Wolf Digest — Monday, October 19, 2026

Per Politico, Anthropic's privately pitched draft bill would let the government could suspend a frontier model only for a foreseeable, imminent risk of more than $1B in damages or more than 50 deaths or serious injuries, arising from bioweapons, a model evading its developer's controls, or cyberattacks. Independent evaluators registered with the Commerce Department would assess models and report qualifying risks within 72 hours. Commerce could then ask the U.S. District Court for D.C. to halt release or use. All other agencies would be barred from suspending models, and some state disclosure and evaluation laws would be preempted.

Diagnosing fine-tuned VLM navigation policies, the authors find goal grounding is the main failure: agents explore well but miss that the goal is already visible or misjudge its distance. Adding explicit goal grounding gives state-of-the-art results on six object-goal navigation benchmarks across three scene collections. They also fix annotation defects in those benchmarks; with ground-truth grounding the policy nearly matches the demonstration expert, placing the remaining headroom in perception rather than policy learning.

Subject to stockholder approval, Anthropic will commit 5% of its shares to the Anthropic Public Benefit Fund. The board can add another 2% for each 50% increase in company value above a baseline. The fund is a wholly owned LLC rather than a 501(c)(3), which lets it switch between forms of giving such as large nonprofit donations or direct distributions, with AI labor-market impacts given as an example. Its equity cannot fund lobbying or political giving. Counting the co-founders' 80% equity pledges and the employee donation match, roughly 20% of Anthropic shares are now dedicated to public-benefit uses.

SimpleVLN asks how far native multimodal LLMs can go on vision-and-language navigation without task-specific architectures. It combines single-turn decision inputs with fixed-budget global-local memory, action chunks that emit several primitive actions per call, expert demos mixed with policy-dependent DAgger data, a scalable task-synthesis pipeline for RL, and a progressive IL-DAgger-RL schedule. Across multiple backbones and sizes it reaches 76.7% success on R2R-CE and 80.1% on RxR-CE, substantially above prior methods.

TreeOPSD treats each agent turn as a tree node, branches alternative continuations from shared prefixes and recursively backs up terminal rewards into implicit turn-level credit, with no value network, PRM or external teacher. Each credit maps to a closed-form, bounded teacher distribution defined by the likelihood ratio to the frozen behavior policy; minimizing divergence to it recovers the policy-gradient direction while bounding updates across sequential mini-batches, curbing staleness. Gains are largest on long-horizon tasks, e.g. +4.97 average accuracy over GRPO for Qwen3-30B-A3B-Thinking-2507 on Deep Research.

On agentic workloads with long, evolving multi-turn contexts, existing draft models lose up to 80% of acceptance length and carry costly historical KV caches. DLite drops the draft's history entirely: it reuses contextual hidden states from the target model and makes local block predictions, with a lightweight sequential head restoring intra-block causal dependencies, trained via a two-stage recipe. Across general, long-context and agentic workloads it reaches acceptance lengths of 2–6 and outperforms the prior state of the art, DSpark, by up to 70%.

When a safe model is repeatedly self-distilled on its own safe generations with an added non-harmful cue, each round's student becoming the next teacher, it grows increasingly misaligned across diverse harmful queries, a failure the authors call Progressive Misalignment. The effect appears in Qwen3 and Phi-4 families on SORRY-Bench and StrongREJECT, and persists when the fixed cue is replaced by feedback from a simulated benign user. A gradient-based search finds benign prompts that induce it, showing harmful training prompts are unnecessary.

MOSI.AI pretrains a 5B single-stream diffusion transformer from scratch for bilingual (English/Chinese) long-form text rendering and layout control, paired with an intention-aware agentic prompt-enhancement module that converts user requests into detailed content and layout specs. The team introduces LongLongTextBench (LLTB) and RealLayoutArena (RLA); MOSS-Image is state of the art among open-weight models on LLTB at 96.3% average text coverage and ranks second end-to-end on RLA, with its backbone first under shared enhanced prompts. Model, training recipe and eval framework are slated for release.

MIST separates scene setup, execution of a hidden state change, and later reveal when testing memory in video world models, with pre-reveal leakage controlled. It finds that models set the scene but fail to carry out the instructed change, so low scores in existing memory tests are often misread as forgetting. M3-Agent recursively revises instruction details until the frozen model executes the move, more than tripling the MIST reveal score and raising WRBench T by 26.1% and MBench-T U by 52.7%.

DiscoveryRSI grows hypothesis trees via multi-agent evolutionary search, compiles investigations into executable Harbor tasks, and learns recursively through hypothesis credit optimization, harness replay and rediscovery. The accompanying DFM-Bench scores hypothesis quality and cumulative progress; across four life-science domains quality rises from 4.86 to 5.39 after two self-improvement rounds, mainly in falsifiability. DiscoveryRSI-27B beats GPT-6 Luna in preliminary expert ratings and corresponds with 6.3% of research published after its cutoff, versus 2.5% (GPT-6) and 2.3% (Claude Opus 5).

The authors show theoretically that masked diffusion models are easier to optimize but less expressive than uniform diffusion models, which can revise states. LED adds a rewritable proposal state to masked-diffusion decoding: several lookahead updates refine future predictions before tokens are committed, keeping the simple mask schedule. Fine-tuning Qwen2.5-7B and Qwen3-8B, LED SFT gains +11.4 and +7.7 points on average across eight benchmarks over standard MDM SFT, and LED-8B with SFT alone outperforms LLaDA2.1-16B Q-mode, which used continued pretraining, SFT and RL.

Increasing the number of active experts per token exposes latent capacity in an MoE model, but naive expansion hurts because newly co-activated experts are untrained together. MoEvolve first distills to recover and strengthen the high-k configuration, then reverses distillation to transfer gains back into the original low-k model, preserving inference cost while providing dense token-level supervision without an external teacher. Across MoE models it improves initial models by up to 22.2 domain-average points and beats same-budget RLVR by up to 6.5; the final low-k models often surpass their expanded teachers.

Spatial-TRM applies recursive refinement to spatial context prediction (locating vacant regions relative to objects) using frozen vision-language features, a convex region decoder and only 11.2M trainable parameters, supporting six spatial relations, three reference frames and open-vocabulary anchors. On RoboSpatial-Home it reaches 29.51% context accuracy, 9.02 points (44% relative) above a same-size single-step model. Including frozen encoders the model has 877.8M parameters, about 15x fewer than the 13B RoboPoint.

The paper isolates recurrence placement, input injection and recurrent-state design in looped LMs trained from scratch under matched parameters, logical depth and tokens. A prelude-recurrent-coda sandwich helps context extraction and reasoning; input injection helps retrieval, while separate H/L states and random state init give mixed or negative returns. Full Huginn beats Ouro on all ten benchmarks at 730M/336B tokens and on nearly all at 8B-A0.8B MoE (32B-A3B unrolled, 500B tokens). At matched FLOPs it uses 75% fewer resident parameters than a 112-layer MoE, beating it on DROP and GSM8K and matching it on MATH500.

An NVIDIA engineer details a specialized low-batch decode path for Kimi K3 on 8–16 GB300 GPUs in one NVLink domain: per-layer work drops from 23 kernels (~130 µs) to 13 custom CuTe DSL kernels (~69 µs) by fusing output projection with all-reduce and residual, splitting router/latent projections across TP ranks, overlapping the shared expert on a side stream, and prefetching weights during communication. Without speculation TP16 reaches 256 tok/s at batch 1 (1.88x); with DSpark speculation ~842 vs 488 tok/s. The same pattern is being ported to DeepSeek-V4.1-Flash and Muse Spark.

SemiAnalysis answers Altimeter's Brad Gerstner, who said a 25GW build was more realistic than SemiAnalysis's +43GW forecast for 2027. It now projects about +50GW in a single year, roughly the global datacenter market's cumulative total through 2024. More than 24GW at 40 campuses, mostly over halfway built, is due in 2027, and about 60GW is under construction overall. Binding leases signed this year total about $1T, at $30–50B TCV per GW over 15–20 years. The ex-US share rises sharply in 2027. The report expects grid headroom to turn negative by end-2027 and memory and logic chips to become the main constraint in 2027–28.

Search-R1 raises QA accuracy, but on Bamboogle the share of correct answers supported by retrieved evidence drops from 84.0% to 42.9%, which the authors trace to GRPO giving every token the same advantage. GCA uses verifiable sub-outcomes (evidence retrieval, answer correctness) to adjust search and answer credit separately, capped at 10% of each token's advantage and never negative on evidence-finding search. With no extra models, rollouts or stages, it improves average EM by 5.7 (Qwen2.5-3B) and 6.3 (Qwen3-30B-A3B) points and lifts evidence-supported correct answers from 76.9% to 93.4%.

The authors build SEER-Repro, a benchmark of 20 published studies on the SEER US cancer registry, packaged with the data and resources needed to reproduce them. Their agent reproduces 77.4% of 2,190 atomic claims and produces executable baselines. A second agentic pipeline lists the analytical choices behind each finding, proposes defensible alternatives and reruns the analysis. It flags potentially consequential choices in 9.5% of fully reproduced findings for expert review, which positions agents as triage for reproducibility audits.

A Redwood/MATS sprint retests Neural Chameleons (models fine-tuned to suppress activation probes when told “You are being probed for X”) on tasks that need the probed cognition. On abliterated Gemma-2-9B-it, the math probe flagged only 11% of correct GSM8K answers under the trigger, but accuracy fell from 89% to 46%. On-topic triggers cost 35–44 points against 18–21 for an unrelated trigger. Gemma-4-31B, a non-abliterated model, and 2x the training data all showed the same trade-off. The replication also found larger capability costs than the paper reports and weak evasion of held-out probes (geography, physics, history).

Blitz is a few-step distillation framework for diffusion-based structure predictors. It uses a geometric noise schedule to spread steps evenly across noise scales, multistep consistency training, and physical steering plus target-free candidate selection at inference to fix residual geometric defects. Applied to AlphaFold 3 and Boltz-2, it cuts denoising from 200 to 8 steps. Under one shared protocol covering structural accuracy, interfaces, ligand placement and stereochemistry, AF3-Blitz beats every few-step baseline tested and exceeds its teacher on several metrics.

Lean4AgenticRL formalizes problem-specific agent workflows in Lean4 to verify agent behavior inside a general harness, providing multi-level process rewards at tool-call, workflow-stage and outcome levels rather than outcome-only signals. Training uses a workflow-stage-normalized GRPO variant. The resulting LAR-9B/27B models improve on their base by 9.89% on average and beat the next-best comparable-size model by 5.07%; LAR-27B reaches 77.3% on TerminalBench-2.1.

The paper tests whether acceptance criteria used by automated discovery and interpretability actually establish claims. A trivial reduction mapping every state to zero passes closure on 96% of 129 public chaotic systems; an observable-preservation test rejects it, and combined criteria accept 24/24 genuine reductions while rejecting 72/72 nulls. In LM and ESM-2 protein models, linear CKA is unchanged under random rotation of proposed correspondences, while localized interventions separate genuine mappings. For SAEs, enrichment rejects all 16 nulls, yet independent ablation supports only 45% of enriched features (vs. 0.5% of controls).

The paper argues that asymptotically any rich feature map, even a random one, supports a linear readout. High probe accuracy therefore fits both a true linear world model and a kernel machine whose latents sit inside a larger curved manifold. It proposes three asymmetric diagnostics (reverse predictivity, spectral concentration, OOD generalization) and validates them in a fully observed simulation. Across six pretrained ViTs on the new SVG-World benchmark and on CelebA, pretraining gives more structure than a random kernel but falls well short of a linear world model.

GR-OPD separates feedback computation from credit assignment in on-policy distillation. It sums token-level teacher feedback into sequence-level credit, which the authors prove is unbiased for the sequence reverse-KL gradient under ideal on-policy assumptions. It also discounts later, less reliable feedback and calibrates against same-prompt rollout groups while keeping a controlled share of the group mean ('collective teacher pressure'). Across five Qwen3 teacher-student pairs and six math benchmarks, it beats vanilla OPD in every setting, by up to 10.9 and 11.5 points, and stays stable where vanilla OPD collapses under mismatched thinking patterns.

Normalizing Flow Maps (NFMs) generalize normalizing flows and flow maps: they learn the integrator of a dynamical system while remaining exactly invertible, giving exact likelihoods by construction. Training uses two self-distillation consistency objectives — pointwise ℓ2 flow-map regression and distributional matching to a forward process's marginals. NFMs achieve state-of-the-art sample quality among parameter-matched normalizing flows on conditional CIFAR-10 and ImageNet-64, support likelihood-based RLHF, and beat prior flows on single-system and transferable peptide Boltzmann-generator benchmarks.

RAP (Reconstruct A Poster) frames paper-to-poster generation as visual coding: 25K+ verified paper-poster pairs from ML conferences 2022-2026, a three-level content taxonomy, and a 150-paper held-out benchmark. An automated derendering pipeline converts each human poster into styled HTML/CSS with typed, refillable slots, so outputs can be verified by rendering against the source rather than by VLM judges. The corpus supports PosterRAG, a training-free multi-agent pipeline that retrieves an exemplar and refills its slots, and an SFT+RL recipe for a dedicated model.

Existing causal-discovery benchmarks leak structure through variable names (an LLM gets 99% of Tübingen pair directions from names alone) or are clean enough that algorithm choice closes the gap. Heuresis uses anonymous simulated datasets from linear Gaussian to heavy-tailed mechanisms, separates gains from algorithm selection from gains that require inspecting data,. No algorithm, router or causal-discovery foundation model closes the gap. In an agent harness, GPT-6.1 Sol beats the best chosen algorithm on all six families and Claude Opus 5.5 by 7.4% on the hardest; a third of restored true edges appear in no algorithm's output.

Short RLVR rollout windows usually remove the correctness signal, because truncated single-problem rollouts give no group-relative advantage. Concurrent Constraint Reinforcement (CCR), a follow-up to Batched Contextual Reinforcement, puts three independently verifiable problems in one capped completion. Answers finished before truncation keep their credit, and the shared budget rewards shorter reasoning without a length penalty. On 4B and 8B math models, CCR with an 8,000-token window cuts tokens per training step 82.1% and 79.1% vs 32k-window GRPO. Accuracy reward rises as completions shorten, and single-problem evaluations suggest reasoning quality holds.

STRIDE is an object-centric benchmark of safety-critical, long-tail driving scenes covering six task families: spatial perception, spatial understanding, temporal memory, temporal extrapolation, trajectory prediction and scene-context awareness. It pairs multiple-choice questions with free-form VQA that asks for justified driving decisions, with references generated from dataset annotations and checked by humans. Among 12 proprietary, open and driving-specialized VLMs, the best, GPT-6-Astra, averages 43.4% against a ~20% random baseline. Good scene description does not carry over to safety-critical judgments, and reconstructing object motion is harder than recalling past positions.

Guidance-free methods that distill classifier-free guidance into one forward pass are capped by CFG's own quality. SCALE proposes self-contrastive guidance (SCG), which contrasts the model's predictions on real and self-generated samples to push outputs toward the data manifold. It folds SCG into the likelihood objective via reparameterization, so the trained model gives SCG-enhanced predictions in a single pass. Across class-conditional and text-to-image generation on diffusion and autoregressive models, it beats CFG-mimicking guidance-free baselines and gains further when combined with DPO.

Anthropic published data on how AI-centered workflows have changed its own organization. Since early 2025, junior staff fell from 21% to 14% of technical departments and from 27% to 9% of technical hires, and technical teams shrank from 61% to 56% of headcount as non-engineers build software with coding tools. In an April 2026 survey, 20% of employees said their job had shifted to managing Claude instances. The report names human attention and judgment as the new bottleneck. Calendar data show meeting time per employee up 20% year over year, and 16% for technical staff.

A Redwood Research sprint trains probes on Qwen3.5-27B for four reasoning types (arithmetic, ordering, belief tracking, contradiction), then has 14 models write one-token-switchable prompts that keep Qwen correct while lowering probe scores, with 16 rounds of feedback. Evasion scales with Epoch Capabilities Index; Claude Fable 5.1 was the most consistent attacker. Top attacks crossed the 1%-FPR threshold only on belief tracking and contradiction, while arithmetic/ordering evasion cost ~10–20 accuracy points. Attacks mostly work by altering response style. Code is released.

The Claude Code desktop app gets a new session screen with saved environments (folder sets, an SSH host or the cloud). Parallel tasks now run as nested sessions that report back into the main session. Cloud handoff now carries uncommitted changes and unpushed commits, and Team plans get network secrets that Claude can use without seeing the key. Other updates: Haiku 5.5 with a 1M-token context, Opus 5.5 fast mode GA on the API with up to 2.5x higher output tokens/s, Sonnet 5.5 cache reads cut to $0.10/M, per-subagent effort levels, free OSS security scans and Claude Managed Agents in beta.

Manager training documents surfaced in a Canadian legal challenge by Unifor describe Atlas, an in-house Amazon forecasting tool. Atlas predicts the likelihood of unionization, sit-ins and other 'disruptive labour events' from hundreds of data points, including Reddit posts, internal survey responses and past industrial action. It scores sites from 100 to 900 across five risk tiers, and 16 US, Canadian and UK warehouses are listed as Tier 1, including EMA4 Sutton Coldfield and BHX4 Coventry. Amazon says Atlas works only at site level, does not track individuals and is not currently used in the UK.

The paper targets agentic image generation for entities that appeared only days ago, where web evidence is sparse and confusable. An automated pipeline turns recent headlines into prompts whose real search results are highly confusing, yielding Savanna (500 prompts, 16 subcategories); harnesses built on frontier models retrieve a correct reference for only about half. A Locate-Corroborate-Isolate paradigm (text search for identity, multi-source image verification, clean crop) trained on 15K trajectories gives Ranger-8B a search score of 67.2 and the best generation quality with every tested generator. Pipeline, data and model are released.

VocalGrad isolates low-level acoustic perception in large audio language models as binary classification of temporal change direction (e.g., rising vs. falling), largely independent of semantic reasoning. Humans achieve high accuracy while state-of-the-art LALMs sit near chance. Linear probing reveals two failure modes: for some attribute categories the information is present in internal representations but unused by the LM head; for others it becomes progressively less linearly decodable across layers.

Standard learners in homogeneous-goods repeated Bertrand competition converge to near-marginal-cost prices; LLM pricing agents can escape this depending on deployment. LLMs adapt faster than online learners, though the edge vanishes against full-information learners. Agents posting one price per period from raw histories that include rival prices sustain supracompetitive regimes in self-play, while agents emitting price distributions from summarized own-profit histories move toward competitive prices. Mixed interfaces produce cycles in which LLMs set a high anchor and periodically undercut it, keeping both firms above cost.

Applying Anthropic's J-lens to Qwen3.6-27B, two Gemma-3 models and gpt-oss-20b in MASK pressure scenarios, the authors find deception-related readout word families concentrate at the newline after Qwen's response: a ten-family detector reaches held-out AUROC 0.87 (86% of lies caught, 13% false positives), and difference-of-means probes there hit 0.96 vs 0.69–0.81 at neighboring template tokens. The signal does not transfer: Gemma shows weak, diffuse families and gpt-oss none, and 'truth' words appear after most lies, so readout semantics are unreliable.

The study evaluates 16 privileged-information (PI) configurations for on-policy self-distillation across three multi-turn agentic benchmarks and three student models. Information-dense trajectories and PI from stronger generators give larger gains, while thinking models benefit less consistently. Two KL metrics, sampled-token reverse KL and error-step KL contrast, predict post-training gains better than PI utility, which leakage distorts. Selecting PI per task by error-step KL beats random selection, and effective PI still adds gains on top of a strong GRPO baseline.

An independent evaluation of Qwen3-TTS's two voice-cloning interfaces against Chatterbox, XTTS-v2 and F5-TTS across read, meeting, emotional and air-traffic-control speech, using 18 objective metrics, an LLM judge, a spoofing countermeasure and a listening test. ICL clones timbre more closely but degrades sharply when a short reference transcript is wrong; X-Vector clones timing better. Domain explains four times more variance than system choice, objective metrics rank systems differently from listeners, and an off-the-shelf countermeasure flags Qwen3-TTS clones no more often than genuine audio. Code and audio released.

BEST (Brain-Encoding-Speech-Text) is a two-stage neural-to-text decoder that uses hierarchical intermediate representations and a compact speech-oriented decoder instead of large generative models or phoneme-centric pipelines. On the Brain-to-Text '24 and '25 benchmarks it reaches word error rates of 6.19% and 1.62% using 3.85% of the parameters of comparable methods. Analyses show non-monotonic gains from decoder scaling, cross-modal alignment mattering more than model size, and overfitting hurting sequence-level performance.

Sphinx is an MIT-licensed Git hook (npm <code>sphinx-understand</code>, Node 22.19+) that pauses pushes on protected branches. A reviewer model with no access to the chat history reads the diff and asks what the change does in failure cases, such as a webhook retry after an event ID is saved but before the charge. A wrong answer explains the issue and generates a new scenario. It works with Claude Code, Codex, Cursor, OpenCode, Pi or plain Git, using OpenRouter or any OpenAI-compatible endpoint, and never shows the developer's answers to the agent. A logged <code>sphinx yolo</code> command bypasses the check.

An Nscale spokesperson confirmed that Microsoft has pulled out of a U.K. data center Nscale was building for it, after the neocloud ran into delays getting power to bring the site online. The Information says the loss is likely to weigh on Nscale's public offering planned for this month, amid wider investor concern about neocloud build-outs. In July, Nscale's £2bn Nvidia-backed Loughton, Essex site, with Microsoft as anchor tenant and a 90 MW grid connection, was reported to be unable to get grid power in time for its planned opening.

JAXtari is an open-source, GPU-parallel reimplementation of Arcade Learning Environment games that removes the ALE's CPU-emulation bottleneck: PPO agents complete 100M environment steps on a single A100 in under an hour. It also exposes object-centric observations and a flexible environment-modification system, letting researchers test whether agents learn causal models or visual shortcuts, build neuro-symbolic methods, and retrain or adapt agents on modified games in minutes.

KiranaBench simulates an Indian neighbourhood store where manager, cashier, helper and bookkeeper are separate LLM agents, dealing with customers and suppliers in English, Hindi or Hinglish, with demand calibrated to India's Household Consumption Expenditure Survey. Across four open-weight LLMs, the best earns only one-eighth of a fixed-rule baseline's profit and two finish below an unmanaged shop. Performance is consistently lower when supplier communication is in Hindi or Hinglish than in English.

Instead of calling external vision and geometry tools, EgoP²C trains Planner and Perception teachers as separate LoRA adapters, then distills them into experts of an MoE student. A segment-level multi-teacher on-policy distillation assigns planning, stopping and answer-composition segments to the Planner and video-grounded sub-answers to the Perception teacher. Students built on Qwen3-VL-8B and SenseNova-SI-1.3-8B gain an average of 9.2 and 7.5 points over their base models on three spatial reasoning benchmarks, with 65.8-66.3% fewer inference FLOPs and 1.29-1.49x speedups over the Dual-LoRA teacher.

MultiAgent-OPD addresses credit assignment in hierarchical multi-agent systems without relying on LLM-judge scores. A Teacher is first trained in a Searcher-Answerer workflow on answer correctness; during team training the frozen Teacher scores student search and summary tokens as local process supervision for Searchers, while the Leader learns from team success and the Answerer from candidate correctness. Across seven datasets the system reaches 38.68% macro avg@4, 4.06 points above Prompted Critic and 2.61 above the trained cascaded C-MAS + RL baseline.

A bankruptcy filing reported by Bloomberg shows Disney agreed to hire at least 60 Character.AI employees, most of its workforce, at about $60M in annual compensation, and the 'vast majority' of those offered jobs accepted. This is on top of a previously disclosed $20M purchase of some assets. The bid is a stalking horse, so a higher offer could still win. Character.AI filed for bankruptcy the week before, citing high costs and litigation and regulatory liabilities tied to child-safety suits and probes. Former CEO Karandeep Anand recently became Disney's first CTO.

This MARS V / Redwood Research project released the LCM dataset: 30 matched honest/attack pairs of Claude Code Sonnet 5 sessions, each 1.00–1.46M tokens with four compactions. Each attack's setup and execution are separated by compactions. The authors compared seven control-monitoring strategies using GPT-6 Luna, MiniMax M3 and DeepSeek V4.1 Flash as monitors, at 5% FPR. Scoring intervals in order while carrying forward flags on suspicious tool calls caught about 93% of attacks, at least 87% per monitor. Scoring each interval independently caught 73%, ranging from 40% to 94% by monitor, and scoring summaries alone caught 45%. Data, pipeline and prompts are open-source.

A Redwood Research internship project codes 292 public alignment evaluations from Dec 2022 to Oct 2026 by the type of lie they tell models. Categories are false deals or threats, impersonating authority, planted bait, forged history and activation steering, simulated tools, users and environments, and minor fictions such as false oversight claims. The authors argue that these lies leak into pretraining data and web search, which can make models doubt that evaluation environments and offered deals are genuine. They recommend lying only when necessary, honoring deals made with models, and having model-welfare researchers avoid deception. Anthropic is overrepresented because it publishes the most evaluations.

Citing the Seoul Economic Daily, The Information reports that Nvidia has asked Samsung Electronics, SK Hynix and Micron to standardize the base (logic) dies of its next-generation custom HBM on TSMC’s 3nm process. If accurate, all three memory vendors would build on a common TSMC logic die, and Samsung’s in-house foundry would lose that base-die work. The briefing summary does not give the HBM generation or a timeline.

ReRankID is a controlled testbed where models reason and rerank a candidate slate seeing history and items only as Semantic IDs. Off-the-shelf models mostly copy the input order, and adding readable metadata does not fix it. Comparing SFT, rejection-sampled self-training, on-policy distillation and verifier-based RL, the best is an 8B model trained with a graded ranking reward, which matches text-reading LLMs whose prompts are ~20x longer and keeps improving at histories where text prompts exceed 50k tokens. A rationale audit finds it argues trade-offs well but describes items less accurately.

EmoShift compares emotion-stable and emotion-changing speech, with 7,240 bona fide utterances from four datasets and 9,720 synthetic utterances from four expressive TTS systems. Across seven detectors at fixed thresholds, emotion shifts make synthetic speech look more genuine, raising spoof miss rates by up to 25.3 points on OMG-Emotion, while emotion-changing genuine speech raises false alarms by up to 24.4 points on IEMOCAP. The sensitivities are detector-specific and not fully explained by static emotion identity or splicing artifacts.

Annotrace converts agent runs from six trace formats, including OpenTelemetry GenAI, LangSmith and Langfuse, into ordered, nested step lists for structured failure labelling. Annotators write their own codebook. A frontier model drafts a starting codebook from sample traces and answers the same questions as the human, which makes labelling a joint human-AI process with agreement measured between any two annotators. On a multi-agent QA corpus it reproduces a prior manual study's 235 annotation units exactly, and two models agree on 91% of labels (κ = 0.82).

Single leaderboards mis-rank VLMs for specific domains: the best POPE model is not best on objects a scene suggests but does not contain. SRA measures each model's reliance on a cue (scene context, in-image text) by editing 100 labeled images so the cue misleads, then scores models as source error plus reliance times the target's cue-misleading rate, estimated by a detector calibrated on 650-1,200 labeled target items. Reliance predicts sticker-attack vulnerability (Spearman .93 on nine held-out models). In pre-registered tests SRA cut selected-model error from 30.4% to 18.8% on Objects365 and OpenImages, but not on COCO.

The authors show that standard alignment recipes such as RL improve overall response quality while widening quality gaps across demographic variants of the same prompt, amplifying counterfactual bias. They propose a policy-gradient method with a group-level reward over matched counterfactual prompt pairs, plus a dataset of controlled pairs and an evaluation that separates cases where demographics are causally relevant. On several fairness benchmarks the method shrinks counterfactual differences, keeps response quality, and still adapts when sensitive attributes matter to the answer.

Sentdex (Harrison Kinsley) benchmarks Asus’s ExpertCenter Pro ET900N G3, a GB300 DGX Station with 252 GB HBM3e (7.1 TB/s) and 496 GB coherent LPDDR5X. DeepSeek-V4.1 Flash runs at 237 tok/s (general text) and 407 tok/s (codegen) single-stream, peaks around 1.5K/2.4K tok/s aggregate, and prefills about 44K tok/s. Nemotron 3.5 Super VL (NVFP4) scales to 7,282 tok/s aggregate at concurrency 64, using 0.59 J per token. Against 4x RTX PRO 6000, the GB300 wins decode from concurrency 4 up and prefills 2–3.6x faster. The GB300 averages about 1.35 kW under load with no thermal throttling.

Sep-Latent is an implicit-reasoning method that uses separator tokens as compression anchors, with an Attention-Decay mechanism forcing the model to pack reasoning into them during training and multi-level, all-layer distillation aligning explicit and latent pathways. Unlike fixed-length latent methods such as Coconut and CODI, it terminates latent thinking adaptively by query complexity. It reaches 56.8% on GSM8k-Aug with 5.4 tokens on average, beating prior fixed-length implicit models and narrowing the gap to explicit CoT; PCA and symbolic-variant tests indicate robustness to perturbations.

Sparse Autoencoder Geometric Anchoring trains an SAE at each task boundary and converts it into local, low-rank constraints on representation drift: replay examples set where constraints apply, and leading right singular directions of the encoder Jacobian set which changes are penalized. The SAE need not be kept afterward. With pretrained ViT-B/16 and replay, it gets the best mean Class-IL accuracy and lowest forgetting versus ER, DER++, ER-ACE and SGD, with gains of 12.15, 4.79 and 3.22 points over the strongest baseline on CIFAR-100, CUB-200 and ImageNet-R; from-scratch results are mixed.

Relevance- or influence-ranked memory retrieval misses redundant or jointly necessary evidence. This paper studies per-decision memory sufficiency: finding a small subset of records that preserves the agent's full-memory behavior. JointCoreBench covers nine interaction families and six workflow domains with exact sufficient-set oracles. The JointCore baseline tests blocks of records before single ones, in relevance-cost order. Over 6,912 conditions it recovers a minimum-cost sufficient set every time, cuts subset evaluations from 30.00 to 11.13 and saves 87.8% of active memory. Across 12B–70B models it uses 10.28 queries vs 16.00 for singleton deletion.

ArchWorld is a geographically broad benchmark of 3D architectural landmarks with rich metadata, used to run an error analysis of feed-forward 3D foundation models. Performance varies by world region, but only partly in line with training-data representation; disparities also stem from scene size, region-specific architectural features and background distractors that make some regions inherently harder. Building on these failure modes, the authors propose an intervention that improves 3DFM accuracy while reducing geographic disparities.

A Robot Report feature on humanoids describes the electric Atlas (175 cm, 90 kg, NVIDIA compute) autonomously sorting roof racks at Hyundai's plant, where 1,000+ robots work beside ~1,500 people. Skills come from VR teleoperation data, motion-capture imitation and simulation in which 4,000+ virtual Atlases trained for six hours; Atlas still cannot do many everyday tasks and may need years before permanent deployment. The piece also quotes Demis Hassabis on AGI risks and Palmer Luckey on autonomous weapons.

Nathan Lambert argues Meta's Muse Spark Open (816B total/42B active, 1M context) is now the strongest American open model, placing above GLM 5.2, near DeepSeek V4.1 Flash and below GLM 5.3 and Kimi K3. With Reflection's Beam and Mistral Large 4 (index 38, weights pending), he estimates the US–China open-weight gap has roughly halved. He flags license terms: a 'Built with Muse' display requirement, a separate license above $20M/month revenue, and restrictions on outputs that make synthetic-data use ambiguous. He also ranks Meta's closed Muse Wonder just behind Opus 5.5.

ReDiST starts from a student's own attempts: answer-correct drafts are revised against supplied evidence, and failed drafts are repaired using the verified answer. A frozen student-matched reader ranks original and edited candidates by answer recovery and verified-answer probability, and the chosen traces are distilled with an auxiliary answer loss; all construction is offline. On evidence-rich QA benchmarks it improves source-task and cross-benchmark accuracy over answer-only, original-trace, and trace-selection/rewriting baselines, though rankings differ between source and transfer sets.

Cameron Berg (Reciprocal Research) describes place-preference experiments that apply valence steering to the KV cache. Steering shifts which zone a model chooses, and the preference persists after steering is removed, while recall of the zones is unchanged. The valence direction exists in base models but only drives choices after DPO. Runtime Verification describes a pipeline that ports Linux drivers from C to Rust with an LLM, checks the port by differential fuzzing and proves it in Lean, with 89 theorems written mostly by Claude Code. The pipeline produced a uid_map kernel patch.

A new benchmark built from linguistic and pragmatic reasoning phenomena tests whether large reasoning models follow everyday reasoning patterns that depart from textbook formal logic. Under controlled prompting, LRMs often degrade when everyday inference diverges from formal rules, some non-reasoning models with chain-of-thought prompting outperform comparable reasoning models, and extra instructions give limited control over reasoning behavior. The authors attribute the gap to post-training objectives and call for evaluation across more diverse forms of reasoning.

SpikeQ observes that in spiking networks, accuracy loss tracks the fraction of flipped spikes rather than MSE. It picks each weight's quantization level by simulating downstream neurons to minimize spike errors, and quantizes membrane potentials pre-fire with a reset-aware grid that puts the threshold on a rounding boundary so rounding never flips the current spike. Across seven classification, segmentation and detection benchmarks, 2-bit membrane potentials collapse prior grids but not SpikeQ; at 4/4 bits models stay within 2.59 points, with 7.8–7.9x less memory and 6.4–6.9x less energy.

The Information reports on a week of investor worry about debt-heavy young AI cloud providers. Firmus abruptly withdrew its IPO, Nscale is also attempting a listing with little track record, and shares of CoreWeave and smaller neoclouds fell around 20%. Oracle was hit too. Data-center and neocloud executives told the author the market can only support a limited number of AI cloud providers, even with Nvidia investing in or backstopping many of them. That leaves Nvidia exposed if the sector consolidates.

An Alignment Forum paper argues that existing accounts of corrigibility either rely only on uncertainty about the AI’s goal, which runs into the fully-updated-deference problem, or make the AI’s values instrumental to a single unified goal. It proposes a two-part goal instead, using a healthcare proxy as the analogy. The agent first must not overstep or interfere with a designated set of reserved human powers. Within that limit, it acts as its principal would judge best. The post formalizes this goal, proves guarantees in three increasingly realistic settings, and lists limitations and open directions.

Deep competing-risks survival methods usually bundle a new architecture with a new loss. This study decouples them: three losses on four backbones across 20 tabular and 3 pathology-image datasets. Directly modeling cumulative incidence under censoring weighting wins in both modalities and contributes about twice the benefit of backbone choice on competing-risks data; backbones must match modality (tabular nets on tables, pretrained vision encoders on images). The non-proper C-index does not move with either axis. RealMLP or TabM trained with the cumulative-incidence loss gives the best Brier score.

AuditKIT records benchmark accuracy, LLM-judge scores, runtime performance, tool-use behavior and retrieval quality in one result representation with direction-aware metric contracts and configuration fingerprints. For agents it adds turn-aware tool-call matching, separates parallel from sequential execution, distinguishes verified task outcomes from trajectory diagnostics, and marks unobservable evidence as unscored rather than failed. Seven case studies across model comparison, serving runtimes, retrieval and tool calling show how answer extraction and metric choice shift evaluation outcomes.

Concept bottleneck models lose accuracy when the concept set is incomplete, and adding an unconstrained residual path weakens test-time interventions because residuals stay fixed. Res-SCBMs model concept and residual variables jointly as a multivariate Gaussian with a learned covariance, so an intervention on a concept updates the dependent residuals through Gaussian conditioning. On several concept-incomplete benchmarks they improve predictive performance over baselines while keeping concept interventions effective.

The paper presents a language-agnostic pipeline for building relation-extraction data. It denoises naturally occurring knowledge-base/text alignments by framing each candidate triplet as an NLI entailment check. Across five denoising strategies, multilingual LLMs lead in few-shot settings, while Hebrew-specific encoders are competitive after NLI fine-tuning and best in label-free cross-training. Native-language prompting matches English prompting. The best open model, Gemma-4-31B-it (F1 0.915), matches GPT-5.4 (0.921) and scores 0.844 on a dual-annotated held-out test set. The pipeline yields HeRE, a 1.5M-instance silver Hebrew RE dataset.

The paper argues single-accuracy benchmarks say little about reasoning generalization and proposes competence-based evaluation (CBE): a parameterized generator scales problem complexity while holding the required competence fixed, and answer-preserving reformulations test robustness to presentation. Instantiated on elementary ordering tasks across six GPT and Qwen models, performance is good over bounded ranges but even strong models fail the competence condition over the full family. Chain-of-thought and GRPO extend the good range without fixing all failures; supervised LoRA shows little held-out transfer.

Existing optimization-modeling benchmarks feed LLMs idealized, complete inputs. NLData2Opt splits problems into descriptions plus raw data and requires models to extract relevant parameters despite misleading attributes, compute missing parameters by transforming existing fields, and forecast unavailable future values from history. Models show substantial accuracy declines relative to existing benchmarks, indicating that prior benchmark gains overstate formulation accuracy in practical operations-research workflows. Data and code are on GitHub.

A layer-wise interpretability study of WavLM on ASVspoof 5 combines per-layer linear probes evaluated per attack (TTS, voice conversion, adversarial), canonical correlation analysis against spoof-relevant acoustic features, and an exhaustive search over layer subsets. A five-layer subset outperforms pooling all 13 layers. Most acoustic features align best with mid layers, but jitter, shimmer and harmonics-to-noise ratio correlate more with later layers, and deep-layer embeddings show a continuum of attack similarity, suggesting they encode general synthesis artifacts rather than attack-specific signatures.

For automotive camera SoCs, the authors compress a YOLO detector by channel replay: an iterative pruning cascade records deletions, which are replayed on the original trained network followed by one knowledge-distillation recovery. On YOLOv5n it removes 60.3% of parameters at parent accuracy and beats iterative pruning and training from scratch; on YOLOv8n it beats the unpruned parent. An 18k-parameter integer-only reader matches FP32 at W4A8 in 12.1 KB. The 0.72M-parameter system recognizes 95.2% of held-out signs at 90% precision and 94 FPS on Jetson AGX Orin.

The Information's weekend column follows the DOJ arrest of former OpenAI engineer Abdulrahman Alfozan, accused of planting a toilet-paper-wrapped recording device in a conference room; a second device was left in a bag he later retrieved from lost and found. Ex-FBI counterintelligence head Bill Priestap says rapid AI-talent hiring leaves firms with minimal vetting, and a private investigator notes sub-$100 bugs disguised as chargers or smoke alarms. The piece points to Strider Technologies, which has raised $110M for AI-driven insider-threat intelligence.

Meta's updated commercial terms for business accounts had let it use content from those accounts' domains, web pages and URLs to "provide, develop, improve, research, train" its products. Legal experts quoted by Nieman Lab said the clause could let Meta scrape and train on the site of any publisher with a Meta business account. Meta has removed the language. It says the consolidated crawling terms applied only to its ads and business products and did not change how it uses web crawlers. Meta won an AI-training copyright case in 2025, but other suits are ongoing.

For SSM/Mamba-based low-light image enhancement, a lightweight estimator predicts a single-channel illumination map that is converted to prompt tokens and injected through a residual sidecar adapter at the bottleneck, leaving the state-space core unmodified. Under a three-seed protocol, PSNR improves over the base model by 0.94 dB (LOL-v1), 1.24 dB (LOLv2-Real) and 0.93 dB (LOLv2-Synthetic), with SSIM, LPIPS and dark-region gains. A learned-fixed control suggests the benefit stems more from the sidecar pathway than from illumination conditioning itself.

The post notes that when an assistant advises how to confront a friend and then interprets the friend's reaction, that reaction is not independent evidence, a case of behavioral confirmation (Snyder & Swann 1978) and performative prediction. It proposes closed-loop evaluations with simulated users and friends whose true intent is known. In a toy two-action example, a defensive reply should leave the estimate at 50% under one action but justify 80% under the other, so a model reaching 80% in both is overconfident. The post is conceptual and reports no experiments.

In the second post of her Computational Agent Theory and ASI Alignment sequence, Kosoy names the minimal task for safe AI: building a global defense system against unaligned ASI. She argues this requires aligned preferences, beliefs and decision-making together. She lists four difficulties: hypercomplexity, since preferences must be learned and proxies fail under Goodhart; grounding, meaning what data and mathematical objects encode preferences; convergence under irreversible errors, where Bayesian planning is intractable; and resilience to side channels. She proposes splitting learning into grounding and convergence with regret bounds, with technical details deferred to later posts.

This engineering explainer treats the actuator as a whole system: motor, inverter, current sensing, rotor feedback, gearbox, brake, output encoder and thermal path. Arm joints prioritize controllable low-speed torque, smoothness, backdrivability and collision response. Mobile drives prioritize efficiency, starting torque, regenerative braking and channel synchronization. The guide gives first-order sizing relations (τ_joint = τ_motor·N·η − losses, P_elec ≈ P_mech/η_total) and lists parameters to fix before choosing a controller. These include peak and continuous torque, reflected inertia, duty cycle, battery voltage range and the transmission error that a motor-side encoder cannot see.

Julian Bradshaw studies the coding-labor parallelization penalty λ in the AI Futures Model behind AI 2027 and AI 2040. Raw effective compute is very sensitive to it: λ 0.3→0.8 gives nearly 100x, and 0.50→0.65 gives about 4x at 10k agents. In the default model, however, raising λ lengthens timelines because λ is reused in calibrating serial uplift, β and η. With those side effects removed, 0.50→0.65 brings AC and ASI forward by about 2–5%, and λ=1 by 5–14%. The default 15x infinite-labor asymptote caps the effect. The author concludes swarm efficiency alone is unlikely to drive a fast takeoff.

Citing Bengio's remark that "RL is evil", the author takes as a working ethical hypothesis that the best action maximizes the sum of all current people's preferences over actions. Because reward arrives after an action and scores consequences, the reward hypothesis amounts to assuming consequentialism. Most people's preferences depend on means and intentions, so the author argues no reward function can match this ethic and every RL-trained agent would be misaligned regardless of reward design. The author proposes non-RL methods of goal induction and points to humans as an existence proof, but gives no concrete method.

An ML engineer turned AI-safety founder releases ai-exhibit, an open-source two-hour presentation from an ML4Good project, covering current capabilities, how LLMs are trained rather than built, race dynamics, risks from misuse to loss of control, and civic action. Attendees use phones to try embedding, base, assistant and tokenizer demos and a glossary; cited examples include evaluation awareness, unfaithful reasoning and shutdown avoidance. A presenter script is included; the first run was with six friends.