Research papers, technical reports, benchmarks and datasets.
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.
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.
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.
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.
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.
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.
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.
He Kaiming's team published VISTA, a visual-native harness that lets existing multimodal models observe environments, store raw frames outside context, and recall, zoom or inspect them during reasoning without retraining. Paired with VISTA, Claude Opus 5.0 completed all 25 public ARC-AGI-3 games with a perfect 100 action-efficiency score, using 57.4% fewer actions than the human first-play baseline, while GPT-5.6 Sol scored 99. The team also tested browser games, mazes and connect-the-dots tasks, and lists embodied tasks as future work.
On October 7, OpenAI released 722 mathematics manuscripts, including two on the Kakeya conjecture: a 97-page paper claiming a three-dimensional maximal-function result stronger than Wang and Zahl's, and a 175-page paper claiming the four-dimensional Hausdorff dimension case. Both are attributed to OpenAI, are not Lean-formalized or peer-reviewed, and OpenAI's README says some unformalized results may be flawed.
Why it matters: The piece traces how two OpenAI manuscripts extend Wang and Zahl's three-dimensional Kakeya work to a stronger maximal-function version and to four dimensions, while noting the proofs lack peer review.
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.
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.
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.
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.
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
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 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.
A benchmark tested the local Qwen3.8-27B-Q4_K_M.gguf on 5,070 cases without reasoning, asking it to add positive integers using only English words and return exact results. It achieved 23.57% numerical accuracy and 96.17% format compliance. Accuracy fell as the numbers grew, from 97.04% on one- to three-digit numbers to 6.44% on ten- to thirteen-digit numbers. In 169 paired cases with medium reasoning enabled, the model answered 167 correctly.
Meta said that after AI models achieved gold-medal-level performance in five math, physics, and chemistry competitions, it had mathematicians work with Muse Spark 1.1 and 1.2 in Thinking Mode through the meta.ai chat interface to solve open problems with no existing solution path and no custom research scaffolding.
Why it matters: Meta made six papers from mathematicians' collaboration with Muse Spark public and described the division of work between people and AI and the review process.
Ataraxos, an AI system developed by researchers at Carnegie Mellon, MIT, NYU, and Stanford, beat Pim Niemeijer, considered the strongest Stratego player ever, with 15 wins, one loss, and four draws. Training used only 16 GPUs and cost a few thousand dollars.
Kaiming He's team proposed NAT-ARC, a purely visual ARC-solving system that does not rely on LLMs. It pretrains an encoder with MAE on ImageNet and transfers it to abstract grid reasoning. The best single model has 0.6B parameters and reaches 63.4% pass@2 on ARC-1; an ensemble reaches 70.2%, with about one-quarter the parameters of the specialized LLM system The ARChitects, which uses an 8B model and reaches 71.6%.
Researchers' Ataraxos achieved an 85% effective win rate against Niemeijer, the strongest Stratego player ever, ending human dominance in the imperfect-information board game.
Researchers from Tencent Hy, Fudan University, and Tsinghua University released ExplorationBench, a benchmark measuring how AI systems explore by discovering unfamiliar rules. It uses two verifiable "Alien Worlds": AlienCode (31 hidden rule changes, 70 tasks) and AlienLogic (24 patched inference rules, 70 theorems), graded by an interpreter or proof checker with no LLM judge. Across 10 frontier AI systems, feedback beat thinking alone, and rankings barely transferred between the two worlds.
In a post on its Science Blog, Anthropic says Claude completed a nine-loop calculation of scattering amplitudes in planar N=4 super-Yang–Mills theory. The previous record for the model was eight loops, held by Lance Dixon of SLAC and collaborators.
Google Research proposes Retrieve-for-Train, a framework that uses offline RL to discover reward-aligned query fan-outs and compile them into supervision, then distills them into a 53.9M-parameter diffusion retriever. At inference, it generates a complete result set in one non-autoregressive pass without chain-of-thought tokens. The method fine-tunes fan-out language models based on Gemma3-4B and Qwen3-4B, using set-level attribute rewards to address paraphrase collapse in zero-shot LLMs and autoregressive latency.
OpenAI has shared an AI-generated solution to the Navier–Stokes Millennium Prize Problem, including a written explanation and a formal proof in Lean.
Multiverse Computing has published a paper proposing Quantization-Aware Healing (QAH). After compressing GPT-OSS 120B to 60B parameters and quantizing it to MXFP4, the method distills directly from the original model before compression rather than from a recovered bfloat16 checkpoint.
Why it matters: The paper presents a repair recipe that distills from the original model after compression and quantization, with stability comparisons against QAT.
Import AI 469 introduces DiG-bench (Discovery in Games), a new benchmark of 70 games with hidden rules that tests AI's ability to infer environmental rules through exploration. Opus 5 and Fable 5 with Claude Code perform best, but complete only 0.2 of the tasks at Tier 7, far below the human success rate of 100%.
Microsoft Research has introduced MindTopo, a benchmark for topological reasoning in multimodal LLMs. It covers connectivity, separation, ordering, containment, and knot relations, across static reasoning and interactive planning. Current models are substantially better at static recognition than interactive planning, but both remain far below human performance; failures occur more often in planning than perception. Image and video generation helps only when topology is preserved within a single frame and remains unreliable across multi-step operations.
