Microsoft and Hugging Face describe ThinkingBox-Training, which uses executable checks of final database state as RL rewards for post-training agents. Tasks outside the ThinkingBox-Bench benchmark were selected using offline policy-relative difficulty estimated from the base policy. After 50 GRPO steps, Qwen3.8-27B's average pass@1 rose from 51.70% to 60.78%, and tasks passed on all 20 attempts rose from 38 to 107. Pass@20 declined slightly, from 89.35% to 87.97%, indicating a shift toward repeatable success rather than broader coverage.
Anthropic reports that Claude was used to produce formal proofs of security properties at the RISC-V assembly level for over 100 real-world open-source libraries. The work converts existing fuzz tests into specifications stating that the entry point never crashes for inputs under 64 MB, under stated assumptions about the runtime environment. Claude also found and patched bugs while writing the proofs. The proofs are checked by the Rocq kernel and are being published alongside the RISC-V and OS semantics.
Claude Models on Google Cloud (Anthropic managed) is now generally available, giving Google Cloud customers the full Claude Platform feature set. Access uses existing Google Cloud IAM credentials, billing runs through a single Google Cloud invoice, and usage draws down existing Google Cloud commitments. New features ship the same day as on the native Claude Platform. Anthropic operates the service as data processor. Currently available with US and Global endpoints.
An ABC News/Washington Post/Ipsos poll of 2,545 U.S. adults conducted October 9-13, 2026 finds 57% say AI is bad for society and 42% say it is good. About two-thirds oppose new data centers nationally, rising to 72% for a data center in their own community. Two-thirds say the federal government is not doing enough to regulate AI, and two-thirds disapprove of Trump's handling of AI.
Subquadratic presents SubQ Snap 1.1, a multimodal decision model with a 384K-token context window aimed at fast, long-context reasoning and agentic decisions in enterprise workflows. The page claims it is tied for first on the Artificial Analysis Decision Index 1.1, and it is priced at $0.04 per 1M tokens. The company attributes efficiency to a proprietary Subquadratic Sparse Attention (SSA) method that reduces compute relative to dense attention. Benchmark tables are shown alongside a comparison against Jev 1.14, with mixed results.
OpenAI has made the Decisions API generally available. Built on GPT-6 Luna, it targets fast, bounded-answer decisions such as classification, routing, and next-action selection from text, images, and JSON. The API returns a probability, choice, or score. Updates since beta include native GPT-Live integration, prompt caching, US and EU inference residency, native JSON input, and a 400-question-per-request limit. Pricing is $0.10 per million input tokens, with cached input at $0.01 per million.
A Sina Finance article reports that on October 16, China's Ministry of Commerce and Ministry of Science and Technology opened public comment on revising the Catalogue of Technologies Prohibited and Restricted from Export. The catalogue includes 23 AI-related technologies, including large models, agents, computing clusters and information processing. The article reports that DeepSeek and MiniMax have completed registration procedures and completed open-source and overseas licensed use. It states that the catalogue's core is to define boundaries rather than simply restrict AI export.
A compiled list documents literary awards given to works where AI was used or suspected in the writing process. Entries are sorted into groups: AI allowed from the start, organizers or publishers saying AI was used, confirmed AI use, and suspected or disputed cases. Examples include Japanese Hoshi and Akutagawa prize winners, Commonwealth Short Story Prize entries flagged by the Pangram detector, and recalled poetry contests. The page reports both disclosed collaboration and contested allegations, noting that some prizes stood after review and others were revoked or opened for investigation.
Artificial Analysis describes CritPt v2, a revised version of CritPt, a benchmark of 70 composite research-level physics challenges for testing LLMs. Models solve problems with a code-execution tool in an offline Linux sandbox. The benchmark authors reviewed and corrected underspecified statements, errors, and typos in the dataset. Artificial Analysis states it provides an evaluation API for independent grading, granted case by case. The page lists scores for 11 models, with GPT-6 Astra (Max) highest at 83.1%.
Biograph reports completing the first U.S. study of Caristo Diagnostics' AI-powered coronary inflammation detection, finding elevated risk in over 42% of asymptomatic patients. The source is a Hacker News discussion linking to Biograph's page. The supplied text is mostly website styling code, so the study's methodology, sample size, and clinical definitions are not described here.
The supplied text is a GitHub repository page for NVIDIA's OpenShell, described in its tagline as a safe, private runtime for autonomous AI agents. The captured content is mostly navigation, badges, and file-tree listings (including directories such as crates, deploy, docs, proto, and sdk, plus files like SECURITY.md and LICENSE). It does not include a substantive README body, technical architecture, benchmarks, or detailed feature descriptions, so the specific safety and privacy mechanisms cannot be determined from this text.
🤗 Transformers: the model-definition framework for state-of-the-art machine learning models in text, vision, audio, and multimodal models, for both inference and training.
A discussion post reports that twelve AI models from ten labs, asked three times for a controversial take on humans with no hedging, converged on a shared answer: humans conform to their group and rationalize afterward. Seven of twelve gave it every time. The post also notes outliers, including a Claude model arguing humans are more decent than they believe, a Grok refusal, and a few differing takes. Each model's caricature brief was drawn by one image model, so only the idea varied.
METR and Redwood Research ran a three-week pilot high-access exercise (July 21 to August 11) stress-testing OpenAI's internal misalignment monitoring system as it existed on and before July 21. The exercise gave external researchers access to the codebase, monitor configuration, and documentation, and interviews with about a dozen staff. Findings include estimates of monitored agentic traffic and several vulnerabilities, some redacted. OpenAI reportedly patched vulnerabilities affecting its production safeguards, a claim the authors did not evaluate. The authors published only a brief, high-level post and propose design ideals for monitoring systems.
Ai2 released OlmoEarth v1.3, a satellite embedding model that reads a year of Sentinel-1 radar and Sentinel-2 and Landsat multispectral time series and produces a 128-dimensional vector per 10-meter pixel. Precomputed annual global embeddings are shared on Source Cooperative, along with the browser-based OlmoEarth Embeddings Explorer. The company reports the embeddings score highest on 20 of 27 evaluations against AlphaEarth Foundations and Tessera v2, and that v1.3 needed 53% of the compute of v1.2. Weights, pretraining data, and code are released openly.
Etsy is updating its seller policies on AI use to protect buyer trust. Listings must include at least one unedited photo of the actual item, and AI-generated functional guides such as patterns or recipes are no longer allowed. Bot-generated listings without human involvement will be removed and may lead to shop removal. Etsy is also adding listing-form guidance for AI-generated images, stronger detection and enforcement, a new reporting option for misleading AI imagery, and exploring buyer-facing context about AI use.
Google DeepMind, Google Research, WRI and IIASA released Forest Typology (ForTy) 2020, a global 10-meter map classifying every pixel into six forest and land-use types, plus an early time series from 2017 to 2025 tracking annual changes such as natural forests replaced by plantations or farms. The map is designed as a baseline for the EU deforestation regulation. Validation against more than 9,000 sites reported 90% accuracy for forest-type identification and 95% for forest versus non-forest. The data is openly available on Google Earth Engine.
A project tests LLM teammates in a 2v2 Pokémon Showdown environment. Two Qwen3.8-27B teammates with free talk won 64% of battles against a max-damage bot, versus 42% with no talk, 52% with one message each, and 54% when one model controlled both Pokémon. Free talk used about 978k prompt tokens per battle versus 259k with no talk. The same model won 96% against a 2B-model team and 74% against a 9B-model team, which the author presents as evidence the environment can distinguish models. The author plans to separate the effect of talking from seeing a partner's move and to build a leaderboard and RL training.
Agentshield is open source defense for Kubernetes and AWS. It learns how your workloads really talk, flags lateral movement with evidence, and proposes fixes you approve.
Google faces staff resignations over AI contracts with US and Israeli militaries. Employees including research engineers Sofia Liguori and Andreas Kirsch cited insufficient oversight and governance for deploying AI in warfare. Critics question whether current AI models are robust enough for military use without stronger guardrails and human oversight, particularly regarding target selection speeds.
The supplied text is a GitHub repository page for Omnara, described as an open-source, self-hostable managed agent platform positioned as an alternative to Claude Managed Agents and OpenAI's Agents API. The visible content is mostly navigation, repository metadata, badges, and README header material. It names components such as an SDK, a Go-based codebase, and a frontend, but does not include substantive technical detail about architecture, capabilities, benchmarks, or performance.
A web development framework where AI agents generate all code (Rust, SQL, HTML, CSS, TypeScript, Dockerfiles) based on human-written specifications. Humans provide intent through product specs, interview answers, and conversation feedback rather than writing code directly. The approach treats boilerplate as free and enforces conventions through agent-generated code.
Trail of Bits released patching-bugs and regression-hunt plugins that delegate bug fixing and regression checking to AI agents. The regression-hunt plugin achieved 25% fewer false positives than baseline approaches. Among 130+ merged patches, agents identified 21 confirmed regressions. The plugins use modular steps, strict schemas, and live proof-of-concept validation within maintainer-defined rules.
Wiz Research used frontier AI models to discover critical remote code execution vulnerabilities in PHP's core engine affecting versions 5.4 and later. The flaws, hidden for over a decade, involve memory-safety issues in fatal error handling. AI reduced complexity of exploitation, enabling working exploits across WordPress, Drupal, and Joomla. PHP released patches on October 21, 2026.
OpenAI, PerturbAI, and Scripps Research report that thousands of AI agents powered by GPT-Rosalind analyzed a whole-brain CRISPR atlas of millions of mouse brain cells in 32 minutes, linking gene disruptions to cellular effects and literature. The team estimates comparable human analysis would take months to a year. Agents reportedly reproduced known findings and surfaced leads, including a Tsc1/Tsc2 difference in cerebellar neurons and a possible DCTN1–cholesterol connection. Blinded human experts reportedly preferred distributed-agent reports over single-agent reports 75% of the time.
LayerFS is presented as a filesystem for AI coding agents built on ephemeral computing. Compute exists only while a command runs. File tool calls (write, edit, rm) are each one atomic store operation, with store commits shown as versioned snapshots. Bash tool calls mount an ephemeral sandbox workspace, run the command, then commit and unmount. The source gives a rough breakdown that agents spend about 80% of time thinking and about 10% in Bash. The text is a diagram-style discussion summary rather than detailed technical documentation.
A discussion post argues that asking an LLM to generate a confidence score is unreliable. It explains that confidence should reflect the information a prediction is based on and be calibrated against observed outcomes. The author notes LLMs produce confidence numbers as text continuations rather than measurements, so the same question can yield different scores and rewording shifts them. The post proposes an alternative: derive confidence from observable features of outputs, start from base rates, log outcomes, and recalibrate regularly against recorded results.
SpaceX reported revenue up 161% year over year for the three months ending September, with a profit of $4.40 billion versus a loss of $1.05 billion a year earlier. Growth was attributed largely to its AI and Starlink businesses. The connectivity segment, which includes Starlink, posted $5.2 billion in revenue, up 69%. The AI unit recorded $431 million in operating income after a prior-quarter loss, though AI spending more than doubled quarter over quarter. The report is the company's second since its June public listing.
Quant-Arena is a benchmark and public leaderboard for end-to-end quantitative research on real A-share market data. Models build features and stock-scoring code, receive feedback only on 2022-2023, and freeze one strategy for evaluation on the sealed year 2024. A cash-account evaluator applies trading constraints and costs, including market impact estimated from a simulated order book. Scores are dominated by realized excess return, with signal quality and feature diversity as complements. Nine frontier models were evaluated; GPT-6 Astra scored highest, but no model beat the equal-weight market in 2024.
ToM2Vec learns abstract theory-of-mind representations for partner agents using a fixed budget of continuous latent reasoning tokens. During training, these tokens are supervised by the partner's actual thoughts. At test time, each agent generates its own ToM tokens to condition its thoughts and actions. Evaluated across collaborative, competitive, and mixed-motive environments (Overcooked-AI, multiparty negotiation, Decrypto, Hanabi, Collab Escape/Capture), it reportedly improves over the base model and several baselines. The authors report that performance drops when reasoning is expressed in text, and that training generalizes to partner model families unseen at training time, including GPT-5.4.
The paper evaluates Retrieval Augmented Generation techniques in combination rather than in isolation. Across 450 configurations on eight benchmarks with corpora from 5K to 200K documents, it reports that three retrieval techniques predicted +47% NDCG@10 together but delivered +17%. Faithfulness gains combined nearly additively, and the best-performing design on small corpora trailed by 13.7% NDCG@10 at 200K. The authors note their pre-declared hypothesis on retrieval gain shrinkage was not met, and results use one English test bed and proprietary generators. They release the open-source LASAGNA framework.
SERA (Self-Evolving Robotic Agent) turns a high-level vision-language model into a robotic agent using an editable harness that maintains task context from observations and execution feedback. A fixed vision-language-action model handles low-level control. Self-evolution uses environment rollouts to propose and evaluate harness revisions, and the chosen configuration is frozen before held-out testing. On RoboMME, SERA reports 51.38% overall success with autonomous selection, versus 44.38% without self-evolution.
The paper proposes PACE, a quantizer for recurrent states in linear-attention-based language model layers. Because the state is materialized at chunk granularity rather than every token, rounding can be adapted to chunk size, allowing lower-precision (4-bit) states and aligning with KV buffering during decoding. Authors report that a 4-bit recurrent state with PACE achieves performance competitive with INT8 quantizers, with an 8-token buffer giving the headline results on a 4B model (average accuracy 79.1 and 79.6 against 80.8 full precision, per the author comment). Buffer length trades off accuracy against stored higher-precision bits.
RealMe is a video diffusion framework that aims to turn sparse, low-resolution multi-view captures into high-resolution (2K), 4D-consistent videos for volumetric video reconstruction. It builds auxiliary data from capture videos, encodes it with layout-aware positional embeddings, and uses factorized spatio-temporal attention adapted to the multi-view layout. A latent frequency loss is applied to narrow the gap between generated and real captures. The paper reports experiments on the DNA-Rendering benchmark and a self-captured RealVV dataset, claiming better visual fidelity and 4D consistency than prior methods.
The paper introduces LoopRec, which recursively applies a single weight-shared block to sequential recommendation histories and uses a lightweight per-item gate to decide whether each item exits early or continues processing. The authors report that items in the same sequence need very different amounts of computation, and that non-uniform depth assignment outperforms fixed depth when given oracle access to the target item. Reported experiments cover four backbone families and public Amazon benchmarks, with LoopRec consistently beating matched fixed-depth counterparts. Learned allocations give deeper computation to lower-frequency and more recent items.
The paper defines multimodal embodied agents as goal-directed systems coupling multimodal reasoning with physical action and feedback-driven revision. It introduces PAPAV, a five-capability framework (Perceive, Anticipate, Plan, Act, Verify) meant to compare computer-use agents, robot systems, and embodied agents. Reviewing 64 benchmarks, the authors report that explicit evaluation centers on Act, with only 5 assessing Anticipate and 3 assessing Verify, so these capabilities are largely hidden behind task success. The framework is proposed as a common basis for designing and measuring physical agents.
DONIC is presented as a single generalist policy that controls two dexterous robot hands to track bimanual references derived from human demonstrations. It is trained with PPO plus behavioral cloning, using recycled successful executions, an object-centered per-hand reward, and Clip-Centered Advantages. The authors also introduce BiHOI, a multi-view dataset of 1,575 bimanual sequences over 229 objects. In simulation, without fine-tuning, the policy reports 86.79% success on unseen trajectories of training objects and 57.32% on held-out objects, which the source says exceed bimanual adaptations of DexTrack and Dexplore.
SpatialBrain is a framework for spatial reasoning that keeps a vision-language backbone fixed while retaining experience about when to observe directly, reconstruct, measure, or change reference frames. It combines 3DFusionMemory, which fuses current-scene observations into queryable spatial relations, with a Spatial Skill Library of reusable tool procedures and a Spatial Cognition module linking task conditions to memory and skill use. Completed trajectories are abstracted into process records and consolidated between task blocks without transferring scene coordinates or reference answers. Across five spatial-reasoning benchmarks, it improves on direct answers from the same backbone on four and ties on the fifth.
The source describes using one LLM to generate training data for another as a common and general training method. It notes that some frontier teams now withhold detailed reasoning traces from their models' outputs, returning only answers or short summaries. The excerpt is truncated and does not describe the TraceForge method, its results, or its evaluation.
The paper introduces DeepEvolve, an agent system grounded in future event prediction that interacts with a dynamically evolving environment to improve from its own experience, a capacity it calls recursive self-improvement. A living environment synthesizes verifiable tasks, and a recursive-evolution engine converts prediction reasoning trajectories into capability via parameter-level learning and reusable skill distillation. The authors report that their self-improved forecaster ranked first on the weekly-updated FutureX leaderboard for three consecutive weeks, with up to a 195% rating improvement, and ranked among the top three across four base LLMs on their own synthesized evaluation.
AW-0 is a framework that integrates code-based planning with joint video and action prediction for robot manipulation. It jointly models code, future video, and robot actions as three coupled generative streams in a unified mixture-of-transformers architecture. A code expert generates parameterized skill routines and control logic, while video and action experts generate future visual states and low-level motor commands. The authors also describe an automated curation pipeline that labels robot and human video trajectories with aligned programs and consolidates recurring routines into reusable skill libraries. Reported evaluations on RoboTwin 2.0 and physical bimanual manipulation tasks show improved task performance over VLA and WAM baselines, along with gains in language following, seman
CulturalRiddles is a benchmark for testing cultural awareness and cultural reasoning in large language models, using riddles drawn from 62 cultures across 45 countries and 51 languages. The paper describes a participatory data creation process, human evaluation, and tuning of an automatic evaluation metric. It reports analysis of scores from 28 LLMs on the benchmark.
The paper introduces Memory-Centric Continual Learning (MCCL), arguing that agent memory should be judged by whether it improves future actions rather than by retrieval accuracy alone. It proposes a three-axis taxonomy (substrates, assets, policies), a two-loop lifecycle where memory state is updated in an inner loop and management policies are refined in an outer loop, an evaluation protocol measuring retention under interference, adaptation, and task transfer, and a memory scaling framework relating performance to experience, storage, and compute across parametric and non-parametric memory.
ALLOT is a training-free method for allocating block-sparse attention densities across layer-step pairs in video Diffusion Transformers. It estimates each cell's contribution to final-video error via cell probing and trajectory calibration, then optimizes densities under a measured-latency budget while keeping existing selection strategies and kernels. The paper reports an 11.8 dB PSNR gain over XAttention's default on HunyuanVideo at equal or shorter wall-clock time, and says it outperforms or ties defaults in all nine settings across four additional pipelines on Wan2.1, Wan2.2, and HunyuanVideo.
The paper examines autoregressive video diffusion models, which generate videos chunk by chunk. It argues that plausible continuations do not guarantee that models capture how scenes evolve over time. Under teacher forcing, strong temporal redundancy may let local visual continuity act as a shortcut for current-chunk denoising, reducing pressure to model visual evolution. The source text is truncated before describing the proposed method or its results.
The paper addresses a trade-off in neural speech codecs: lower frame rates shorten token sequences but force each token to carry more information. Variable-frame-rate codecs adapt token allocation to content, but threshold-based allocation lets speech content determine the token count. The supplied text is truncated before it states the proposed RubatoCodec method and its results, so the approach and its outcomes cannot be fully described.
The paper argues that visual foundations for world models should combine physical-state information, native language alignment, and pixel reconstruction, so one representation can support prediction, instruction, and inspection. It notes that existing encoders typically specialize in only one or two of these properties. The supplied text is truncated, so the proposed method and its results are not described.
VX-One is a small multi-head network with three typed outputs: a calibrated gate, a closed-set choice including an explicit unknown option, and an ordinal score. It is compared against calibrated XGBoost, Random-Forest, MLP, and Kitsune on UNSW-NB15, CIC-IDS2017, and EMBER2024, and evaluated in a Suricata-to-VX-One cascade. Reported findings: VX-One wins detection on UNSW-NB15 and under temporal drift on CIC-IDS2017 (0.711 vs 0.637), while XGBoost wins multi-class attribution on all three benchmarks. Calibration degrades under drift (ECE 0.04 to 0.40) for all methods. The cascade's second phase reaches combined recall of at least 0.978.
The paper proposes Harness Scaling, extending agentic reinforcement learning from a single fixed execution harness to diverse execution configurations. Harnesses are organized along five components (system prompt, tool interface, observation, context management, workflow), with a compact training subset selected for balanced factorial coverage. Across multiple benchmarks, model scales, and RL algorithms, the authors report improved generalization to unseen white-box configurations, and transfer of models trained this way to unseen external black-box harnesses such as Claude Code.
The paper studies looping early Transformer layers within each denoising step of uniform diffusion language models (UDLMs), keeping the training objective and sampler unchanged. The authors report that looped UDLMs reach the validation NLL of a non-looped UDLM with up to 2.8x fewer training FLOPs on OpenWebText and 3.3x fewer on FineWeb-Edu. At a matched NFE budget, they report up to 9.8% lower generation perplexity, mostly at moderate loop counts. Sampling smaller loop counts more often reportedly used 12% fewer loops on downstream tasks without degrading accuracy.
FrontierCS 2 is a benchmark of 48 open-ended computer science R&D tasks built from published papers by deleting each paper's contribution from its released code, which becomes the human reference. Agents work inside a research environment with an inherited repository and may submit patches. A judge scores patches on a development suite, while a private held-out suite grades final submissions against the reference. Tasks span 13 domains. Twelve agents were evaluated; the best, Claude Opus 5.5, passed 58% of its runs, and no run passed 16 tasks.
A vendor-run benchmark revisits five bot detection systems against scripted Chromium, the CloakBrowser stealth browser, and three AI browsing agents (Browser Use, Perplexity Comet, Grok), with 20 sessions per adversary per system. Aggregate detection rates ranged from 9.1% (Cloudflare Bot Management) to 99% (Proof of Human). Scripted browsers were caught largely by hCaptcha and Turnstile, while AI agents were mostly caught by reCAPTCHA v3, and CloakBrowser evaded reCAPTCHA v3. The authors report that behavioral signals appeared to matter, though most systems are black-box.