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.
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.
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.
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.
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.
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.
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.
Can AI automate Epoch? We're introducing Epoch Automation Reports to evaluate frontier models on realistic, open-ended tasks drawn from our own work. Claude Fable 5.1 and GPT-6 Astra lead, yet they are far from fully automating Epoch’s work.
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.
Hinton, Bengio, Andrew Barto, Jakub Pachocki, and 18 other authors published a paper titled 'What if automating AI R&D triggers an intelligence explosion?'
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.
A recent paper involving Shing-Tung Yau acknowledges GPT-6 Astra and Claude Pro for helping explore some proof strategies and calculations. The paper addresses whether the 27 exotic spheres in seven dimensions can be given a metric with strictly positive sectional curvature, a problem Yau added to his own list in 1982.
A Graphite study found that Claude Opus 5.5 uses 'this matters' 116 times more often than people and 'why X matters' 92 times more often. Its most distinctive word is 'dependable', used 23 times more often. OpenAI's Astra favors corrective frames such as 'another dimension' and 'not simply X'; the latter appears more than 100 times as often as in human writing.
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.
Anthropic says Claude discovered a previously unknown enzyme system in bacteriophage DNA, next to a stretch of repetitive DNA resembling CRISPR. The system's function is not yet known, but the few known systems with similar features can cut, copy, and paste DNA. Anthropic says more research is needed to determine what it does and whether it could be used for gene therapies like CRISPR.
The UK's AI Security Institute (AISI) is adopting EvalEval's Every Eval Ever schema and Evaluation Cards platform to share evaluation results publicly and improve reproducibility in evaluation science.
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%.
Epoch and METR have released MirrorCode, a benchmark testing AI systems on coding tasks that take humans a long time to complete. Of 25 target programs, 17 achieved a perfect-score run at least once, while eight never reached the 100% threshold.
Researchers at Oxford, the UK's AI Safety Institute, Stanford, and the London School of Economics ran four experiments with 6,923 participants and 18,978 conversations. They found that AI systems consistently outperformed human experts at text-based persuasion, even when experts chose their own topics, researched in advance, and were offered a £1,000 prize incentive.
Jan Leike reports that Claude makes fully autonomous progress on scalable oversight research, measured by performance gap recovered (PGR). Claude iterates on several different techniques and ends up significantly outperforming human researchers for $18k in credits.
Google and Cornell University tested GPT-4o, Perplexity, Claude 3.5, Gemini Advanced Pro 1.5, NotebookLM, and a custom RAG system on 67 expert-level questions about high-temperature superconductivity. An international panel of experts scored the answers blind across six dimensions.
