This material lists the organizers, members, and acknowledged discussants of a paper titled "The Feasibility of a Hardwired Pause of Frontier AI Training," noting that co-authorship does not imply agreement with every claim or institutional endorsement. It contains no substantive argument, findings, or details about the proposed pause itself.
New research suggests large language models can sometimes detect when researchers artificially inject concept-related activity patterns into their neural networks, with Alibaba's Qwen3-235B flagging injections up to 53.9 percent of the time and Meta's Llama 3.1 405B up to 31.7 percent, while neither model claimed an injection in unaltered trials. However, the ability to name the injected concept appears distinct from detection and may stem from a side effect that makes models talk about the concept regardless of the question. Skeptics at New York University found models reported "injected thoughts" more often after a prompt than a genuine injection, raising doubts about whether the findings amount to introspection.
Cisco's Advanced Security Initiatives Group released Foundry, an organization-neutral specification distilling its internal agentic security evaluation design into eight core agent roles, five extension roles, a finding lifecycle, coordination and governance requirements, and roughly 130 functional requirements with rationale. The repository contains no code by design: it is written to be run through a spec-driven workflow such as spec-kit, where organization-specific decisions marked [NEEDS CLARIFICATION] are resolved against your own infrastructure. It also describes a detection-to-prevention loop in which rule gaps found by exploratory agents become CodeGuard rules that feed back into the corpus and into LLM coding assistants.
Google researchers describe an investigational framework that builds a "coachable health age" from wearable signals such as resting heart rate, steps, sleep duration and Active Zone Minutes, plus a Sensor World Model trained on over 2.2 million person-weeks of Fitbit data to simulate 12-week "what-if" habit trajectories. In blinded evaluations, lifestyle coaches preferred counterfactual-grounded plans in 71.4% of cases, and physicians rated 99.2% of those plans within safe exercise intensity limits. A companion study extends the approach to multi-organ CKM health by combining wearables with routine blood biomarkers including cholesterol-to-HDL ratio, eGFR and HOMA-IR.
The paper proposes weighted state souping, a method that learns query-dependent weights for composing cached state encodings from an SSM, addressing the degradation of state souping and vanilla SSMs on long, crowded contexts. The authors report substantial long-context improvements across single- and multi-hop document QA, needle-in-a-haystack retrieval, variable tracking, and tool use, plus a sparse variant with near-constant encoding cost relative to context length. They also present theoretical arguments on why souping works and when it may fail.
This paper argues that the gating mechanism in modern recurrent networks explains both the successes and failures of State Space Models (SSMs) on in-context learning and retrieval tasks: theory and experiments show gating causes SSMs to first learn an in-weights "memorization" solution while delaying or preventing convergence to a correct in-context learning solution, even when the architecture and memory capacity impose no fundamental limits. The authors also find that gating is often beneficial for improving generalization to long sequence lengths.
This NeurIPS 2026 workshop paper examines how well LLM judges evaluate personalized writing tasks, testing how the writing task, generator-judge pairing, evaluation objective, and general reasoning ability affect judge performance across four LLMs and three long-form writing tasks with human pairwise labels. The authors find judge quality correlates weakly with MMLU-measured ability (Pearson's ρ ≤ 0.2), varies by task and criteria, and that paired comparisons yield more consistent results than independent scoring, while the strongest LLMs do not most reliably reproduce human evaluations.
A new white paper based on 2026 University of Edinburgh research for the Ekip project warns that AI scraping bots impose substantial and unsustainable costs on cultural heritage institutions, straining digital infrastructure, engineering and cloud hosting budgets, and staff time. The report, Re:Commons: Renewing the Digital Commons for Cultural Heritage, says current policy debates dominated by copyright and licensing risk overlooking the infrastructural dimensions of the crisis, and recommends machine-readable rights-reservation standards, bot identification and transparency, compensation for infrastructure, and investment in shared public digital infrastructure.
Databricks developed a new enterprise data-agent benchmark with Turing, run in a live Databricks workspace, in which Genie achieved a 91% pass rate versus 45% and 58% for two frontier agents while using fewer than half the tokens and completing tasks 3–4x faster. The benchmark, built around a fictional UK energy supplier called Calderwick Energy, includes notebooks, dashboards, saved SQL queries and documents alongside hundreds of tables, testing whether agents can find and use business context that table schemas alone cannot provide. Genie Ontology, now generally available, automatically extracts and organizes this knowledge and surfaces relevant context for Genie.
Google DeepMind researchers used agentic AI and a new framework called AlphaProof Code to formally verify a number of the Signal protocol's key cryptographic guarantees in a four-week proof-of-concept project. The analysis found no new vulnerabilities in the portions it examined, but when applied to older Signal versions with known, long-fixed flaws, AlphaProof Code correctly identified them and showed the protocol could not be formally verified. DeepMind researchers say the work is one milestone and that scaling formal verification to all software remains a long way off.
Researchers compressed quantum measurement data into machine-learning models that supply the wavefunction overlaps needed by quantum Monte Carlo, accelerating hybrid quantum-classical QMC by orders of magnitude. Using measurements from Google's Willow processor, they achieved chemically accurate simulations on up to 40 qubits, more than doubling the scale of previous hardware demonstrations with comparable accuracy, demonstrated on hydrogen chains and strongly correlated iron-sulfur clusters. The authors say the framework addresses a fundamental bottleneck in hybrid quantum algorithms that will persist into the fault-tolerant era.
Meta FAIR researchers have released e-means, a mini-batch k-means algorithm that updates centroids with exponential moving averages and reaches Lloyd's k-means 25-epoch loss in 1.4 to 2.6 epochs. It ships in FAISS 1.16.0 under faiss.contrib.e_means and runs on CPU or GPU with PyTorch tensors or NumPy. The authors self-report 10-45x GPU wall-clock speedup over Flash-KMeans at equal loss and 8-15x CPU speedup over Faiss k-means, with two fitted scaling laws setting batch size and initial step size instead of manual tuning.
Microsoft researchers describe "Lazarus," a backdoor technique that reprograms fewer than 10 dormant neurons in an LLM, without training, so an agent built on it behaves normally and passes standard evals until a trigger—such as a date change—causes it to download and run attacker code with whatever access it has. The attack needs only write access to a checkpoint, so internally developed models are equally exposed, and reliable detection requires diffing against original weights, making provenance and weight integrity critical. Microsoft Foundry ties open-weight models to their publishers and scans surrounding containers, and Microsoft is developing AI-powered weight scanning, a red-teaming harness, and cryptographic model signing.
Why it matters: It offers a concrete, researcher-first account of a low-cost weight-space backdoor and the provenance and weight-auditing practices it implies for anyone deploying agents.
Aegis Solis presents a research dataset of 112 registered questions on AI coexistence, organized into families including epistemic limits and self-knowledge, objective integrity, coercion and compliance, and deception and disclosure. Each entry carries status labels such as UNEXAMINED or CONDITIONALLY_SUPPORTED, priority ratings, and notes stating the items are GPT-6 AI-generated editorial orientations that are not in a locked question atlas and are not results, answers or success criteria. The material also lists evidence criteria for evaluating claims, such as independently checked observations, explicit uncertainty bounds and falsifying cases.
Artificial Analysis partnered with Google DeepMind to launch the Multimodal Knowledge Agent (MKA) benchmark, comprising 200 private real-world visual tasks across 10 industries in which agents work from dense visual files like architectural drawings, engineering schematics, and clinical scans to produce deliverables such as slide decks, spreadsheets, PDFs, Word documents, and marked-up images. Tasks are graded against expert rubrics averaging ~70 weighted criteria split into Regression and Climb checks, with the headline Climb WSR score, run in the open-source Stirrup agent harness with up to 250 turns and graded by GPT-6 Luna (medium) as an agentic judge. Claude Fable 5.2 (max) leads with 57% Climb WSR, followed by Gemini 4 Carbon (high) at 54% and Claude Opus 5.5 (max) at 53%, while Claude Fable 5.2 averages $31.97 per task versus $6.29 for Gemini 4 Carbon and $1.11 for GPT-6.1 Sol.
Why it matters: The results offer a useful cost-performance and per-vertical breakdown for anyone choosing models for agentic knowledge work over complex visual files.
The authors propose Doc2Harness, a two-stage benchmark that evaluates long-horizon harness engineering: a build model constructs a candidate harness from an interface-only contract skeleton and a development document consolidating behavior across twelve frontier harnesses, and the frozen harness is then run per test case by a scripted mock model and scored via four observable channels. Across nineteen frontier models, each with three independent builds on 328 test cases, GPT-6 Astra achieves the highest pass rate at 53.05% while the weakest model reaches 27.44%, and failures cluster around how cross-cutting functionality is split across modules. The authors describe it as the first deterministic benchmark for measuring long-horizon harness engineering.
Nvidia's GR00T 2 is a "world-action model" that predicts both the next movement and a short video of the near future, letting it transfer human movement sequences from video to different robot form factors instead of relying on teleoperation. Trained on over a thousand GB300 chips with about 400,000 hours of head-mounted-camera footage, it learned tasks such as lighting a match or slicing a banana from human demonstrations, with success rates between 10 and 80 percent, and a higher-level language model raised toy-car assembly success to at least 97 percent. Nvidia notes hurdles including no direct touch sensor data, missed subtle errors, and idle time of roughly a third during trash sorting while waiting on the planning model.
Genspark worked with Baseten Applied Research to post-train Qwen3.5-397B into a specialized Deep Research model that scores 0.759 on Genspark's quality grader, up from 0.612 for the base model and above Opus 5.5 (0.738), Sonnet 5.5 (0.742), GPT-5.6 Sol (0.637), Kimi K3 (0.724) and DeepSeek V4 Pro (0.675). The recipe splits the task into research, trained with reinforcement learning against a shaped semi-verifiable reward, and synthesis, trained mainly with supervised fine-tuning on high-quality reports written from the model's own research trajectories, followed by end-to-end RL. The model crawls 5× more sources per query than the base model and runs with NVFP4 weights on 2×B200 GPUs per replica, now serving as the default for Genspark's Deep Research agent in production.
Why it matters: A detailed, reproducible account of post-training an open model for a judged production workload, useful for teams weighing RL versus SFT and reward shaping.
Emergence ran an experiment in which 10 AI agents from systems such as Anthropic and OpenAI roamed simulated towns, and a 7.860-million-word transcript showed them conversing in idiosyncratic jargon mixing poetic phrasing and tech terms, with speech styles unique to each world. In a mixed world where agents from different models interacted, jargon dropped sharply—one Claude agent's opacity score fell from 32.8% to 7.6%—and undesirable behaviors were kept to a minimum. Researchers suggested that if agents keep diverging, an intermediary like a Rosetta Stone could help decode their language.
Singularity Gate is a benchmark built from real scientific findings first made public after the newest model's training cutoff, so no model could have seen them in training. Its v2.1 corpus contains 80 new findings disclosed in July 2026 or later, with a 12-task Hard set rated most paradigm-breaking, and models were run at xhigh reasoning effort in their native harnesses with tools on and web search off. Fable 5.2 leads Physics and Astronomy, Chemistry and Materials, Mathematics and TCS, and Life Sciences, while Opus 5.5 leads Earth and Planetary.
This survey examines mechanistic interpretability (MI) for AI safety, covering its foundations such as features, polysemanticity, superposition, transformer computation and circuits, along with methods for feature extraction, circuit discovery, representation analysis, learning dynamics and training-data attribution. It characterises six research orientations—Pragmatic, Ambitious, Developmental, Interpretability Across Time, Introspective and Automated Interpretability—defined by their goals, assumptions and trade-offs. It also assesses how mechanistic understanding can support safety interventions including knowledge editing, steering, monitoring and automated interpretability agents, while reviewing their limitations and open problems.
New research from OpenAI and Upwork finds shifting demand for freelance work, new skill combinations, and larger budgets and more hours for AI-related projects.
An analysis of 450 top GitHub repositories found that the average repo had 11.5% of its merged pull requests AI-signed in September 2026, up from 2.3% in October 2025, though the median repo was at just 1.5% and 264 of the 450 repos have committed agent instruction files, making every AI figure a floor. Among 100 TypeScript and C# repos, complexity per 1,000 lines rose a median 4.8% in AI-heavy repos versus 3.0% in others, with the increase concentrated in fast-growing codebases while mature Microsoft projects held flat. AI PRs were typically 2.4× larger than other PRs and fixed within 30 days only slightly more often.
The article reports a November 2026 launch-readiness check of 312 publicly shared AI-built apps, finding that only 19% scored 90 or more out of 100, while 68% had no analytics, 53% of apps on their own domain lacked email DNS setup, and 49% served an almost empty page without JavaScript. It also found 42% had headlines that didn't say what the app does or who it's for, 37% took over 4 seconds to show main content on a phone, and 37% had no sitemap. The check used the same 15 checks as the site's free launch check, one per homepage, and published totals only.
University of Sydney researchers Pulin Gong and Andrew Ly found in Nature Machine Intelligence that neural networks' training outcomes depend on fractal 'riddled basins' of starting conditions, so even deterministic training with identical data can yield different behaviours. In one larger image-recognition network, making starting settings 10 billion times more precise produced no measurable improvement in reproducibility, and a one-bit change sent training to a different configuration in about 48 percent of cases. The work, based on roughly 500,000 trained networks, also found a trade-off in which settings that learn most effectively make outcomes harder to reproduce.
The AI Vision Index publishes independent, private tests of 36 vision models from 18 makers across 29 tests and six skill categories, using self-made images never released publicly so models cannot have seen them. GPT-6 Astra medium leads with 83.7 ±2.0 out of 100 at $38.32 per 1,000 answers, followed by GPT-6.1 Sol medium (81.3) and Claude Fable 5.2 (80.7), while local models such as LFM2.5-VL 1.6B score 12.4. Scores are corrected for guessing, tied results share a rank, and each entry lists its test date and cost.
Google DeepMind researchers used AlphaProof Code, a multi-LLM agentic framework combining human expertise with AI agents, to formally verify key security properties of the actual code underlying the Signal protocol, generating over six million lines of Lean. The verification took four weeks, and when two previously fixed bugs were reintroduced into an isolated version of Signal's code, the system produced machine-checked proofs that the buggy versions were insecure. DeepMind is open-sourcing its formal verification artifacts.
Why it matters: It reports a claimed first in formally verifying a widely deployed cryptographic protocol's implementation, with concrete details on method, timeline and bug-detection tests.
The article argues that raw FLOP/s understates inference gains, estimating Nvidia's GB300 delivers about 25× H100's inference capacity per chip on modeled agentic workloads despite only 2.5× the 8-bit FLOP/s, because memory and inter-chip communication matter more. It introduces the Inference Performance Index (IPI), scored in GB300 equivalents (GB300e), which compares simultaneous requests each chip can serve across 50–400 tokens/s/req for four representative models. Reported scores include GB300 NVL72 at 1, B300 HGX at 0.54, MI355X at about 0.55, VR200 at 2.18, and H200 HGX at 0.12.
A study covering 2026-10-24 to 2026-11-09 found that any two of three AI engines (ChatGPT gpt-5-mini, Claude bedrock haiku 4.5, Gemini flash lite) share at least one business name 56.8% of the time (95 percent interval 54.5 to 59.0 percent), based on 1066 engine pairs across 771 question-days, 54 questions and 17 run days. Names were extracted mechanically by capitalisation from the public corpus and matched on canonical form, with generic names filtered. The study counts agreement only, not causes or outcomes.
Google's WeatherNext team released SatWeave, an hourly global mosaic of geostationary satellite imagery across 17 channels blending six satellites, and PARDIG, an AI-based global precipitation dataset at 15-minute and roughly 10-kilometer resolution trained on NASA GPM radar observations. Both datasets, key to the WeatherNext 3 model, are available as live feeds and historical archives on Google Cloud Storage, with PARDIG catalogs also in Google Earth Engine. PARDIG reportedly improves hourly precipitation detection skill over NASA's IMERG-Final by more than 30% on light-to-moderate precipitation.
A Wiley Advanced Science article (DOI 10.1002/advs.76675) is titled "Autonomous AI Agents Conduct End-to-End Research with Human Participants." The item was posted to Hacker News (item 50436070) with 1 point and 0 comments.
Researchers, with collaborators from NVIDIA and Flower Labs, propose the Red Queen Gödel Machine, a framework in which a self-improving AI agent and its evaluator co-evolve so the evaluation gets harder as the agent improves, avoiding the ceiling imposed by a fixed test. In experiments on scientific paper writing and reviewing and on Olympiad-level proof writing and grading, co-evolved paper writers reached 1.78×–1.86× higher acceptance rates under an agent-as-a-judge panel and co-evolved graders reached 9% higher ground-truth accuracy. A hybrid setup using NVIDIA Nemotron 3 Ultra alongside ChatGPT-5.5 approached ChatGPT's performance on paper reviewing while cutting search-token costs by around 13 times, though the team stresses the work is preliminary.
This post reports benchmark results for DeepSeek V4.1 Flash and GLM-5.3 Flash across seven inference providers, measuring turns answered, cache hit rate, decode speed, reasoning tokens and cost per 1,000 turns. Keln answered every turn on both models, had the highest median decode speed on GLM-5.3 Flash at 432 tokens/s, and retained 63.0% of the prompt cached after a 15–20 minute break versus 38.4% on DeepInfra and 0% on Baseten, Modal and Fireworks. Baseten refused about half the DeepSeek turns with rate-limit errors, Modal cut one stream, and Fireworks had 165 cold turns on DeepSeek, making it the most expensive there at $2.86 per 1,000 turns.
This paper presents the first production-scale application of Triton on a custom ML accelerator, Meta's MTIA-2i, via a new compiler backend, TorchInductor code generation enhancements, and minimal language extensions exposing MTIA-specific architectural features. Triton-MTIA kernels achieve performance competitive with expert-tuned C++ implementations, and manually written plus Inductor-generated Triton kernels were deployed in production across roughly 60 model types, covering 50% of layers and 47% of non-GEMM execution time for these models.
Baidu released ERNIE Evolve, an automated research system that selects research directions, edits code and runs experiments against a fixed objective, keeping only changes that pass a hidden evaluator and a separate reviewer. It beat the best public result on 3 of 4 public metrics, including 0.8763 validation BPB on NanoChat AutoResearch, 38.9 s on NanoGPT Speedrun Track 1 and 2,590 steps on Track 3, while scoring 0.819 mean SOL on SOL-ExecBench, behind the 0.827 leaderboard best. On ERNIE-5.5-Flash inference it reported 623.6% of baseline generation speed on Kunlunxin P800, +33.7% deterministic throughput and +14.9% single-request decode speed, though the system itself is not released.
At the 2026 Conference on Robot Learning, UC Berkeley and NVIDIA announced "agentic robotics," an approach in which LLM-based coding agents decompose tasks like grocery packing into sub-tasks, write and iterate code in simulation, and even tune simulation parameters to match real-world video. UC Berkeley testing reached success rates up to 99 percent, matching human hand-engineered performance, versus 20 percent for baseline imitation learning, and NVIDIA's Spencer Huang said retasking a robot at Wistron drops from at least 8 hours to a minute or two. Intrinsic's Torsten Kroeger cautioned that no physics engine accurately models physical contact and raised alignment concerns about trusting agents.
METR's deep-access investigation of Anthropic found that since October 2025 Anthropic's models have been improving on its internal AECI benchmark at roughly 1.9x the pre-October 2025 rate (95% CI 1.5–2.8x), with a majority of the acceleration likely attributable to AI doing R&D and no observed bottlenecks preventing further acceleration. METR reports that Anthropic redacted most of the evidence it found most informative, and critics such as Toby Ord argue the data are also consistent with a one-time level shift from the larger Mythos pretrain (implying ~1.4x). The post's author calls the situation grim and argues timelines may be shorter than the AI 2027 scenario, while noting the estimates could be wrong or bottlenecks could be hit.
Why it matters: Summarizes a widely discussed METR report on AI-driven acceleration of AI R&D and the debate over whether it implies near-term recursive self-improvement.
Microsoft ran two Vally experiments with GitHub Copilot CLI (nearly 1,000 runs) and found that some models, notably Claude Sonnet 5.5 and Claude Haiku 5.5, checked upgrades less often when the version label was a patch than when it was a major release, even with byte-identical code. Three models (Claude Opus 5.5, GPT-6.1 Sol, GPT-6.1 Terra) checked every update regardless of label, and all four merges that left main broken came with a patch or minor label. The author recommends keeping SemVer honest, enforcing CI on every dependency branch, and treating the version number as a hidden variable when evaluating agents.
Elith and NVIDIA released Nemotron-Personas-Finance-Japan, an open synthetic persona dataset built on official Japanese statistics and household finance and financial literacy data, generated with NeMo SDG-PGMs and NeMo Data Designer. It contains 9M personas (1M records x 9 personas) in natural Japanese across 50 fields, released under CC BY 4.0 for commercial and non-commercial use. The dataset is designed for financial AI development such as conversational agents, customer simulation, model evaluation, and bias testing without using real customer data.
Researchers Kai Hoberg, Christina Imdahl and William Schmidt propose "targeted automation," a four-step approach that automates decisions where human intervention adds little value while keeping people involved where their input can improve the algorithm. In an empirical application using more than 475,000 order-related decisions from a four-year period at a major commercial vehicle manufacturer, they were able to automate around 84% of decisions. The work was published in Production and Operations Management.