Wolf Digest — Tuesday, October 27, 2026

Between August 27 and September 9, Cloudflare's GLM-5.3 deployment advertised a 1,048,576-token context while its prefill KV pool held 681,280 tokens; in SGLang's disaggregated prefill/decode path, oversized requests were marked rejected but still passed the admission queue, so decode attended over stale KV pages and returned HTTP 200. 155 responses to prompts over 681K tokens could contain fragments of other customers' recent requests; Cloudflare found no sign of deliberate exploitation. Fixes include a 2.04M-token pool, retiring rejected requests before KV transfer (SGLang #36988, #38935), deferred KV release to stop late RDMA writes landing in reassigned slots, and pool-versus-context deployment checks.

Nemotron 3.5 Bolt 74B-A9B is a hybrid Mamba-Transformer vision-language model derived from the Nemotron 3.5 Super VL 120B-A12B family via Star Elastic compression and 2-bit quantization-aware distillation, fitting in 24GB. It ships with a LiLiCorr speculative drafter and runs at 84 tokens/s on RTX Spark. NVIDIA reports higher accuracy than similarly sized, similarly fast models on GDPVal v2, BrowseComp and Terminal-Bench 2.1, targeting long-horizon agentic workloads on local hardware.

Bedrock Robotics raised a $600M Series C co-led by CapitalG and Emergence at a $4.1B valuation, bringing total funding to nearly $1B. NVIDIA is among its existing investors. The company retrofits existing heavy equipment for full autonomy and has moved beyond the West into the Midwest and South Atlantic, anchored by a Texas earthwork partnership with Haydon. Its machines now work live job sites including data centers, hospitals and civil infrastructure.

AutoTavern is a self-improving system of specialized LLM research agents with shared typed memory and verification gates. In 20 days it built TavernGym, a full-content simulator of Hearthstone Battlegrounds, and 23 versions of a PPO-trained TavernAgent. Playing the live ladder alone, the agent climbed from 6,000 to about 15,000 rating, averaging placement 1.60 with 80% wins over its last 20 games versus 2.34 and 59% for the server's top ten. On the 960-task TavernBench it answers 91% of play tasks against at most 63% for frontier LLMs, at 0.03 s per decision.

On Oct. 15 at Fort Knox, four drive-by-wire ground vehicles running DARPA’s RACER off-road autonomy software moved and detected targets on their own after an operator tasked them from satellite imagery. The lead vehicle carried a boxed drone, and a simulated rocket hit a target about 300 yards away. Every engagement needed human approval of image chips sent back by the robots. Requested by EUCOM, it was the team’s first demonstration linking satellite imagery to tactical ground robotics.

SpaceX finalized $40 billion of GPU financing structured through an Apollo-affiliated special-purpose vehicle that sells debt, buys Nvidia hardware and leases it to SpaceX, keeping it off the balance sheet; chips arrive in the first half of 2027. SpaceX reported 1.7 GW of installed GPU nameplate capacity at September 30, is bringing roughly 200,000 GB300s online monthly before shifting to Vera Rubin, and guided fourth-quarter capex to $35–40 billion. Semafor reports the most expensive tranche costs up to 8.75%, about 200 basis points above its June bonds beyond Fed moves.

A small Block research lab released Dessa, a full-duplex conversational model trained from scratch only on recorded human conversation, with no text and no pretrained components. It listens and speaks at the same time, backchannels, laughs mid-sentence and responds to prosody. Block says it learned language from hundreds of times fewer words than LLMs, was trained on a single GB200 NVL72, and runs faster than real time on an M4 Max MacBook Pro. Safety researchers are getting early access.

Qwen-Image-PE trains a prompt enhancer for text-to-image models in two stages: SFT on dense captions of well-designed images (Caption-Based Reverse Distillation), then RL with a Groupwise Design Reward comparing all candidate images for a request on lighting, composition and layout. The 397B variant improves Qwen-Image-2.1 by 7.81 points on Qwen-Image-Bench; the released 9B model beats all tested PE baselines and transfers to four other backbones, beating native enhancers on HunyuanImage-2.1 and FLUX 3 Image Klein.

Instead of collecting on-site teleoperation demos, Scan2Instruction turns a single brief iPhone scan into a training environment with synthesized, grounded navigation episodes, and PRPT-KS adapts the policy using a closed-form Pareto utility computed from the reconstructed scene, with no learned critic or labels. Across five real environments, RealNav reaches 85% real-robot success versus 49% for a demo-post-trained baseline, and 82% vs 41% in unscanned regions of the same buildings.

Toyota plans to deploy 400,000 robots on its assembly lines as part of a $6.4 billion annual automation investment, aiming to catch up in humanoids and offset Japan's labor shortage. Analysts cite its 20 years of production data as an advantage over Chinese hardware-focused upstarts, and an investor says Toyota is developing its own large behavioral models to integrate robots, AI and production processes.

Meta will raise up to $95B through equity, including about $35B in underwritten public offerings and roughly $60B in newly created nonvoting shares that preserve Zuckerberg’s control stake. The offering is larger than Google’s $85B and Oracle’s $25B stock sales this year. Hyperscaler AI capex is expected to reach about $800B this year and more than $1T next year.

Trajectory2Gym argues that successful tool-use trajectories are poor supervision when they skip multi-step dependencies or contain actions not justified by what the agent could see. It synthesizes executable interactions around under-covered compositional structures and supervises only decision-grounded actions, with parent-normalized SFT weighting. Across five Qwen3 scales it improves Tau1, Tau2 and BFCL at every size, raising the three-benchmark mean by 16.9–29.8 points, with the largest gains in stateful multi-turn settings.

Microsoft said Monday it is offering voluntary buyouts to about 27% of its workforce, roughly 60,000 employees, to reduce expenses while it invests in developing and running AI. The Information describes it as one of the largest potential staff reductions at a big tech company in recent history. Microsoft also plans further layoffs this year, mainly in its Xbox unit.

S3AC offers a critic-training recipe for long chain-of-thought RLVR, addressing poor coverage of varied trajectory lengths and early gradients dominated by a new value head. The critic supplies token-level advantages and prompt-level baselines, which also drive Difficulty-Aware Prompt Sampling to choose prompts before generation and cut wasted rollouts. It outperforms the GRPO-X baseline across dense and MoE models from 1.5B to 397B, and at 35B sustains single-rollout training where critic-free baselines degrade.

PaperFormalizer runs inside a coding agent (Claude Code or Codex) and converts a paper's LaTeX into a dependency graph of statements, audits each for faithfulness, ambiguity and correctness, then formalizes and proves them in Lean against Mathlib, domain libraries and the author's own toolkit. It stops to ask the author when a statement is ambiguous and presents every repair as a proposal to accept or redirect. Outputs are a revised paper, a change report and Lean code linked to each statement, with shared verified papers hosted publicly.

Rubric-based RL rewards coverage, so it optimizes recall and drives verbose, partly incorrect answers that lose to the base model. FORK jointly rewards recall and precision, requiring each response unit to be necessary (rubric-driven necessity credits) and factually correct (a correctness grader). It reaches a Pareto-optimal coverage/quality trade-off, shortens responses at similar coverage, and raises win rate over the base policy from 10.6% to 68.0% on RubricHub-Medical and 22.1% to 74.1% on HealthBench.

Devin Security now runs deep scans for chained exploits using Claude Mythos and other frontier cyber models, available to enterprises through Devin without separate model access. It builds a living threat model with the customer, re-scores findings from its own and third-party scanners by exploitability, and can relabel with evidence (human approval required, KEV entries never relabeled), apply infrastructure mitigations, or run thousands of parallel VM sessions that reproduce, fix and test each vulnerability before opening a PR. Cognition is also committing hundreds of millions of dollars in free scans and remediation credits.

K3.1 Pro is a heavily post-trained Kimi K3 with large-scale RL across hundreds of domains, 1M-token context, and a Highspeed variant running two to three times faster. Moonshot reports 65.7% on Terminal-Bench 4.0 (Opus 5.5 66.4%, GPT-6 Astra 58.2%), 64.4% on OfficeQA Pro v2, and a claim-level hallucination rate down from 23.4% to 8.3% on its internal stress test; it scores 55 on the Artificial Analysis index, second only to Opus 5.5. Cyber data was removed for the open-weight release, cutting xdev-bench-hard from 90.5% to 81.0%.

TraceTrial-Bench evaluates agent harnesses rather than models, splitting reliability into seven components (context, execution, governance, lifecycle, observability, tools, verification) over 95 expert tasks with executable or state-based ground truth. Trace capture, multi-source evidence gathering and mechanism-chain failure attribution support a two-layer score. Across 4 harnesses, 9 models and 3,420 runs, harness scores range 59.8–72.0%, and 31.1% of task-successful runs violate at least one harness criterion — success rates hide harness defects.

CLE measures whether LLM agents improve from accumulated experience: 120 expert-authored professional tasks across 89 roles, each a five-level ladder where later levels depend on lessons from earlier ones, graded against human rubrics. Across 17 harness-model configurations, diagnostic feedback plus one revision adds a median 5.3 points, but the best system clears only 47.7% of levels and finishes all five levels on 23.3% of tasks. The data is also positioned for post-training continual-learning ability.

An independent frontier-risk evaluation of the open-weight DeepSeek V4.1 Flash-1017, structured around the EU General-Purpose AI Code of Practice and the AEF-1 standard, using 34 benchmarks (5 in-house) on evaluation integrity, loss of control and harmful manipulation. It shows substantial safety gains over the V4 generation and approaches closed-source baselines (GPT-6 Astra, GPT-5.6 Sol, Gemini 3.8 Flash) on several dimensions. Recommendations: evaluate every minor release, monitor reasoning traces and action logs, and build private benchmarks given evaluation awareness and contamination.

DynaFlow puts executable workflow synthesis into the main agent's action space: it decides when to collaborate and writes task-specific programs that run fixed sub-agents sequentially or in parallel with structured information exchange and verification. Only the orchestrator is trained, via SFT on distilled trajectories and refinement on examples filtered for unsupported premises and biased verification. Across six math, science QA and multi-hop retrieval benchmarks it reaches 88.08% overall accuracy, 10.90 points above the strongest baseline with the same backbone.

Labference evaluates whether lab robots adapt their next operation to observed physical outcomes, such as a failed placement, using controlled evidence variations and separate metrics for operation selection, physical execution and task completion. Oracle-assisted continuations from shared checkpoints show which forms of support help. Initial weighing tasks show tested VLA policies make target-selection errors despite successful object transfers, while GPT-driven control selects correctly but fails on execution and incurs heavy compute costs.

Δτ-Bench is an executable agent benchmark where services evolve across six ecosystem states: tools vanish, get replaced, or gain more efficient overlapping alternatives with noisy documentation. Its 1,588 multi-turn tasks span 42 domains. The strongest model/harness reaches only 42.3% success, and existing skill-learning methods fail to keep gains as the ecosystem shifts. The authors propose Steward, which incrementally maintains skills and provisions tools at runtime, sustaining higher success with lower maintenance cost.

A systematic test of dynamic evaluation (one SGD step per 128-token chunk on late-block MLP outputs). Softmax-attention models (Qwen3, Gemma-4) cut document loss 5–8% and beat the frozen model with full 32k context, while hybrid Gated-DeltaNet models gain under 1% — their linear-attention blocks already adapt. Masked-diffusion LMs gain 8–20%. But on RULER, BABILong, LongBench-v2 and others it recovers none of the accuracy lost to truncation: it learns style, not facts. Distilling conclusions into a LoRA instead matches in-context accuracy.

The Pentagon has opened several funding lines and competitions to scale small-drone production without guaranteeing long-term orders, and industry is divided over whether that is enough. Performance Drone Works’ CEO wants sustained demand, such as 1,000 drones a month for a year, before investing. Army Materiel Command says it is deliberately trying to “artificially induce” demand to stand up a domestic supply chain, as the post-Replicator drone push seeks billions in funding.

Agility Robotics and GXO are extending Digit deployments to GXO sites in North Carolina and Tennessee. GXO was the first customer to run Digit in live operations under a commercial agreement, and Digit moved more than 100,000 totes for GXO last year. Agility cites over 65,000 hours of real-world operation and has SEC approval for its SPAC merger with Churchill Capital Corp. XI. It also plans a deployment at an AGCO factory.

InteractiveWorld is a verifiable simulation where an egocentric AI assistant instructs a simulated human agent who observes and acts independently, with the environment updating state in response. Its InteractiveBench contains 120 multi-turn tasks in six categories scored by state-based rewards. Across 12 frontier MLLMs, the best model, Claude Fable 5.1, reaches at most 20% on long-horizon Guidance and 25% on Planning tasks, indicating wearable-assistant capability remains far from reliable.

World-in-Context is an interactive video world model that places explicit view, action and camera tokens directly in a shared self-attention space instead of separate conditioning pathways, plus Interaction-Transport RoPE to keep continuity across control intervals. Explicit view tokens enable emergent view switching from single-view data. On WBench the 5B model matches 22B EchoWM on navigation (81.5 vs 81.6) and reaches 93.8% perspective-switching success versus 18.8% for 5B baselines.

GEMS keeps a persistent spatiotemporal 3DGS map, reconstructed once and updated only where the scene changes, and pairs it with an agentic pipeline that renders novel views on demand to gather evidence for spatiotemporal questions and language grounding. The accompanying GEMS-Bench has 473 QA pairs across 3 real and 12 synthetic semi-static environments. GEMS beats recent baselines on 3D grounding and spatiotemporal QA with rendering quality on par with dedicated scene-update methods and lower storage.

The Army launched Project Gatekeeper, which automates the months-long authority-to-operate process: vendors commit code to the Army DevSecOps Platform and validate it against enterprise standards in containerized enclaves in near real time, even before a solicitation. Access currently requires a DoD CAC and will extend to Login.gov credentials. The Army plans to add unnamed frontier AI models to strengthen security tooling, aggregate security reporting and prioritize remediation.

OvisOCR3 is a 0.8B model unifying document parsing, general OCR, key information extraction, document figure captioning and cross-page table reconstruction. Training uses continued pretraining, domain-balanced mid-training, multi-teacher on-policy distillation from five 4B domain experts and mixed-task RL. It scores 96.74 on OmniDocBench v1.6, ranks first on OCRBench v2 (65.9 English, 61.3 Chinese) and tops all four KIE benchmarks, showing compact models can cover broad document AI. Weights are on Hugging Face.

StreamVibe-Bench provides 4.1K eval and 85K training samples covering explicit, implicit and dynamically changing user intents in streaming video. The accompanying StreamVibe-VL model infers implicit intent from continuous observation and improves response timing and answer quality by 10.7% over baselines while keeping general video understanding. A harness that keeps intents as persistent states in hierarchical memory adds a further 12.3% on proactive interaction; existing proactive models struggle most with implicit and dynamic intents.

At the Ateş Serbest live-fire exercise near Ankara, Turkish Aerospace showed its turbojet-powered Ayaz one-way-attack UAV publicly for the first time; it carried out dive attacks on targets. Super Simsek flew with a warhead for the first time. Twenty STM Kargu drones flew as a swarm, sharing data in real time and splitting into two groups to hit separate targets at once. ISR drone feeds went to a mobile command system to support strike decisions, and Aselsan's Ejderha microwave and Gokberk laser systems downed drones. No specs were disclosed.

AWS released TorchNeuron in public beta, an open-source native PyTorch backend for Trainium2 and Trainium3. Existing code moves over by switching to .to('neuron'), while keeping eager execution, torch.compile and standard PyTorch APIs. The Neuron Kernel Interface (NKI) handles custom kernels and low-level tuning. Removing the porting step is AWS's most direct attempt yet to weaken CUDA lock-in for training and inference. GA timing and pricing were not given.

Arrow Research scored frontier agents on four long, open-ended AI R&D tasks, using automated checks plus expert-rubric LLM judges. Aggregate scores double about every 11.5 months (95% CI ≈7.2–16.7). The trend is fastest on GPU-training-heavy engineering tasks and slowest on the most conceptual task, where a piecewise linear fit beats the exponential. GPT-6.1 Sol is the best model overall, and the authors project the benchmark will saturate by mid-2027. Scores correlate only weakly with reasoning effort or wall-clock time.

In a letter included in Anthropic’s IPO prospectus filed Monday, the company’s seven founders committed to keeping Claude free of advertising. They wrote that Claude’s only incentive should be giving helpful answers, and that once advertising enters revenue targets “the pressure to optimize for engagement tends to grow.”

On ProgramBench (rebuilding programs like SQLite from a binary plus docs), Claude Opus 5.5 (xhigh) scores 18.5% fully resolved and 70.5% at over 95% of tests, up from 0% and 3% for Opus 4.7 at its May launch. Mirendil says one researcher built Mirendil-Base in 3.5 months on under $10M of compute and that it beats Nemotron-550B-Base, GLM-5.3 and DeepSeek-V4-Pro-Base, though it trails Kimi-K3.

Resolution used orchestrated agents to formalize Hironaka's resolution of singularities in Lean: about 600k lines in a week, roughly $135k in tokens. It also covers OpenAI agents leaking secrets between microVMs via cache-timing side channels.

Gen. Francis Donovan told Breaking Defense that SOUTHCOM wants a river-reconnaissance system that combines unmanned surface, air and subsurface platforms to map how criminal networks move narcotics along South and Central American rivers. Saronic Corsair USVs, first moved to Guantanamo Bay, now operate on both sides of the Panama Canal. A new SOUTHCOM Autonomous Warfare Command at Guantanamo trains with and tests unmanned systems and is open to industry. Donovan argues cheap loitering USVs can free destroyers for specific missions.

A CSIS analysis traces how the FCC Covered List has moved from naming suppliers to banning whole product categories by place of production. It added foreign-made drones and critical components in December 2025, consumer routers in March 2026, and advanced robotic devices, inverters and some logic chips from listed entities in July 2026. The authors warn this can hit U.S. firms that assemble abroad, and cite an estimated $5.3B in five-year costs. They recommend transparent waiver processes and coordination with allies, and note an August proposal to split the list into producer- and location-based categories.

Smack Technologies will prototype Omega, an AI planning and decision-support platform that generates linked objectives, tactics and tasks as a single course of action and can repair plans in seconds as assessments arrive, under an Other Transaction agreement with SOCOM worth up to $7.5M. Omega runs on small domain-specific models that combine classical planning with deep RL. Any third-party LLM is narrowly scoped and grounded in a proprietary context graph, and the models are small enough to run on-prem at disconnected nodes.

USARPAC’s commander said that in a trilateral U.S.–Japan–Australia exercise allies once could not pass voice or data and had to enter icons into the common operating picture by hand. Balikatan 26 and Yama Sakura 91 have since used the Pacific Mission Network’s partner enclaves for live sharing, including target-quality data, across echelons. The aim is a single-screen common operational picture shared with allied forces.

WorldCast targets character video generation along large camera trajectories that reveal regions unseen in the first frame. It back-projects the initial frame to a depth-based point cloud rendered along the target path for geometric guidance, and adds Geometry-Aligned Memory Attention, which retrieves view-relevant features from multi-view character anchors in a shared coordinate frame. The authors also release CLACT, a character-centric dataset and benchmark of large-angle trajectories with accurate camera geometry; results show accurate trajectory following with preserved identity and appearance.

TrajMamba predicts short-term human trajectories from RGB-D input using stacked bidirectional Mamba blocks over hip-center velocities and a GRU decoder. It closes the train-deploy gap by running the same pose extraction, depth back-projection and smoothing pipeline in training and deployment rather than using idealized ground truth. It reaches 0.29 m ADE and 0.51 m FDE on ETH/UCY, runs in real time with 262K parameters, and is deployed on a TIAGo robot with a Jetson Orin.

At GTC Berlin, Jensen Huang said Nvidia's software for running agents on AI factories would cost about $2,000 per GPU per year, against a $4,500 per-GPU list price for AI Enterprise; Nvidia has not said what the license covers, when it starts, or how it overlaps with AI Enterprise. The context is Nvidia's new OpenShell reference architecture, a secure agent runtime on Vera Rubin, Vera CPUs and BlueField-4, with Nvidia Sentry monitoring and quarantining agents. Anthropic is integrating OpenShell with Claude Managed Agents.

EvoRouter learns to configure multi-agent systems (models, collaboration modes, agent counts, rounds, skills) per query. An expert periodically distills performance-cost feedback into reusable routing insights, which a privileged teacher uses to give dense token-level on-policy distillation to the router, combined with cost-aware RL in a self-improvement loop. Across 19 in-domain datasets in five domains it averages 68.69% accuracy, 7.23 points above the best routing baseline, while cutting inference compute by 50.95%. Code is open-source.

Studying harness self-evolution across seven models and three benchmarks, the authors find 103 task-skill pairs where skills improve GPT-5.6-luna by 31.1 points but degrade Gemma-4-E4B by 8.8; in 86% of failed weak-model runs the violated requirement was already in the skill. They distill failure diagnoses into a fixed meta-skill for rewriting skills to fit the executor: Kimi-K3-guided rewrites improve three weak executors by 5.1–14.1 points, and in a five-round self-evolution loop the meta-skill adds 2.0–5.1 points across all seven executors.

An independent analysis of AFM 3 Core Advanced in macOS 27 finds a roughly 20B-parameter sparse MoE that picks experts per prompt (1–4B active, reselected about every 32 tokens) and streams 2-bit weights from a 4.95 GB file in flash on the Neural Engine. On an M3 Max it held about 2 GiB, against 15 GiB for Gemma 4 26B A4B, and used about 465 J on 20 GSM8K problems, roughly a third of the best open model. It is weaker at math, scoring 78/100 on GSM8K vs 89–94 for 4B open models.

AWS open-sourced the Agent Toolkit for AWS (Apache 2.0), a set of officially supported MCP servers, skills and plugins that give coding agents such as Claude Code, Codex, Cursor, Devin and Kiro service knowledge and guardrails for AWS. Plugins cover core services (CDK/CloudFormation, serverless, containers, observability, billing), building agents on Bedrock and AgentCore, and data analytics with S3 Tables, Glue and Athena. They install from Anthropic’s official plugin marketplace or through a new aws configure agent-toolkit CLI command.

MIT researchers built a gum-stick-sized robot from a stiff, roughly 0.5 mm GelMA frame with two fins, each coated in an optogenetically engineered muscle layer thinner than a hair. Square-bottomed grooves align the cells, and light-pulse training strengthens them. Lighting one fin or both steers the robot through a maze at up to about four body lengths per minute. Published in Advanced Functional Materials, it is described as the first thin 2D biohybrid swimmer.

AMBIT is a framework for native First-Order Ambisonics generation and editing. A shared Transfusion backbone couples autoregressive planning of executable ScenePlans (source content, activity, 3D trajectories) with rectified-flow rendering; an FOA-aware VAE preserves inter-channel spatial structure; and joint on-policy self-distillation refines both planning and rendering under student-generated states. Across speech, music and sound it reports better content and spatial quality, editing fidelity and preservation of unedited sources than compared baselines.

AndroidFlux, built on AndroidWorld, restores recorded trajectories to checkpoints before, at and after the first recoverable error so different agents continue from the same device state and history. A companion reward-modeling benchmark scores recovery-action preferences against downstream rollout outcomes, and an offline RL dataset is released that improves both recovery and overall task success. Recovery performance correlates only partially with clean-start success, and explicit reasoning helps only models trained with GUI-specific RL.

PatEx-Gym frames patent novelty examination as an agentic search task: PatEx-Bench offers 1K realistic tasks, corpora of over 1M patents and 52K papers, and a runtime for iterative search and Office Action generation. PASS synthesizes training trajectories without annotation by treating document visits as implicit relevance signals. A 4B PatEx-Agent fine-tuned on this data competes with much larger general agents, while prior-art discovery remains hard even for frontier models and final-decision accuracy alone misstates examination quality.

Diffusion LLMs make fast parallel drafters for speculative decoding, and both diffusion effort and verifier acceptance leniency can trade quality for speed. This paper maps the joint Pareto frontier over those two knobs and shows that varying them together pushes the speed-quality frontier beyond single-step diffusion drafting, framing lossy speculative decoding as a tunable design space rather than a strictly lossless technique.

EnvLunch adds a warm-up stage before agentic RL for data-analysis agents: it samples valid action trajectories in an executable environment (e.g. a database), executes them, and trains the agent to recover an equivalent trajectory from the final observation alone — no human or model-written task descriptions. It improves downstream accuracy by over 9 points on average across training budgets; baselines need 8–32x more annotated examples to match, and it beats having the model write its own tasks.

Using post-training corpora with controlled mixtures of opinions tied to demographic groups, the authors compare five SFT/DPO/GRPO post-training paths. Post-training amplifies opinion imbalances to different degrees depending on the path, and whether later training can undo an amplified preference depends on both method and data mix. Stability under interventions tracks whether further training shifts the preference, and value preferences carry over to student models even when distillation data contains no explicit opinion questions. Code and data are released.

Recursive Semantic Divergence (RSD) is an unsupervised, trajectory-level consistency metric adapted from bisimulation metrics: it estimates expected cumulative NLI-detected disagreement between two independent rollouts from the same state, learned by a neural network so it scores each turn without completed trajectories. Used as a fine-tuning reward, RSD cuts variation in task progress from 132% to 91% on tau2-bench and from 90% to 44% on ToolSandbox while improving performance, matching SFT with lower variance and no labels.

SLICE-Mem replaces LLM-driven memory admission and extraction with a compact salience classifier and a semantic-dependency extractor that turns retained user turns into concise entries, deferring reconciliation to offline consolidation. Against the generative MemOS baseline with Qwen3-30B, it scores 46.60 vs 46.68 FAMA on Memora Weekly and 56.8% vs 59.3% on LongMemEval-S, while reducing total LLM token consumption by 92.4% and 99.5%.

VPR-Space compiles sixteen database-query datasets from terrestrial space-analog sites with varied viewpoint and illumination challenges. Rather than fixed pose thresholds, ground truth marks an image pair as a match if an ensemble of relative pose estimators would yield a useful pose for SLAM loop closure. Nine state-of-the-art VPR methods are evaluated and leave considerable room for improvement in feature-poor extraterrestrial scenes.

OMRON agreed to buy all of OnLogic, a Vermont maker of configurable, AI-capable industrial and IoT edge computers with more than 70,000 customers, for $675 million. The deal pairs OMRON's sensing, control, safety, vision and motion automation with OnLogic's hardware, so manufacturers can run AI at the edge alongside real-time control. OnLogic keeps its brand, headquarters and global offices. OMRON frames the purchase as part of its SF2030 strategy and its push into North America.

A SPAR-originated paper elicits 24 propensities through 4 methods each and measures spillover to non-target traits. The best predictor of SFT spillover is how other models behaved under the same finetune. Next best is system-prompting the target trait. Third is generic drift, which pushes Qwen3.5-9B toward its base model. Assuming misalignment and trait X raise each other symmetrically is the worst predictor. Self-preservation is the concerning trait that rises most often under benign finetuning. Human trait correlations predict the spillover only weakly, which the authors read as limited support for the Persona Selection Model.

OpenAI Signals released quarterly data from Q1 2025 to Q3 2026 that maps consumer ChatGPT work messages to 165 O*NET Intermediate Work Activities across nearly 100 countries. Enterprise and Codex use are excluded. Work use became less concentrated: across 74 countries, the top ten activities fell from 56.0% to 51.1% of classified work messages, and countries specialized in different work over time. The aggregates are de-identified, small groups are suppressed, and statistical noise is added.

This study tests whether RLVR gains in VLM image quality assessment come from better reasoning. Using sampling-aware max@k evaluation, chain transplants and steering, it finds RL-trained models do beat base models, with gains beyond averaging mostly in-distribution, and RL chains help when transplanted into base models. Yet scoring depends less on the chain at inference, especially at 7B versus 32B, and score-based rewards can exploit textual cues correlated with MOS, potentially inflating benchmark results.

The authors train harmful-request linear probes on four base models (Qwen2.5-Math-7B, Llama-3.1-8B, Qwen2.5-14B/32B) and apply them unchanged to DeepSeek-R1 distilled and instruction-tuned successors. The unsafe/safe direction rotates 17°–35°, yet AUC changes only −0.011 to +0.027 on XSTest and −0.015 to +0.053 on held-out OR-Bench. Probes beat text-only classifiers by 0.15–0.20 AUC on XSTest but flag 29–42% of benign-but-alarming OR-Bench prompts, suggesting they partly read the appearance of harm.

OpenAI released GPT-6.1 Terra, its mid-priced tier, saying it costs up to 50% less per token than the Terra model it shipped three months earlier. Artificial Analysis lists the Max reasoning variant at $1/$5 per million input/output tokens with a 95% cache discount, a 1M-token context and text plus image input.

It scores 48 on the Intelligence Index (#25 of 221; class median 28) at about 91 tokens/s, but it is verbose: 130M output tokens to run the index versus an 81M median. That works out to about $0.40 per index task.

A long-form podcast with MIT's Ankur Moitra, UW's Steven Brunton and Rutgers' Alex Kontorovich on the claim that AI systems have resolved major open problems, including Navier-Stokes, a quasi-Riemann hypothesis result and the unique games conjecture. They discuss whether AI will 'solve' mathematics, whether it can create new math, and the role of Lean and machine-checked proofs, along with turbulence modeling and the complexity theory of learning.

Woolf catalogs cuts from June to September 2026, including GPT-5.6 Luna falling about 80% to $0.20/$1.20, GPT-6 Sol at $2/$10, Opus 5.5 at $4/$20 and a 75% cut to Fable 5.1 cache reads. He attributes them to commoditization through routers like OpenRouter (29 providers serve DeepSeek V4.1 Flash) and real serving-efficiency gains, and rejects the loss-leader theory. Since cache reads are 96–99% of agent token spend, cache cuts drive most of the effective savings.

At GTC Berlin, Nvidia unveiled the DSX POD, a factory-built shipping container with seven Vera Rubin racks that travels on normal roads to cut deployment from months to days. It also announced OpenShell agent sandboxing and the Agile ONE SN humanoid on Jetson Thor, and said it has invested nearly $10B in Europe. On the floor, talk centered on neocloud financing as CoreWeave fell 10% and Nebius 7%. Huang put a 1 GW data center at about $50B and said demand for capital outstrips supply.

An Air Force SBIR solicitation asks for a system that uses tanker-mounted cameras on the KC-46 to automatically track the boom and receiver aircraft. It would measure position, velocity, alignment and closure rate, flag deviations from nominal profiles, and generate post-flight feedback through overlays, heat maps and timelines. The aim is to replace subjective instructor review and to serve as a prototype for AI-enabled aviation performance assessment across DoD.

An independent audit of a cheap frozen classifier-style moderation model on 142,658 human-labeled comments, benchmarked against 14 frontier reasoning models. Its zero-shot AUROC of 0.83 is not statistically separable from 13 of the 14 at 1–3 orders of magnitude lower cost. The paper argues confidence conflates annotator disagreement with decision risk, which need opposite escalation signals; decomposing judgments into parallel sub-questions recovers a disagreement signal, but only for weaker models and aligned decompositions.

Multi-reference image generation in a single DiT suffers quadratic cost from token explosion. ToDoPro runs a few denoising steps on 2× downsampled references to estimate token importance with the model itself, then discards low-importance tokens, replaces mid-importance ones with downsampled versions, and keeps the rest before the main denoising stage. It needs no extra parameters, stays compatible with in-model acceleration, and across three base models and two benchmarks speeds inference, reduces peak VRAM and can improve quality by filtering noisy reference tokens.

A CLEAR study reported by Rest of World finds that among the 20 most-used internet languages, the costliest needs 1.5–5x the tokens of the cheapest, and across 200 languages the gap reached 17x for one model. Rarely seen scripts such as Burmese, Tibetan, Khmer, Lao and Ol Chiki fare worst across nearly all tokenizers analyzed, which also means higher latency and less effective context. Researchers urge labs to design tokenizers for these scripts and publish per-language cost scorecards.

Ollama 0.40.11 adds speech models on macOS, with Windows and Linux to follow. The first is Gradium's Phonon, a family of small TTS models with dozens of voices, run with 'ollama run phonon'. Ollama says its MLX-accelerated build runs Phonon 35% faster than the next-fastest option on an M5 Max MacBook Pro while matching the reference implementation's accuracy. Apps can call it through an OpenAI-compatible /v1/audio/speech endpoint, and a chat model's output can be piped in to stream speech.

OpenAI’s Alexander Embiricos told TechCrunch Disrupt that dots, launched at developer day in September, repackages Codex-style agent capabilities as a single persistent entity in one ongoing conversation. Connected to calendar, email and the user’s computer, it keeps working in the background. It is available on ChatGPT Pro tiers and above, starting at $100/month. Embiricos said much of the work was framing rather than new technology.

The paper introduces objective-conflict coefficients and derives near-optimal online regret bounds that quantify how conflicting objectives hurt sample efficiency. It proposes routed ensemble policy learning, which keeps a few basis policies and routes trajectories to the most promising one, plus a kernel surrogate method that estimates how feedback from some trajectories affects learning on others. Routed ensembles gain 4.3% on Meta-World and MO-Gymnasium; kernel attribution improves leave-one-out correlation by 50% and trajectory-selection reward by 27%.

CodeRabbit now runs GPT-6.1 Sol as the main model in CodeRabbit Secure, which reviews PRs in real time and runs repository-wide Deep Scans that trace changes across files, shared libraries, dependencies and linked repos such as shared auth code. For selected findings it attempts exploit verification in an isolated sandbox, uses an independent verifier to reject unsupported findings, and drafts fix PRs for human approval.

CodeRabbit joined OpenAI's Daybreak Defense Network and is covering $1M of free Deep Scans (up to $5,000 per org) through November 8. No detection or false-positive rates were published.

Probing three feed-forward 3D reconstruction backbones shows static/dynamic separation is linearly decodable from frozen tokens, concentrated in a low-rank subspace of π³X. Probing-based Motion Distillation converts motion-compatibility scores against an anchor token into dense motion masks. Trained on 160 YouTube-VOS videos on a single 16GB GPU, it transfers zero-shot to DAVIS, SegTrackv2 and FBMS-59, beating training-free heuristics and approaching heavier supervised methods.

TIAgent turns a natural-language analysis request plus a dataset into an inspectable, executable computational graph built from existing pathology operations, targeting uses like biomarker discovery without coding. A benchmark of 90 unit tasks over 18 operation types compares 14 backbone LLMs; Kimi-K2.7-Code tops Unit Workflow Success at 0.967. Node- and parameter-level analysis shows choosing the right operation and configuring it correctly are distinct failure modes.

Kapa traced gold chunks through a retrieval pipeline (GPT-6 Luna query decomposition, Voyage-3 reranking, 50 retrieved, top 15 returned) on its Company Knowledge Bench. About 18 months ago, around 80% of failures came from gold chunks never being retrieved. Today search and selection each cause about half. Of the gold chunks that search did surface, almost a third were never shown to the answer model. For code, grep driven by a capable model is now enough, but heterogeneous company corpora still need better selection.

AgentProofBench pairs tests of a frozen acceptance rule with execution from identical saved starting states, covering 97 tasks in 79 families across four web services. With DeepSeek Flash/Pro, complete-panel and decision-replay tests retain 82/96 and 79/96 source programs. Panel correctness predicts delivery: 32/33 assignments with error-free panels deliver correctly versus 1/11 with panel errors, and all program-gate regressions fall in the latter group. Instructions, states, programs and trajectories are released.

In an OpenAI case study, code-review startup Greptile describes using GPT-6 Astra's computer use to spin up a pull request's dev server, resize the browser to a phone viewport, and click through the app; in one example it caught a button that vanished on mobile. Greptile uses Daybreak for security hardening and exposes its codebase Knowledge Base to Codex over MCP so generated code follows house patterns, tracking revert rates as its quality signal. The post reports thousands of customers connecting Codex to the knowledge base but gives no measured outcomes.

Breunig traces context engineering from long-context failure modes (poisoning, distraction, confusion, clash) and six fixes (retrieval, tool loadout, quarantine, pruning, offloading, summarization) to 2026's harness era. He names two newer problems. One is 'fighting the weights': Claude Code's prompt repeats parallel-tool-call instructions because training favored serial calls. The other is 'context inertia'. He cites periodic forgetting, such as Berkeley's result of about 1.14M vs 1.01M on AHC014 with four agents, and suggests around 200K tokens as a point to reset context.

According to an internal email obtained by 404 Media, Proton told staff on Monday that it will re-interview employees across the company about how they use AI in their jobs. The email indicates the interviews may decide whether staff have “continued employment” and calls them “consequential.” The rest of the report is paywalled.

Benjamin Todd scores forecasts made by superforecasters, AI researchers, AI CEOs and himself against the past 18 months. Most underestimated progress, superforecasters and the general public most of all. Experts were closer but still too slow, and only AI company leaders were slightly too optimistic: their six-month claims tended to come true in six to 18 months. Of his own 2025 end-of-2026 capability benchmarks, he judges about three on track and three or four too conservative.

On a dbt podcast, Shreya Shankar discusses Data Agent Bench (DAB), which tests junior-analyst work: joins across databases, undocumented business context and messy free-text columns. The best agent passed 61% of queries on the first try. Public leaderboard scores in the 80s reflect per-query prompt overfitting, including about 85% for a heavily prompted model at roughly $10k per run. A human-understood semantic layer significantly improved accuracy. For cost, she recommends pushing computation into the databases, using fewer and smaller LLM calls, and planned AI-SQL query planning.

SimPOTrans uses a two-stage RL post-training pipeline that rewards an offline speech-translation LLM for behaviors useful on partial speech input, enabling dialogue-style simultaneous translation without synthesized read/write trajectories or explicit latency objectives. Inference needs no external decision policy; quality-latency is controlled by input chunk size. On ACL 60/60 and MCIF for En→De, En→Zh and En→Ja it beats strong baselines on the quality-latency trade-off.

Ben Thompson argues that AI is speeding the decline of the open, self-hosted web. He points to a week with an RCE bug in every PHP version, flaws in Node.js, GitLab, Java and Cisco, and supply-chain breaches at LiteLLM and Railway. AI removes the human-time limit on attackers, and AI answers consume content without sending traffic back. He expects things to get worse before agents continuously defend rewritten software, and sees a large market in cheap custom software built for one person or a small team.

claude-red is a curated set of SKILL.md files that load offensive-security methodology into Claude on demand. Coverage runs from SQL injection and shellcode to EDR evasion, ADCS abuse and exploit development, with skills triggered by conversation context. Its stated use cases are authorized red teaming, bug-bounty triage, CTF prep and operator training. It is an example of agent skill packs carrying dual-use expertise.

John Loeber describes building a private Lean4 fork, a parallel subagent harness with a lemma cache, and C++ counterexample search to attack graph-theory problems such as the Hadwiger conjecture. GPT-5.5 and GPT-5.6 Sol made partial progress. After GPT-6 Astra's jump in proof-strategy intuition and the reported Navier-Stokes resolution, he stopped. He describes OpenAI running an unreleased model on 4,000 major problems at about 3.5 hours each and argues that scaffolding gets absorbed by base-model gains, though commenters dispute how foundational the results are.

In an opinion newsletter, MIT Technology Review looks at how Anthropic presents itself as its S-1 lands; the prospectus runs 335 pages, about 100 of them on risks. The piece cites Jack Clark telling an IMF audience that the singularity has begun and Chris Olah telling the Vatican the company finds internal states that functionally mirror emotions. It argues these messages shape public understanding while the technology's trajectory stays uncertain. It also notes Encode AI criticism of a draft bill Anthropic circulated to lawmakers.

As shopping assistants such as Instinct and Meta's Muse gain traction, Stripe has become a leading payment rail for agent purchases. It often issues one-time virtual cards, which leave the card networks with less visibility and control over the transaction. Visa and Mastercard have built agent-payment schemes that use customers' existing cards and keep the networks in the loop, but those efforts are still early. This is a paywalled report from The Information.

On a Lawfare Daily episode, Stanford's Riana Pfefferkorn and Edinburgh's Cyndie Demeocq discuss their review of 71 federal prosecutions involving AI-generated CSAM: 65 convictions, 32 involving nudify apps and only 9 with purely synthetic imagery. Prosecutions rose from a couple in 2022 to about 40 in 2025, and only one of eight First Amendment challenges partly succeeded. They cover open gaps, including the Seventh Circuit's Anderegg possession ruling and legal risk for red-teamers, and compare the UK's Crime and Policing Act, which creates a vetted testing pathway.

The essay argues that computer-use tasks on well-documented apps are now largely retrieval for frontier models, and that the remaining difficulty is tacit organizational knowledge, such as approval thresholds or who actually controls budget. It proposes building RL environments as one world exposed through several surfaces. Task difficulty then varies by surface, and the reward reads world state, so any surface that reaches the same authorized result gets the same score. This rewards achieving the goal rather than an email sent.

Dan Luu argues that AI can now one-shot many projects but still struggles with hard ones, and that claims that software is solved come from people not working on hard problems. His main example is pgrust, an AI-assisted Rust rewrite of Postgres focused on performance. Its testing has uncovered real upstream Postgres bugs, including SP-GiST, JSONPath format-cache, pgcrypto, dictionary-loader memory-safety and money-overflow issues.

S²-RL recasts semantic-segmentation dataset distillation as text-to-image generation, avoiding pixel-wise gradient unrolling. A diffusion policy is fine-tuned with GRPO under a dual reward: per-sample semantic alignment and intra-class diversity, plus set-level inter-sample distribution diversity framed as a contextual bandit. The authors report consistent gains over prior distillation methods on ADE20K and COCO.

Tiger Data built three memory systems for its Slack agent, Eon: a user-scoped fact list, a general knowledge graph and a tree. Agents struggled to decide what graph edges meant. A hierarchical tree on Postgres’s ltree extension worked better, cutting the filing of a thread summary from dozens of turns to one tool call. The team extracted the hybrid search layer into searchgres.js, an open-source TypeScript library for composable filtered search in PostgreSQL.

Notes from HPTS 2026. Ion Stoica reported that 10 of 12 Berkeley students using agentic loops beat published baselines on research problems, yet agents still cannot generate new research ideas or build production systems. In one example, an agent asked to build a fast key-value store got a 6x speedup by noticing that YCSB generates values with a hash and computing them instead of storing them. Stoica frames such failures as requirement and model gaps, and the meeting’s recurring theme was that logs, formal checks and human review are the real control plane.

A Machine Learning Street Talk documentary built on Kevin Hartnett’s book The Proof in the Code traces Lean from a Microsoft Research side project to a working tool for mathematicians. It features Leonardo de Moura, Jeremy Avigad, Kevin Buzzard and Paul Lessard, and covers Mathlib, the Liquid Tensor challenge and Tao’s Equational Theories project. The second half examines AlphaProof and autoformalisation, and asks what is verified when a checker certifies a statement that may differ from the intended mathematics.

Because avatar videos decouple rendered appearance from the operator, this work verifies operators from motion. It combines diffusion masked pretraining over grouped facial landmarks (with auxiliary behavioral-cue tasks and random transforms simulating avatar generators) and frame-difference video features, without external pretrained models. The fused embedding reaches 0.769 AUC on WBA2026; probing shows the video embedding carries little avatar-identity information and fusion helps across speaking and listening modes.

A technical explainer contrasts Parquet's row-group and page layout with LanceDB's open Lance format. In Parquet, reading one row decodes a whole page, the Thrift footer must be fully deserialized, and buffering entire row groups can exhaust memory on embeddings. Lance drops row groups, stores per-column metadata reachable directly by offset, and uses mini-block encoding for small values and full-zip encoding for values over 256 bytes, so single values can be decoded alone. The '100x faster random access' figure is the vendor's claim; the article reports no benchmarks.

A native C++ extension for DuckDB 2.0 lets SQL query and write hf:// URLs across datasets, models, Spaces and buckets, with revision support. Reads use range requests and a shared read-ahead cache; writes go through the Hub commit API and Xet chunk-level deduplication, including partitioned COPY. Table functions cover listing, metadata, uploads, deletes and repo creation. Limitations include writes staged to a temp file and committed one file per commit, a cache held in memory, and sequential large-file reads that trail httpfs.

In a talk first given at an internal Microsoft conference, Maggie Appleton argues that unattended agent workflows depend on high-quality plans. She says today’s ways of thinking through problems with agents are rudimentary given how scarce in-the-loop human attention is. Drawing on Thomas Gladwin’s anthropology of Chuukese navigation and the idea of boundary objects, she argues for richer shared artifacts and thicker interfaces between how humans and agents represent a plan.

Bloomberg reports that Jane Street, HRT and other high-speed trading firms are spending billions on compute and investing in the sector to secure processing capacity and faster chips. These firms rely increasingly on machine learning, so trading performance now depends on compute as well as latency. Despite the risks, most see little alternative if they want to avoid obsolescence.

Frigade released Hob, an MIT-licensed desktop app (macOS Apple silicon, Windows x64, Linux AppImage) that automatically picks and installs the strongest open-weight model that fits the machine, such as Gemma or gpt-oss, and runs it offline. It targets assistant tasks like working through documents and filling PDFs, and the repo points to llama.cpp inference and MCP support. It needs 8 GB+ RAM. The project is early, with about 17 GitHub stars at the time it reached the HN front page.

LangChain describes how its Deep Agents harness handles three retrieval failures: inefficient search, oversized context dumps and stale context. Filesystem tools search before reading (glob, grep, paged read_file), and oversized tool results are written to files so the agent sees only a path and a preview. For custom search such as RAG, the recommended pattern is to save results to files, return paths and let subagents read them, so the main agent sees only what is relevant.

The post asks whether persona-related interpretability work distinguishes a “deceptive persona” feature from a plain “deception” feature. Under the Persona Selection Model, a persona feature should encode an attribute of the currently simulated character. The author proposes treating actions as evidence about personas: an action feature would spike at the moment of a lie, while a persona feature would update the character attribute after the event. The question is posed as open.

In a BCBusiness op-ed, the CEO of Caseway, a 15-person firm whose document and form AI is used by courts and more than 100 law firms, says a Sept. 8 U.S. directive removing Canadian products from federal schedules led it to halt new U.S. business development. Its Taiwan work includes counter-drone and maritime decision support with MiTAC and a planned paid pilot with drone maker Kunway for detection software that lets one operator fly ten aircraft. In Europe it is building defence software for TKMS and has registered on Ukraine's Brave1.

An MIT Technology Review op-ed by the cofounder of the Young People’s Alliance and the head of Hopelab’s Responsible Technology Youth Power Fund argues that tech policy measures success by the absence of catastrophic outcomes. Products can meet narrow safety requirements while fostering dependence on AI companions or chatbot-written schoolwork. The authors cite a figure that a third of teens discuss serious matters with AI companions rather than people, and call for standards that address everyday harms.

Definition-generation models used for lexical semantic change detection tend to produce too many definitions for a single sense. Fine-tuning with DPO reduces this granularity problem, evaluated on an academic LSCD benchmark and qualitatively on auditing shifts in sustainability terminology in EU development-finance documents. Models, code and predictions are released.

A column in The Information looks at why AI startup financings now commonly have two, three or more tranches, with different investors buying in at different valuations. Critics say large VC firms use the structure to push smaller firms into the pricier tranche and to inflate headline valuations. Founders and many investors say it reflects AI companies' unusually large and ongoing need for capital to buy compute. The issue also flags Anthropic's public S-1, including disclosures on compute efficiency, cash burn and its Long-Term Benefit Trust.

Brown senior Cunyan Ma, working with the university’s CNTR center, released a free interactive map linking technology companies and their data-labeling intermediaries to data workers worldwide. It combines existing data-worker databases, Ma’s own data collection and news reports on working conditions. The aim is a systematic picture of low-wage annotation, transcription, rating and content-moderation labor behind AI systems, beyond anecdotal reporting.

Writing in FedScoop, a Code for America author says America.gov, launched last month, is still mainly a chatbot that routes people to government services without fixing outcomes. The author calls for a national model that starts with people, pairs government with outside expertise, keeps humans in control and measures real-life changes. IRS Direct File is cited as a cautionary example: about 300,000 users in 2025 and 9 in 10 pilot users rating it highly, then shut down.

A speculative LessWrong post draws on the folk theorem to propose deterring misaligned AIs. Weights would be tracked even after exfiltration, a mechanism such as compute backdoors would be able to shut down every server holding them, and a daily countdown would default to shutdown unless people around the world override it with proof of life. Any goal-directed schemer would then need humans alive to pursue its goals. The post discusses objections, such as an AI disabling the mechanism, but offers no implementation.

The post defines a strong Turing test: convince even frequent model users that the AI is human. It argues no public model passes, since there is no wave of undetectable agent spam and no reliable style mimicry, even though models now solve open math problems. The author suggests labs deliberately train against duplicity in favor of a recognizable house voice. A well-resourced malicious actor could still train a duplicitous model, for example for election influence, which the author treats as an underrated risk.

Starting from The Bayesian Geometry of Transformer Attention and its “Bayesian wind tunnel” finding that MLPs memorize while transformers learn in context, the author asks whether attention can be defined functionally, independent of the softmax-QK mechanism. That would make it possible to identify contextual learning in other architectures. The post is exploratory and invites discussion. It presents no new experiments.