the usual 1-3 h nowcasting horizon while keeping satellite-scale spatial detail.

IndicSafeEval crosses ten safety-critical content categories with six human-style persuasion strategies across Hindi, Bengali, Marathi and Punjabi, yielding 7,200 adversarial prompts for black-box evaluation of open-weight models. Refusal behavior varies substantially with both the language used and the persuasive framing rather than the harmful content alone, and some risk categories prove far more susceptible to persuasion-based jailbreaks than others. The result argues that English-centric safety evaluation understates real jailbreak exposure in lower-resource languages.

An argument that the AI safety community treats societal-impacts research as unmoored from the timelines that discipline technical alignment work, and that the absence of a deadline changes what the field prioritizes and how it is funded. The Cohere automation dataset published the same day is a concrete instance of the gap: an empirical measurement of where automation has actually landed, arriving well after the exposure-index literature it partially refutes.

RATL freezes a base forecaster, turns its historical forecast residuals into a train-only memory keyed on context, and at inference retrieves residual trajectories from similar past contexts under causal availability constraints. A set-aware router over forecast blocks and variables selects and combines them, with validation-chosen correction strength limiting residual over-injection. Using iTransformer as the frozen base, it improves accuracy in most settings on real-world multivariate benchmarks and transfers across backbones — the retrieved object is model-specific error rather than raw target values.

A PAC learning framework for general-sum concurrent stochastic games under transition uncertainty. The algorithm keeps data-driven L1 confidence sets over transition kernels, solves the robust game for a social-welfare-optimal epsilon-Nash equilibrium, and explores via a robust MDP objective to cover joint state-action space. A Nash margin characterization lets it either return an epsilon-NE within epsilon of optimal welfare or certify that no exact equilibrium exists. Under a minimum reachability condition the sample complexity is O-tilde(R_max^2 H^4 |S|^2 |A| / (p_reach eps^2)).

Sentinel-RL keeps topological reasoning out of the LLM entirely: a heterogeneous graph attention encoder compresses the live authentication subgraph into a fixed-size state, a PPO policy maps that state to a constrained action set, and the LLM is confined to narrating the policy's recommendations under a critic gate. On LANL cyber-security events, a two-phase CREATE ingestion loads a 24M-edge subgraph into Neo4j in 14.2 minutes, about 24x faster than the MERGE pipeline; PPO converges over 200 iterations to 0.91 precision and 0.87 recall on held-out red-team events, with a median 6.3 s detect-to-human-approval loop.

Extending bundle adjustment beyond points and lines to relations like coplanarity, parallelism and wireframes normally costs runtime and numerical stability. This work models higher-order groups as camera-like entities and expresses both group constraints and cross-feature (point-line) relations as 2D reprojection errors, which keeps the sparsity pattern of classical point-based BA intact under Schur elimination and avoids the direct 3D regularization that wrecks conditioning. On real and synthetic data, runtime matches point-only BA while recovering richer structure and better geometric accuracy.

A walkthrough of the Claude Fable 5.1 paper focused on the behaviors that did not make the headline coverage. The framing is useful as a counterweight to benchmark-first reporting: the model-card observations that get compressed out of launch coverage are frequently the ones that predict how a model behaves in deployment, which is the same lesson the Astra monitorability numbers deliver from a different direction.

VisCAD is a CAD-generation model suite whose core is VisCAD-M1, a 27B model mid-trained and post-trained for part-level design from renders, text, 2D drawings, and real photographs, targeting the generalization gap between narrow specialist CAD models and inconsistent frontier models. On PubCADBench and RealCADBench it averages 0.5540 at part level against 0.5496 for the strongest frontier model, and reusing it as a test-time verifier lifts that to 0.5797, roughly 5% relative. A separate domain harness drives frontier models for assembly generation, where mating relations and pose estimation are required.

Mike Pound works through the mechanism behind recent text-watermarking schemes and demonstrates his own implementation, including where detection degrades. It pairs directly with today's arXiv work on inference-speed watermarking and on embedding-space watermarks for auditing third-party retrieval systems, both of which attack the same throughput and robustness constraints he demonstrates.

The first round of the Corrigibility Research Fund closed applications on August 23rd and the grantees are now published. Corrigibility is the property directly at stake in one of the three prongs of this week's superintelligence bill, which names subverting shutdown commands as a qualifying dangerous capability, so the technical and legislative tracks are converging on the same definition from opposite directions.

DIFFINT structures an autoencoder bottleneck as soft axis-aligned interval memberships learned end to end from raw numerical features, so each latent unit is a readable hyper-rectangle and reconstruction error remains the anomaly score, with no discretization step. The paper proves a certified reconstruction-error lower bound outside the learned support under a Lipschitz decoder and derives a closed-form label-free importance ranking every unit-feature pair. On 48 ADBench datasets against 22 baselines it takes the best mean rank on both metrics (4.10 ROC-AUC, 4.16 AUPR) and is the only interpretable method in the tied leading cluster.

A pattern the author encountered repeatedly during outreach and lobbying: an argument stated at one level of abstraction gets rebutted at another, so both sides leave convinced the other missed the point. The relevance is practical rather than philosophical, since the same equivocation shows up in legislative debate, where a definition like the one in this week's superintelligence bill has to survive being read at whichever level of abstraction its opponents prefer.

A survey of what changes when agentic AI moves from pilot to deployment: agent-to-agent coordination, connection to systems of record, and the governance surface that appears once agents take actions rather than produce drafts. It reads as the practitioner-side companion to the Cohere measurement, which finds that outside a handful of software-mediated occupations the end-to-end tools simply are not built yet.

Multi-agent debate amplifies rather than corrects errors when most agents start out wrong, and existing fixes address peer influence while leaving each agent's biased concept prior in place. R2-MAD adds a memory of past debates, retrieving historical evidence with a policy conditioned on the current consensus level to recalibrate the prior, then using those retrievals to estimate per-agent reliability and reweight peer influence. Reported gains are consistent over single-agent and MAD baselines, though the paper gives no headline numbers in the abstract.

TAP-Path restructures a pretrained Virchow2 pathology encoder in place instead of distilling into a separate student: validation-driven selection physically removes 8 of 32 transformer blocks, input-adaptive pruning keeps 70% of patch tokens, and multi-depth feature recovery feeds a lightweight gated task head. Encoder parameters fall 24.96% (631.24M to 473.70M) and analytical compute 35.20% (340.13G to 220.40G FLOPs), while 32-class histopathology accuracy is 87.98% against 86.89% for full Virchow2 and 87.67% for UNI2-h. Calibration holds (Brier 0.1800, failure-detection AUROC 0.9047), with 91.22% on 433 frozen external CPTAC samples.

A feed-forward transformer for dense 3D shape correspondence that adaptively tokenizes meshes into patches using curvature guidance, then applies self- and cross-attention to learn patch-level and point-level relations between shape pairs. Trained only on BeCoS, a non-isometric partial-to-partial matching dataset, it generalizes to full-shape matching without retraining or fine-tuning. Across CP2P, PSMAL, BeCoS, FAUST, SCAPE and SHREC'19 it mostly beats prior partial and full matching methods on mean geodesic error and IoU, at sub-second inference, avoiding the cost and opacity of generative functional-map approaches.

Fusing several Bayesian networks into one structure trades dependency preservation against treewidth: unrestricted fusion keeps every dependency but inflates inference cost, while limited fusion prunes edges and risks overfitting input-specific noise. The proposed consensus framework prunes before fusion, prioritizing structure shared across the input networks under an explicit treewidth constraint, and searches with genetic algorithms using tailored initialization, operators and fitness. On synthetic and real-world networks it outperforms adapted prior methods and greedy baselines.

Two financial agencies, the National Credit Union Administration and the Federal Housing Finance Agency, currently have no authority to examine the technology providers their regulated entities rely on, a gap a new House bill would close. As credit unions and housing-finance entities adopt model-driven underwriting and fraud detection from third parties, examination authority over the vendor is the only mechanism through which a regulator can inspect the model at all.

RARF is a region-aware rectified flow for masked 3D data generation, submitted to the BraTS Inpainting Challenge 2026. Stochastic interpolation is confined to the inpainting region while observed voxels stay fixed as patient-specific anatomical context; a 3D network sees the voided volume with Gaussian noise in the hole plus the mask and timestep, trained with masked flow-matching and reconstruction-consistency losses. Under the BraTS protocol it produces competitive reconstructions that stay anatomically consistent with the surrounding tissue.

A post-processing stage for BraTS local-synthesis brain-MRI inpainting: ensemble the two 2025 co-first-place models, then train a lightweight residual refiner on the ensemble's own outputs with an L1 loss plus a weighted structural-similarity term. SSIM rises from 0.8767 to 0.8780 on a held-out reproduction of the official scorer and 0.8555 to 0.8572 on the validation leaderboard with MSE essentially unchanged, improving 62.6% of cases (signed-rank p = 2.2e-7). Adding a third ensemble member hurts and unsharp masking never helps, so the gain is learned rather than indiscriminate sharpening.

A cold-start study for industrial tool inspection where the only labeled imagery is manufacturer catalogue photography. Frozen off-the-shelf features fail to separate head shape and tooth profile; metric learning gets adjusted Rand index 0.94-0.97 for unsupervised clustering on catalogue images, but under half of that transfers to field photographs. The gains that do transfer come from reducing domain sensitivity rather than scale: grayscale conversion adds 0.22 and constraining retrieval with the known order sheet via Hungarian assignment adds 0.11.

Guidance for conditional generation normally leans on score functions, which are unavailable when the diffusion coefficient is singular and the conditional densities either fail to exist or are not smooth. This paper uses causal optimal transport, characterized through the predictable representation property of conditioned diffusions with a well-posed martingale problem, to build approximate loss functions that identify a minimum-entropy control for guidance under minimal regularity assumptions.

RobustSeiz standardizes CHB-MIT, TUSZ, Siena and SeizeIT1 into BIDS-EEG trees and sweeps environment, noise and adversarial transforms over predefined hyperparameter grids to stress-test subject-independent seizure detectors before deployment. Each run reports sample- and event-level sensitivity, precision, F1, false positives per 24 h, onset lead and lag, and Monte Carlo dropout predictive agreement, packaged as a Dockerized GPU pipeline with an experiment registry and full or research-subset modes. A demonstration on TUSZ shows how additive white Gaussian noise severity degrades detection quality, onset timing and predictive agreement together.

Restaurant owners reaching for generative models to refresh menus produce output that customers viscerally register as off. The mechanism is the familiar distributional one — sampling near the mode of a training distribution yields images that are individually plausible and collectively identical — and it is a useful reminder that diversity collapse shows up as a commercial problem long before it shows up as a benchmark number.

Numerical reasoning has been a gap in HTN planning. This extends the standard SAT encodings for totally-ordered HTN to SMT so that numeric fluents are handled directly, and contributes a benchmark suite for numerical TOHTN planning, which did not previously have a shared evaluation basis. The straightforward encoding is already competitive as a baseline, which mostly establishes a starting point for more expressive numeric HTN approaches rather than a strong result.

An Arc Institute research feature on anaerobic electron-transfer pathways that persist in modern human cells, framed around the observation that everything from single-celled organisms to human neurons runs on electron flow. It is basic biology rather than computational work, included because Arc's wet-lab program is the data source feeding its virtual-cell modeling effort.

For indivisible goods under identical additive valuations, the EF1-constrained Nash social welfare threshold problem is shown strongly NP-complete, inheriting hardness from NSW maximization. Beyond the known e^(-1/e) approximation that any EF1 allocation attains, uniform valuations make every EF1 allocation NSW-optimal and an epsilon-small-item condition gives a ratio approaching 1 - O(epsilon^2). For the harder requirement that EF1 hold after every assignment, PriorityNet trains a PPO policy with prospective EF1 action masking, reaching mean normalized NSW of 0.9911 offline and 0.9701 online over 3,000 instances each.

IRWOZ 2.0 rebuilds the industrial human-robot dialogue dataset whose first version carried enough noise in states and utterances to cap state-tracking accuracy. Mistral and Claude-3.5 generation plus manual correction and automated typo removal expand it to 390 dialogues across assembly, delivery, position and relocation domains. On dialogue state tracking, GPT-2's BLEU-4 rises from 0.1651 to 0.5604 relative to the original IRWOZ; the dataset is released on IEEE DataPort.

An evaluation of zero-shot and few-shot LLM inference for early chronic kidney disease screening, using clinically selected tabular features rendered into structured prompt templates so no task-specific training is needed. Against standard ML, DL, tabular foundation models, and existing screening tools, LLMs are competitive from a handful of examples and often win in the low-data regime, but results vary by model and degrade as input complexity grows, while the trained baselines improve steadily with more data. The framing is data efficiency versus stability, not replacement.

Instead of collapsing translation to one output, this study runs three decoding pathways as separate agents over a shared multilingual backbone (zero-shot, lightly fine-tuned dialect-stabilized, and English-pivot) and reads their divergence as a behavioral signal rather than error. On 5,000 Turkish-Syrian Arabic dialogue sentences, with stabilization trained on 5,000 further pairs from television dialogue and MADAR-Turk, lightweight tuning nearly doubles dialect marker frequency from 0.2266 to 0.4988 and cuts structural variance, while pivoting through English imposes normalization and compression and zero-shot shows the highest decision variance.

An extension of the cooperative multi-task semantic communication framework — a shared common encoder unit plus per-task specific units — from homogeneous classification on simple datasets to mixed classification and regression on Cityscapes. An InfoMax objective is adopted so the encoder handles discrete and continuous semantic variables together. Benchmarks cover independent single-task training, task-agnostic digital transmission and single-encoder multi-decoder SemCom, with an ablation on how common-unit capacity trades off across the joint tasks.

An audit of zero-shot species recognition over 10,321 images spanning seven Bangladeshi freshwater fish categories from two sources. BioCLIP2 reaches 72.36% with English common names on BFF-15 and 68.91% with scientific names on SylFishBD, against 25.15% and 14.40% for generic CLIP — but Bengali prompts sit at chance (14.22–14.29% balanced accuracy), with multilingual Jina CLIP v2 recovering only partially (21.89% and 16.36%). Masking interventions show a large white-mask artifact and strong species dependence, so headline zero-shot biology numbers confound visual knowledge with nomenclature, language and prompt formulation.

A curated set of 7,607 recipes - 3,807 suitable for diabetes and 3,800 not - tests whether LLMs can decompose a dish into ingredients and cooking methods, retrieve the relevant dietary guidelines and judge suitability. Three prompt regimes supply increasing guideline context: direct query, context-guided, and exemplary context. Models skew conservative, preferring to label recipes unsuitable to avoid harmful advice, and those that reason explicitly over the supplied guidelines do better; Mistral-7B and Llama 70B lead the models tested.

A position paper arguing that knowledge graphs state crisp assertions while the foundation models and agents consuming them reason in probabilities, which is why the two integrate as a data pipeline rather than a shared reasoning substrate. Semantic Bayesian World Models would instead hold beliefs over knowledge graphs, with ontological axioms constraining priors, observations updating by Bayesian conditioning, and actions modeled as interventions. The authors sketch worked cases such as a home-security agent judging courier versus burglar, and list the missing infrastructure: belief annotation over RDF 1.2, probabilistic entailment regimes, calibration layers, and protocols for agents to exchange and disagree over calibrated beliefs.

Rather than asking how many synthetic examples to generate, this study asks where they should sit in representation space. Using 410 annotated instances of English "look" from the BNC across four discourse-pragmatic functions, Llama 3.1 generates augmentations partitioned by cosine distance from real training data in RoBERTa embedding space, with augmentation quantity held constant across six conditions. All conditions beat the real-only baseline; near-boundary examples give the largest macro-F gain (0.113) while a distance-balanced mix gives the best accuracy (0.748). No condition improves AUC, indicating augmentation shifts the decision boundary without improving underlying probability estimates.

Filling missing typological feature values from URIEL+ and Glottolog via in-context learning, with the aim of producing justifications alongside predictions. Zero-shot prompting is insufficient, but supplying phylogenetic and geographic neighbour evidence in context puts LLMs substantially ahead of all baselines without disadvantaging low-resource languages. Most generated rationales are consistent with the evidence provided, which makes the predictions auditable rather than opaque.

A survey of work that treats rendered depictions of graphs as first-class model inputs rather than symbolic structures only. It organizes the area into three threads: vision for graph reasoning, where visual depictions support structure understanding and multi-step reasoning; vision for graph learning, where visual features complement graph encoders beyond known message-passing limits; and scientific graphs, where standardized drawing conventions such as molecular diagrams anchor both. The framing is what current vision and vision-language approaches can and cannot do relative to GNN pipelines.

Buildability simulation for 3D concrete printing usually models deposited layers as simple rectangles. This work couples ShapeGen3DCP, a learned filament shape predictor, with layer-activation FEM to build geometry-aware models directly from material and process parameters, avoiding experimental filament characterization and fluid-flow simulation. Filament geometry matters most for free-flow deposition and much less under layer-pressing strategies; an elliptical cross-section balances fidelity against meshing simplicity, and where rectangles are retained for regular meshes, sizing them by volume conservation is more reliable than calibrating to maximum width or interlayer contact width.

Session-level injury labels cannot be pushed down onto every minute of monitoring data without implying that onset time is known. The alternative here builds one representation per athlete-session at each fixed landmark (10, 20, 30 minutes) from information observed up to that point, keeping the prediction target at session level. On 2020 SoccerMon data - 3,743 athlete-sessions from 48 elite women's football athletes, of which only 22 are injury-associated - cumulative-plus-dynamic logistic regression reaches ROC-AUC 0.367-0.607 and PR-AUC 0.0080-0.0150 across landmarks with wide uncertainty, so discrimination is near chance at this class balance.

OBER+ extends a deployed outcome-attainment platform to cover what happens after a shortfall is detected: accumulate attainment across course deliveries, flag persistent shortfalls against cutoffs the regulator already uses, record the corrective action against a catalogue of evidence-annotated practices, log the change, and quantify subsequent movement. A continuity rule compares successive outcome statements so attainment is never read across a redefinition. On two live courses every outcome of a core course had been redefined between deliveries, and recomputation flagged six of ten figures differing beyond rounding, a defect since reported.

A pure analytic-number-theory note that lands in the cs.LG feed. It studies the ordered Bernoulli-word kernel p^k(1-p)^(n-k) and the geometry of its inverse-integer level sets, where the binary level fixes p=1/2 as the unique split-independent anchor and complex continuation produces the vertical critical-line pair. Critical-line zero ordinates decompose into integer shells plus a first-Bernoulli residual phase, which unique factorization resolves into prime-generator coordinates. The authors state explicitly that no proof of the Riemann Hypothesis is claimed, and no learning experiments appear.