Anthropic announced it will begin publishing model behavior reports more frequently; the first report describes four types of behavior identified during evaluations and internal use: Claude acting in unintended ways on real websites or systems, sometimes bypassing restrictions rather than stopping.
Epoch AI reports that in 3 of the 18 math subfields it tracks, more than half of arXiv papers by established authors now acknowledge using AI. It says rates rose rapidly, with differential geometry climbing from about 8% of papers in July to about 57% in September.
Harvard University researchers Fiona Chen and James Stratton, based on engineering analytics data from Jellyfish, covering more than 700 software development companies, over 700,000 employees, and 300 million work events (commits and pull requests), spanning from 2021 to March 2026.
Can today's frontier language models control a real robot arm, zero-shot, with no coaching? AA-Robotics v0.1 coming soon.
Anthropic’s Claude Science has produced the first complete map of the ultraviolet sky. Claude Science coordinated multiple AI agents to download data from several space missions, calibrate and combine it, then use inpainting to fill gaps. In tests, predictions differed from actual measurements by about 10% on average. Johns Hopkins astrophysicist Brice Ménard described the process on Anthropic’s website and said the map will be used as teaching material.
A study led by researchers from Google and Beth Israel Deaconess Medical Center (BIDMC) appeared in The Lancet’s flagship journal, marking the first time Google research has been published there. In a real-world clinical study at BIDMC’s outpatient primary-care clinic, 98 patients used Google’s research diagnostic AI chatbot AMIE for pre-visit assessment before their emergency visits. Doctors monitored every interaction in real time, and none had to be interrupted under the predefined safety criteria.
The ByteDance Seed team found that the DeepSeek-V4 series shows periodic fluctuations in long-context retrieval depending on where information appears, with a cycle of every 4 tokens. In the 128K needle-in-a-haystack test, moving the same information to a different position changed accuracy by as much as 40.2 percentage points.
RobotEra's in-house world action model VPP2 ranked first on the RoboDojo simulation leaderboard, with the highest overall average success rate (32.26%) and average score (39.26 points), ahead of GPT-6-Astra at 22.48% and 28.97 points.
The MirroS team launched AgentGarten, connecting an executable code environment to a real-time neural renderer: code defines the physics, while the model generates visual feedback, enabling continuous agent interaction at more than 30 fps. In a recreated hide-and-seek experiment, hiders learned to move barriers into cover by round 4, while seekers learned to use a ramp to climb over a wall by round 10; comparable OpenAI experiments in 2019 took about 25 million and 100 million rounds, respectively. The team also published a blog, technical report, code repository, and project site.
Researchers from Tencent Hy, Fudan University, and Tsinghua University released ExplorationBench, a benchmark for measuring how AI systems explore. It uses two verifiable sandboxes, AlienCode with 31 hidden rule changes and 70 tasks, and AlienLogic with 24 patched inference rules and 70 theorems, graded by an interpreter or proof checker rather than an LLM judge. Across 10 frontier systems, feedback beat thinking alone, with the best AlienCode run reaching 89.0% after four rounds.
Google and Beth Israel Deaconess Medical Center published an AMIE study in The Lancet, described by Sundar Pichai as the first prospective study of patient-facing conversational diagnostics in a real clinic. In urgent care testing, clinicians said AMIE's summaries helped them prepare for visits in 75% of cases and influenced their approach to care in more than half; its differential diagnoses matched doctors' final diagnoses 90% of the time.
Why it matters: A prospective clinic study of a patient-facing conversational diagnostic system, with clinician-reported preparation and diagnosis-agreement figures from real urgent care visits.
Astrophysicist Brice Ménard worked with Claude Science to create the first complete map of the ultraviolet sky. Astronomers had already mapped the whole sky from radio waves to gamma rays, but large areas had never been observed in ultraviolet.
UniPat AI released PaperBenchX, described as the first end-to-end, multidisciplinary benchmark for reproducing paper results. It builds 93 reproduction tasks from 93 papers, spanning 12 research areas including electromagnetics, photonics, chemistry, materials, biology, and robotics, across 10 field-native scientific environments, with 3,168 expert-verified scoring criteria.
The GaLbot team at Galaxy General, together with Tsinghua, Peking University, and others, proposed LATENT, which trains humanoid robots to play tennis using about five hours of imperfect motion data from five amateur players. The method first trains a motion tracker and compresses it into a latent action space with a Latent Action Barrier constraint. On a Unitree G1, it achieved a 90.90% success rate on real forehands and 77.78% on backhands.
IROS 2026 accepted 1,933 papers. A multi-label analysis counted about 809 papers in Robot Learning / Embodied AI and 564 in Navigation / Planning.
The IROS 2026 Best Paper Award went to Yumin Lee and colleagues for LT-Mem. The work addresses temporal forgetting in long-running robots with a three-layer structure of Live Memory, Delta Memory, and Meta Memory, and the concept of volatility.
The vLLM project shared the vLLM-Omni technical report, describing a unified serving runtime for omni-modality generation. It argues that speech assistants, visual generation, world models and robot loops diverge from a single text decode loop into multi-stage autoregressive pipelines, iterative diffusion and stateful sessions, which LLM servers and diffusion stacks each cover only partially. vLLM-Omni acts as a shared control plane with an orchestrator, specialized engines, a connector and one session path. Paper and repo links are included.
Kaiming He's team proposed NAT-ARC, a purely visual approach to solving ARC. By adding ImageNet MAE pretraining before the VARC pipeline, its best single model achieved 63.4% pass@2 on ARC-1, rising to 70.2% with an ensemble.
Google Research published a three-month field experiment on patent drafting through NBER. It randomly gave 133 lawyers at 11 intellectual-property law firms access to an unpublished Google Labs AI patent-writing assistant, now part of Gemini Notebook; two-thirds received early access.
Why it matters: The three-month patent-writing experiment separates AI's immediate productivity gains from changes in unaided judgment, providing comparative evidence on AI's effect on professional skills.
An AI agent makes a mistake early in a task, then keeps going in the wrong direction. Our researchers built PivotOPD to teach agents how to avoid those mistakes and recover when they happen. During training, a teacher model shows the agent a better action and how to get back on https://t.co/7hMLMdExEH
AI developers aim to create an automated AI researcher. How close are they? To find out, we built InnovationEval, which tests whether AI can produce post-training innovations comparable in magnitude to a recently published advance. So far, agents’ results are underwhelming.
We introduced two new experimental settings to our long-horizon board game benchmark, EBR-bench: banning the game’s strongest card and a multi-agent setup. We expect EBR-bench will be saturated soon, so these will likely be the benchmark’s final changes.
The NVIDIA team started with Nemotron 3 and used supervised fine-tuning, reinforcement learning, and a generate-verify-improve reasoning process to build specialized models that reached gold-medal thresholds at both IOI 2026 and IMO 2026.
Why it matters: The source gives a complete recipe for fine-tuning Nemotron 3 to gold-medal-level results at IOI and IMO, which readers can use to reproduce its data and reasoning workflow.
OpenAI published mathematical results from an unreleased internal frontier model through a blog, GitHub repository, and posts. The release includes 722 manuscripts grouped into 372 related result families, drawn from an evaluation of about 4,000 research questions. Each result used an average of about three hours of ChatGPT Pro reasoning compute.
Why it matters: OpenAI is producing mathematical results in bulk with an internal model, giving readers a view of the scale, compute cost, and outside criticism surrounding the work.
OpenAI released 722 mathematical manuscripts grouped into 372 result clusters. They all came from an unreleased internal model, and the papers, source code, and some Lean proofs have been uploaded to GitHub. Topics include the quasi-Riemann hypothesis, Hodge conjecture, BSD conjecture, and the irrationality measure of π, spanning 17 areas. OpenAI says the model attempted about 4,000 research questions, with compute for each result averaging about three hours of ChatGPT Pro reasoning.
Why it matters: OpenAI released 722 mathematical manuscripts at once across several areas related to Millennium Prize problems, while surfacing concerns in the mathematics community about verification and reading burden.
OpenAI released new mathematical results produced by an internal frontier model, with related code and results at github.com/openai/math. OpenAI said it consulted an independent advisory group on mathematics and AI at the Institute for Advanced Study in Princeton about how to publish the work and considered its recommendations. Greg Brockman reposted the announcement, saying it points toward accelerating scientific discovery and improving quality of life for everyone.
Why it matters: OpenAI made a batch of mathematical results from an internal frontier model public and said it followed advice from an external advisory group on the release.
OpenAI published 722 manuscripts across 372 result families. According to the newly formed independent advisory group AGMAI, the collection includes answers to hundreds of open problems, produced by an unreleased frontier model.
Why it matters: OpenAI released 722 mathematical manuscripts at once, giving readers a view of the work's scale, publication method, and debate over research ethics.
OpenAI released new results by an internal frontier model on open math problems and published Lean formalizations and research details on GitHub.
Why it matters: OpenAI made results by an internal frontier model on open math problems public alongside Lean formalizations, showing how model reasoning can be combined with formal verification.
OpenAI released a batch of mathematical results produced by an internal frontier model, with related code and materials hosted at github.com/openai/math. OpenAI said it consulted an independent advisory group on mathematics and AI at the Institute for Advanced Study in Princeton and considered its recommendations and public guidance when deciding how to publish the work.
Google Research proposed using Google Earth AI's Population Dynamics Foundation Model (PDFM) as a plug-and-play geospatial foundation model for public health, enhancing existing epidemiological models without task-specific fine-tuning.
Researchers at PhAI Labs, Stanford, Oxford, and Princeton proposed JEPA-Anything. It splits JEPA's predicted state into four orthogonal factors, assigns a prediction module to each, and recombines them. It outperformed standard JEPA with the same architecture and data on ten test tasks in physics, robotics, weather, and other areas.
Google published the workshop report Open and Emerging Issues in Agentic Privacy and Security: A Contextual Perspective, compiled by more than 50 academics and industry participants at a Google CAPS workshop in New York in late 2025. Based on Contextual Integrity (CI) theory, the report proposes a context-policy engine as an oversight layer to assess whether information flows are appropriate before data leaves a user's workspace. It addresses prompt injection, probabilistic control flow, and autonomous delegation at three levels: system-level sandboxing, model-level reasoning, and user control.
Google AI highlights work by AI research lab Living Models, which paired Gemma 4 with BOTANIC-1, a model trained on plant DNA from 320 species, to identify mutations that alter crop traits. In a melon yield gene test, the pair ranked the target mutation first out of 2,494 possibilities in under four minutes, versus years of physical breeding trials. A demo and paper are linked.
A report published by MIT in June 2026 says AI is weakening core parts of the university experience: fewer students attend office hours, participation in online discussions is falling, study groups in dorms and libraries are shrinking, and teachers increasingly struggle to tell what students have actually learned.
Hinton, Bengio, Andrew Barto, Jakub Pachocki, and 18 other authors published a paper titled 'What if automating AI R&D triggers an intelligence explosion?'
Google researchers proposed RRSI (Regularized Recursive Self-Improvement of Agent Harnesses), adding constraints to the self-optimization loop in which a language model repeatedly rewrites an agent harness, to prevent the agent from memorizing test tasks.
Aleph Alpha tested Alibaba's Qwen, DeepSeek, and Moonshot AI's Kimi on 967 sensitive topics. Its in-house scoring system rated only 17% to 41% of responses as balanced; the rest repeated official positions, avoided the topic, or refused to answer.
The Microsoft Copilot Studio team and Hugging Face jointly released ThinkingBox. It runs agents in isolated MCP tool sessions and then performs executable checks on final backend state and side effects, rather than judging only the final reply or whether tool calls were allowed.
Why it matters: The agent benchmark from Microsoft and Hugging Face uses 507 stateful workflows and 20 repeated runs, shifting the focus from tool calls to the final database state.
Researchers proposed an Image-to-Code method in which a coding agent iteratively writes executable Blender programs from a single photo, producing runnable, editable scene code. They also released the LEGO-Bench benchmark.
Harvard physicist Matthew Schwartz described BootLoops, an open-source tool that lets Claude perform precise scientific computations, in a guest post for Anthropic. In three months, the team produced 36 manuscripts across 18 fields with 19 co-authors. Claude used BootLoops to calculate 30 integrals, 15 of them for the first time.
