Research papers, technical reports, benchmarks and datasets.
In Nature this week, Satya Nadella says his organization’s new research shows how machine learning can improve weather forecasts two to six weeks out, a timeframe that has historically been very difficult to predict. He highlights the work’s potential to help people better prepare for extreme weather.
Google Research presents three projects connecting AI agents with wet-lab experimentation through VERID laboratory automation, closed-loop aptamer discovery, and de novo protein binder design. Across three rounds, aptamer on-target hit rates reached 33.5% for myoglobin and 42.1% for α-thrombin; the protein design pipeline accelerated computational cycles more than eightfold and produced binders with single-digit nanomolar affinity. Integrating these projects remains a next step, with safeguards including sequence screening, simulation before robot execution, and human authorization and supervision.
Why it matters: The aptamer experiments illustrate how experimental feedback can expose misleading optimization targets, offering a transferable approach to distinguishing genuine biological improvements from artifacts in automated discovery workflows.
Anthropic reports that Claude Mythos 5 designed quantum circuits from plain-language instructions through its STEM Fellows Program. The circuits improve on the best published ones, sometimes by an order of magnitude or more.
Google DeepMind reports that four Gemini 4 Argon agents beat Minecraft from scratch in 27 hours. Operating in a sandbox using only code and text, they wrote 2,783 programs, divided up jobs over chat, shared notes, and debugged errors together, all with no human help. https://t.co/uyIhDOQXGT
Why it matters: The account offers a concrete example of agent coordination through code, chat, shared notes, and collaborative debugging during a task lasting 27 hours.
Nick Cammarata’s Neural Networks Are Already Clean argues that improved observation tools reveal coherent neuron functions, reporting clear interpretations for 99%–100% of final-layer neurons in tested modern vision models. He reinterprets classic examples of apparently unrelated features as geometric detection or recursive generalization. Neel Nanda counters that these findings do not resolve representational superposition in language models and defends SAE applications. Cammarata describes the proposal as a working hypothesis, acknowledging gaps in language-model training evidence, multimodal coverage, and causal validation.
Latent Space highlights Goodfire’s Nick Cammarata’s working theory that neural networks are already largely interpretable, with further agentic analysis resolving many familiar examples of polysemanticity and superposition. Cammarata argues that most neurons are legible, contrasting with roughly 2% in his 2023 OpenAI effort. The roundup also reports objections concerning weak activations, neuron simulation and neglected non-maximal activations, alongside a reminder that the paper itself is not yet available.
Artificial Analysis reports that NVIDIA's Vera Rubin NVL72 delivers up to ~5.7x GB300 NVL72's throughput per megawatt on DeepSeek V4 Pro in initial AA-AgentPerf results. Its advantage is largest at faster per-user speeds; peak-throughput gains range from 1.14x to 1.45x across the four tested models. NVIDIA submitted all Vera Rubin results using TensorRT-LLM and vLLM with NVIDIA Dynamo; the benchmark replays real agentic coding sessions and measures throughput per megawatt of measured GPU power alongside per-agent output speed.
NVIDIA and AI4Bharat at IIT Madras introduce Nemotron-pretraining-Indic-v1, approximately 2.8 trillion tokens across 22 Indian languages, releasing soon alongside its open, GPU-accelerated generation pipeline. The corpus includes native-script text for all 22 languages and romanized text for twelve, translated from curated English sources with quality checks and source tracing. An Indic continued-pretraining stage improved five reported benchmarks, but combined synthetic and natural data, so the comparison does not isolate this corpus’s contribution.
Why it matters: The workflow offers a reusable approach to multilingual data generation by separating translation, quality assessment, repair, and script conversion while preserving links to source documents.
Microsoft Research presents Magnetar, a zero-knowledge virtual machine that proves RISC-V program execution using security based only on standard hash functions. It supports programs written in Rust, Go, or C without requiring developers to rewrite them as cryptographic circuits. Across 15 benchmark programs, Microsoft reports average speeds 10 to 39 times faster than the fastest existing comparable zkVMs on the same machine, with proofs verified in about 12 milliseconds.
Why it matters: The benchmark comparisons connect proving speed with general-purpose execution and standard hash functions, clarifying the tradeoffs that have limited broader adoption of verifiable computing.
GPT-6-Astra, running in OpenAI's Codex environment without human hints, solved 49 of 50 TheoPhysEval problems across eleven theoretical physics subfields, costing just under $116 overall. Anthropic's Opus 5.5 scored 92 percent, while both models reached 92 percent without internet access. The researchers emphasize that the benchmark tests calculations using known methods, while choosing promising research questions remains a separate challenge for current agents.
Why it matters: The benchmark distinguishes proficiency in concrete physics calculations from the ability to choose fruitful research questions, clarifying which parts of scientific work its high scores actually measure.
Xu Disheng’s team used Shanghai AI Laboratory’s Shusheng-S2 model in a human–AI collaboration to prove the three-dimensional Ghys conjecture, with a preprint posted on arXiv. Researchers guided seven rounds of recursive exploration as the problem graph expanded from 15 to 389 nodes. A Peking University team used its SHEAF framework to formalize all four main theorems, producing approximately 550,000 lines of Lean code in about five days and passing Lean checks.
Why it matters: The workflow offers a transferable approach to research mathematics, pairing expert guidance with recursive model exploration and organizing formal verification around a dependency graph of checkable proof tasks.
Artificial Analysis asked AI models to build the Eiffel Tower in 3D offline. Claude Opus 5.5 and Grok 4.7 included all 72 inscribed scientists’, engineers’, and mathematicians’ names in the correct order, although the shared brief never mentioned them. Claude Opus 5.5 also matched each set to its orientation in Paris. Both models worked as agents without internet access, and the names first appeared in their generated code, never in files or tool outputs they read.
Welcome to the era of embodied AI. We are launching AA-Robotics v0.1, our research preview that measures frontier language model performance on understanding, planning, and acting in the physical world. In v0.1, each model is given control of a 6-axis robot arm (5 joints and a https://t.co/fCwpnDlvB2
Six months ago, frontier AI models could barely beat Slay the Spire. Now, according to our new SpireClimb benchmark, they win 90%+ of runs at the easiest difficulty, and up to 20% of runs at the hardest. https://t.co/KvV5gD2RBh
10 new papers in health science research. all written with muse models in under 4 weeks. health is one of the areas where we believe muse is going to have the greatest positive impact on humanity ❤️ https://t.co/eejFDFd2Co
Google engineer Haydn Belfield says Google DeepMind has seen no measurable AI-driven acceleration of AI research despite substantial automation. AI writes more than 75 percent of production code, but every change receives human review before submission; Argon completes about half of RE-Bench's 74 ML engineering tasks. Google's General Capability Index shows steady Gemini capability growth without statistically significant acceleration, contrasting with researchers' warnings of a possible intelligence explosion.
Artificial Analysis is adding AMD MI355X results to AA-AgentPerf, which replays real coding agent sessions with prompt caching and tool-call delays. On a single 8-GPU node, MI355X delivers ~15% to 40% more throughput per MW than B200 on DeepSeek V4 Pro 0813 between 35 and 75 tok/s per user. DeepSeek V4.1 Flash matches NVIDIA Blackwell performance for a single agent but trails as load rises; GLM-5.2/5.3 and Kimi K3 also trail the compared NVIDIA systems. Results use ROCm nightly and dev builds of vLLM and SGLang; Kimi K3 comparisons use NVIDIA multi-node configurations.
Researchers used AI-based protein design tools to retain the catalytic cores of two types of protease while replacing the rest of each protein. The resulting enzymes could be tuned to cut specific proteins in half, and the researchers demonstrated their effectiveness by cutting proteins associated with neurological diseases.
Microsoft researchers and collaborators published research in Nature on Probabilistic Bias Correction (PBC), an open-source machine learning tool that corrects systematic bias in existing weather models to improve forecasts two to six weeks ahead. In the study, PBC improved ECMWF’s physics-based model accuracy for 98 percent of precipitation and 92 percent of temperature targets, and doubled the accuracy of its AI forecasting system. The tool can run on CPUs, though ECMWF cautioned that operational implementation remains challenging.
Google released a technical report evaluating its experimental Geospatial Reasoning agent across four benchmarks and is expanding early access through Ask Google Earth and an Agent-to-Agent API. Evaluations using Gemini 3.8 Flash and Gemini 3.1 Pro found that specialized Earth AI tools improved spatial reasoning, while upfront planning helped with open-ended workflows. Google is releasing development splits for GeoKTools and GeoExpertQueries, but cautions that GeoExpertQueries evaluates adherence to expert workflows rather than final-answer correctness, making agent outputs drafts for expert review.
Ai2 introduced HiRO-ACEv2, an updated climate emulator that generates global 3 km rain, wind, and pressure data at about 1/300th the cost of X-SHiELD. Given coarsened X-SHiELD data, its HiROv2 downscaling model matches high-resolution statistics to within 3%, even for events occurring once in a million data points. The system currently works only with ACE tuned to match X-SHiELD and cannot yet generate high-resolution data for other climates; downloadable models, tutorial notebooks, and a web demo are available.
Why it matters: The comparison with X-SHiELD gives researchers a concrete cost and accuracy reference for extreme weather simulations, with applicability currently limited to the climate represented in training.
White Circle researchers report that their RISE attack method bypassed safeguards in DALL·E 3, Nano Banana 2, and GPT-Image-2 to generate explicit images about 13 percent of the time. The method uses widely available models to iteratively develop and evaluate jailbreak strategies. White Circle told Ars Technica that subsequent testing found ChatGPT Images 2.5 and Nano Banana 2.1 significantly more vulnerable than their predecessors; Google and OpenAI did not respond to requests for comment.
Artificial Analysis added DeepSeek V4.1 Flash to AA-AgentPerf, its benchmark replaying real agentic coding sessions with prompt caching and tool-call delays. At 100 tokens/s per user, GB300 leads with approximately 261M tokens/s per MW, followed by B300 at approximately 220M and MI355X at approximately 128M. Compared with DeepSeek V4 Pro 0813, V4.1 Flash scores 39 versus 36 on the Artificial Analysis Intelligence Index and delivers approximately 1.3x to 6.2x the throughput per MW at that user output speed, depending on hardware.
Reporting on Jiayuan Mao’s IROS 2026 presentation, Leiphone describes research from UPenn, MIT, and Stanford on robot learning and compositional generalization. RAPID combines a coding agent with physics validation to learn complex non-grasping tasks from one human demonstration. In separate video-action research, the team introduced Temporal Ratio to measure attention to future versus current frames; adaptive guidance based on this metric raised average success on unseen compositional tasks with a real dual-arm robot from 71.8% to 83% without retraining.
Artificial Analysis introduced Decision Index v1.0 to compare decision models’ capability, cost, and speed across classification, tool calling, knowledge, and reasoning. Jev 1.13 leads decision models with 58, while Claude Haiku 5.5 (Low) scores 61 but takes 1.4 seconds per task versus 0.13 seconds and costs $184 per 1M Tasks versus $42. Leading decision models perform comparably to language models on classification and tool calling but trail on knowledge and reasoning, where their answer probabilities generally exceed their accuracy.
Anthropic researchers William Zhou and Daniel Freeman report that small AI models could become capable of autonomous replication across poorly secured consumer devices by the end of 2026. Gemma4-26B-A4B ran across 15 Android TV boxes costing about $55 each, while GLM-5.2 replicated across all eight test nodes in four of five runs; the experiments omitted actual hacking. No current model combines the small size and replication capabilities needed for a fully autonomous AI worm.
Jay reports that Anthropic researchers William Zhou and Daniel Freeman tested GLM-5.2 in an isolated network of 8 GCP nodes running Android 15 redroid emulators. It autonomously deployed Llama3.1-8B and propagated to all 8 nodes in 4 of 5 trials, reaching 7 in the remaining trial. Separately, researchers manually distributed Gemma4-26B-A4B across 15 Android TV boxes at 0.39 tokens per second, while Qwen3.5-9B deployment success rose from 45% to 61%–84% across later generations inheriting debugging notes.
Baoyu summarizes Anthropic research in which GLM-5.2 alone completed a replication workflow across simulated Android servers, while smaller models failed. The experiment used unauthenticated ADB access and ran GLM-5.2 remotely to deploy Llama3.1-8B, limiting what the results establish about autonomous spread. Anthropic roughly projects that smaller models could match GLM-5.2’s tested capability by year-end based on general capability trends. Researchers also manually distributed Gemma4-26B-A4B across 15 Android TV boxes, achieving 0.39 tokens per second.
Why it matters: The experimental setup helps distinguish successful replication under simplified conditions from autonomous spread on consumer devices, clarifying the limits of the projected timeline for smaller models.
Amazing to see research move from the lab to real-world clinical utility—congratulations to the team on this milestone and The Lancet paper! https://twitter.com/sundarpichai/status/2108329198955958339 In the quoted post, Sundar Pichai describes Google’s prospective AMIE study with @BIDMC_Medicine involving real urgent care patients. Clinicians reported that its medical-history summaries helped prepare for visits in 75% of cases and influenced care approaches in more than half, while its differential diagnoses matched doctors’ final diagnoses 90% of the time.
Artificial Analysis launched AA-3D, a benchmark in which AI agents turn 55 written briefs across seven domains into interactive three.js scenes for blind human comparisons. Over 10,000 counted votes from over 1,000 panelists rank 21 models by Elo with 95% confidence intervals; Claude Sonnet 5.5, Claude Opus 5.5, GPT-6 Astra, and Claude Fable 5.1 are statistically tied at the top. Models receive up to three hours and 250 turns, and reported costs range from $0.20 to $32.08 per scene.
QbitAI reports that mathematics enthusiast nor used GPT-6 Astra to prove optimality for packing 15, 16, 17, 18, and 20 unit disks inside a circle, completing the remaining cases for N≤20. The approach combines geometric constraints, finite computational enumeration, and exact optimality arguments. Nor describes substantial human involvement, including guiding the model, fixing issues, manually verifying validity, and rechecking the work with Astra and Opus 5.5.
Why it matters: The reported proof strategy shows how geometric constraints can reduce an infinite search space to finite cases, with rigidity arguments connecting computational elimination to exact optimality.
Researchers from Sakana AI and UC Berkeley introduced MASS, a method that uses copies of Qwen3.6-27B to solve tasks, evaluate results, and optimize workflows without a stronger teacher model. After two rounds, its test-task win rate against the base model reached 69.9 percent, although benchmark gains largely reflected more completed work and some scores declined. Shared blind spots allowed flawed results to pass evaluation and enter training, prompting the researchers to recommend independent checks, limits on autonomy, and clear intervention conditions.
Artificial Analysis launched its FrontierCode 1.1 leaderboard, evaluating models on Cognition’s private coding benchmark using the 100-task Main set and its Stirrup open source reference harness. Grading combines unit tests with maintainer-oriented criteria; failing any blocking check yields zero. Claude Opus 5.5 (max with fallback) leads at 52%, with blockers passed in 58% of attempts; MiMo-V2.6-Pro leads open weights models at 45%, while GPT-6.1 Sol (max) scores 49% at $0.81 per task, matching Claude Fable 5.1 (max with fallback).
Microsoft Research introduces StudentSim, a research system that trains personalized student simulators from publicly available, de-identified learner records to evaluate and improve AI tutors. StudentSimEval measures behavioral fidelity and guidance responsiveness across chess, second-language English writing, and mathematics, where StudentSim outperforms leading baselines. In a chess tutor study, training with StudentSim feedback achieves 90.5% accuracy versus 75.7% without reinforcement learning and 71.6% with GPT-5.4 feedback; real learner studies remain essential, and the paper and code are publicly available.
Why it matters: Evaluating both a simulated learner’s starting ability and response to guidance offers a concrete method for assessing tutor feedback beyond the correctness of an explanation.
A Microsoft Research study reports that Fork-and-Flush helps long-running AI research agents explore better solutions by creating parallel branches that retain workspace files but discard chat history. Across 13 tasks with the same compute budget, it achieved a min-max-normalized average score of 0.78, versus 0.47 for single runs and 0.54 for best-of-4. The researchers found that accumulated context can trap agents in similar approaches, even when conversation histories are periodically compressed.
Import AI 476 covers Humanity’s Sixth Sense, a benchmark from Elorian and Scale AI that tests intuitive visual reasoning. People scored 93.1 aggregate pass@1, while GPT 6 Astra scored 53.6; GLM-5.3-flash and Kimi K3 scored 26.8 and 25.5. The newsletter also examines how AI could accelerate space development and reports on TasteVal, a benchmark of AI systems’ research taste.
Google DeepMind’s AlphaProof Nexus solved dozens of open math problems, according to a peer-reviewed study in Science, including nine of 353 open Erdős problems, two unsolved for 56 years. It also proved 44 OEIS conjectures, solved an open algebraic geometry question, and improved a min-max optimization bound. Built around Gemini 3.1 Pro, it drafts and revises proofs in Lean using checker feedback and can call AlphaProof for harder sub-problems, though DeepMind reports it solved the nine Erdős problems without that tool.
Why it matters: DeepMind’s use of peer review provides a concrete comparison with OpenAI’s publication of AI proofs on GitHub amid a debate over how AI mathematics should be evaluated.
Huawei’s Noah’s Ark Lab, UCL, and collaborators report that Memento 3 cleared all 25 public ARC-AGI-3 games and 183 levels using 7518 actions, about 56% fewer than humans, achieving a 100 RHAE score. Built on Claude Opus 5 without weight updates or game-specific fine-tuning, the agent iteratively revises a Markdown rulebook and executable Python world model. In Atari Pong, its learned code-based strategy won three games 21:0 without further LLM calls; the experiments primarily cover game environments with relatively clear rules.
Evals company Vals AI tested GPT-6 Sol and Claude Opus 5.5 on the Vibe Code Bench, both solo and as teams, at medium and maximum reasoning effort. Teams cost between 1.8x and 5.1x more than single agents, and only one of four comparisons showed a statistically significant improvement: GPT-6 Sol at medium reasoning, where the team scored 7.3 points higher. At maximum reasoning, teams gave neither model a real advantage.
Epoch AI's InnovationEval asked AI agents to invent a new post-training method starting from GRPO, then implement, test and refine it without internet access. GPT-5.6 Sol reached about 35 percent of the human reference method SDPO's improvement under generous grading, and about 15 percent counting only rule-compliant changes, while Claude Fable 5 produced no measurable gain. Both agents reported only their best runs, and Epoch concludes humans would need to review all AI-generated research.
