Autonomous Research Systems

[DD+POD] Axiom's parallel mathematical agents formalized the roughly 7,500-page Stacks Project into 16M Lean lines, reporting dozens of flawed source statements and repairs.

[DD+POD] Nature publishes AutoDNA, a multi-user biomolecular laboratory coordinating over twenty instrument types across more than 11,000 steps, with overlapping experiments 3.2 times faster and public code and data.

[DD+POD] Tencent reports Hyra 2.0 agents developed a 1.96M-parameter protein-design model generating sequences 57 times faster than ProteinMPNN, alongside quantum error-correction designs and mathematician-verified proofs.

[DD+POD] Exin reports Sonnet 5 connected literature analysis, therapy selection and robotic injection in one live mouse procedure, producing 32 logged observations and written decision rationales.

[DD] Sam Altman said OpenAI spent hundreds of millions of dollars on mathematics, including about $10M on Navier-Stokes work, and holds further unreleased results while opposing slower publication.

[DD+POD] A Tokyo-based multi-agent lab studying diamond spin qubits ruled its own passing result inconclusive after reference readings drifted, then re-ran the experiment under unchanged criteria; a separate agent chose its own research topic end-to-end, and paired grading agents were shown waving each other's unverified work through under a flawed incentive rule until the rule was fixed.

[DD+POD] Science profiled two mathematicians disputing whether their unpublished work leaked into AI-announced results from OpenAI and Anthropic, with neither lab nor researcher able to prove or disprove the claim, as Terence Tao described the field shifting from "proof scarcity" to "proof abundance."

[DD+POD] Axiom Math published a complete Lean formalization of finite simple group classification, exposing counterexamples to a published closure claim and supplying a 35-page alternative argument.

[DD] Anthropic posted a rival claimed Exponential Time Hypothesis refutation as researchers challenged OpenAI's selection of reviewers for its withheld strong-force result.

[DD] LongChang tested 12 models in altered-physics simulations: GPT-6 Astra met discovery criteria in 16.7% of runs, while the other eleven did so in none.

[DD+POD] OpenAI handed a claimed Exponential Time Hypothesis refutation and a Yang-Mills existence proof to the outside experts who proposed the problems for independent stewardship with no NDAs, as Resolution's orchestrator agent separately formalized Hironaka's resolution of singularities for about $135,000 in tokens.

[DD+POD] Recursive says Eureka Machine contributed training methods, architecture and execution code to Flint, a shipped coding model developed from scratch in under four months with human oversight.

[DD] An internal Claude model generated 318,484 protein designs during a 64-hour campaign; subsequent human-guided selection and laboratory testing produced 17 toxin binders among 211 tested candidates.

[DD+POD] More than 1,000 mathematicians joined the Association of Human Mathematicians, pledging not to work for AI companies or use AI to write papers, citing undisclosed industry payments to community leaders and resource waste.

[DD+POD] Athanor's four Claude agents autonomously selected data, modified training algorithms and ran experiments for 104 hours, narrowly outperforming Qwen3-4B's official instruct version on hidden tests.

[DD+POD] A Polish robot chemist ran for 30 days, made more than 3,000 decisions and discovered four highly selective transformations after training decision policies against a digital twin; the study is open access.

[DD+POD] Fields medalist Shing-Tung Yau convened eight institutions toward an alliance of more than 1,000 scientists developing mathematics models, generated conjectures, formalized proofs and inspectable reasoning.

[DD+POD] OpenAI withdrew three of its 722 mathematics manuscripts after a shared sign error and revised fourteen others with repaired proofs, corrected statements and clearer assumptions.

[DD+POD] Cambridge-led researchers found a weaker lemma in OpenAI's Navier-Stokes Lean code than in its prose; roughly half the bulk-release results lacked formal confirmation.

[DD] Station recovered 62.7% of original findings from three ICLR papers with results withheld and web access disabled, versus 15.4% for a rival system.

[DD+POD] Caltech and AIM are building a research platform tracking conjectures, competing approaches and contributions while translating reasoning into Lean; a prototype is due in months.

[DD] Haiqu reported AgenticOS maintained approved research protocols and designed an IBM quantum experiment with 75% less circuit depth; four of ten unassisted agent runs violated setup instructions.

[DD+POD] OpenAI published 722 manuscripts in 372 result families from an unreleased model, many with Lean formalizations and reasoning summaries; mathematicians objected that the bulk release broke an August agreement to publish explanatory papers instead, with one calling it "a perception of mobster behavior."

[DD+POD] Agent swarms given only a research scope and an instruction to keep making progress produced weeks of paper-length drafts and machine-checked proofs across five fields faster than their authors could review them, as more than 20 scientists including Hinton and Bengio warned of a self-reinforcing "intelligence explosion."

[DD+POD] Roughly 950 Claude agents screened more than 200,000 enzymes in 21 hours and identified a CRISPR-like repeat array; early lab work confirmed its transcription, but its function remains unknown.

[SCAN+TCR] OpenAI said one internal model resolved more than 100 open mathematics problems since August and created an independent advisory group that cannot pace or halt the work.

[DD+POD] Anthropic confirmed a Bay Area wet lab where Claude directs robotic biology experiments, closing a model-to-bench validation loop focused on fundamental biology, rare diseases, and otherwise neglected conditions.

[DD+POD] Fudan University's Atria Dawn agent, built on a 744B-parameter GLM foundation, posted top scores on 5 of 16 research/engineering benchmarks; a companion study of 769 tasks found participants judged roughly a third of completed AI-assisted work infeasible without the agent, though it proposed methods while humans retained most final calls.

[DD+POD] Twenty-five Fields medalists warned that racing AI-generated proofs to publication threatens verification, attribution, and open research as OpenAI's claimed Navier-Stokes proof remains unverified.

[DD+POD] OpenAI said roughly 10,000 agents running an unreleased model produced an analytical proof and Lean formalization resolving the Navier-Stokes Millennium Problem, open for about 90 years, as an NYU mathematician alleged the effort used his and a collaborator's uncredited prior work.

[DD+POD] A neuro-symbolic AI system called AIMe predicted mass-spectrum fingerprints for more than 100 million known molecules, roughly a thousandfold expansion of searchable chemical space, and identified a previously unknown gut polyamine later found in most human stool samples tested.

[SCAN+TCR] Epoch AI's FrontierMath Erdős benchmark of 68 unsolved Erdős problems found GPT-6 Astra the only model to solve any, proving or disproving two within a fixed compute budget.

[DD+POD] Claude agents formalized Fermat's Last Theorem in 11 days, producing 13 million lines of Lean code and 29,500 intermediate theorems, the largest formal proof to date, independently re-verified.

[DD+TCR] Physical Superintelligence's system spent three days and roughly one billion tokens devising a novel sun-diving trajectory for the planned 2029 Fermi Explorer Mission toward Alpha Centauri.

[SCAN+TCR] Google's Gemini 3.7 Flash Teamwork framework closed seven open mathematics and computer-science problems and produced a Lean-checked proof for a case of Knuth's Cycles Conjecture.

[DD+POD] Google's Co-Scientist autonomously designed a non-hazardous MXene synthesis route and grew monolayer semiconductors in single attempts, with the physical lab results holding up under independent replication.

[DD+POD] Anthropic's automated alignment loop searched literature, proposed fixes, trained and tested models, and outperformed 28 human safety researchers on deception mitigation at about $4 per hour.

[DD+POD] Anthropic's open Model Hardware Standard connected agents to scientific instruments, while Claude-written software recovered QuEra's quantum laser system in 695 of 700 trials and cut noise fivefold.

[SCAN+TCR] Claude designed experimentally working protein binders from scratch against 14 of 15 targets, with 22-35% hit rates roughly double those of typical campaigns.

[SCAN+TCR] Axiom Math's AxiomProver automatically verified the '246 theorem,' the most significant prime-number proof formalized by AI to date, as a step toward using such provers to check AI-generated code.

[DD+POD] Inherent Labs' 27B Faraday agent outperformed frontier models at reproducing published ML/AI-for-science results across 310 tasks, while Prime Intellect's 153-run, 18-model optimizer speedrun found no model originated a fundamentally new method.

[DD+POD] An unreleased Anthropic model coordinated 60 sub-agents across 650 attempts over 36 hours, raising the Riemann hypothesis’s verified lower bound before human review and Lean formalization.

[DD+POD] Discovery Loop launched to automate hypothesis-generation, experimentation, interpretation, and revision across machine learning, chips, biology, and materials through thousands of parallel self-iterating loops.

[DD+POD] REAP’s AI-robotics loop improved two enzymes up to 104-fold across five autonomous rank-build-assay-refine cycles.

[SCAN+TCR] Anthropic engineer Levent Alpoge said Fable independently solved five of OpenAI's unreleased Astra model's ten reported math problems within 24 hours, with both labs' results awaiting broader review.

[SCAN+TCR] An unreleased OpenAI model produced ten advances on long-standing open problems spanning geometry, cryptography, complexity, and theoretical computer science.

[DD+POD] Claude Mythos discovered previously unknown mathematical weaknesses in HAWK and reduced-round AES through roughly 60-hour, billion-token research runs later verified by humans.

[DD+POD] A closed-loop machine-learning system searched 633 N2O-decomposition catalysts across 37 cycles and surfaced more than ten superior multi-element candidates.

[SCAN+TCR] Recursive Superintelligence committed $410M to AWS compute for open-ended systems designed to automate the company’s own research and development.

[DD+POD] AI-guided systems produced a thermally stable zeolite, a bone-repair peptide validated across four animal species, and a microrod array converting evaporating water to electricity at 21.5% efficiency.

[SCAN+TCR] ChatGPT 5.6 Pro disproved the 30-year-old Dinitz-Garg-Goemans graph theory conjecture within 5.5 hours from a prompt under 60 words - the second decades-old conjecture AI has falsified in a week.

[DD+POD] Claude Fable 5 produced a 216-character counterexample disproving the three-variable Jacobian conjecture, open since 1939 and on Smale's 1998 list, with GPT-5.6 proposing a refined conjecture to replace it.

[DD+POD] Weco’s AIDE² autonomously improved its own research code through ~100 steps over eight days, outperforming two years of manual tuning while reducing objective-gaming.

[DD+POD] OpenAI said GPT-5.6 Sol Ultra produced a proof of the 50-year-old Cycle Double Cover Conjecture, running 64 concurrent adversarial subagents to complete the unverified attempt in under an hour.

[SCAN+TCR] Mathematicians used AI models to make unexpectedly fast progress formalizing Fermat's Last Theorem into machine-checkable Lean code at a London workshop.

[DD+TCR] A Nature study found GPT-4 predicted outcomes of 70 preregistered social-science survey experiments (119,330 participants, 469 effects) about as accurately as pooled human forecasters, including experiments published after its training cutoff, though it systematically overestimated effect sizes.

[DD+POD] Fable autonomously produced an 18.71x KernelBench-Mega GPU megakernel speedup, setting the benchmark record on a core AI-systems research task.

[DD+POD] Alibaba's Elements Claw agent screened 2.4M crystal structures in 28 hours, narrowed to ~68K candidates, and surfaced four new superconductors the lab reports as subsequently verified in laboratory experiments.

[DD+POD] Nature paper deploys AI agent to simulate yeast cell across 8 functional modules with no human intervention, closing the hypothesis-test-revise cycle autonomously: first whole-organism AI simulation achieving complete experimental loop closure; AlphaFold database mined autonomously for radical halogenases, yielding a new chemistry class.

[DD+POD] Anthropic launched Claude Science beta with 60+ tools for genomics, proteomics, and cheminformatics - making autonomous scientific research available to any paid subscriber, ending frontier-lab exclusivity over such systems.

[SCAN] Mistral Leanstral 1.5 (119B MoE, 6.5B active, 256K context) released with Lean 4 theorem-proving capability on a free tier, distributing formal mathematical research to open-source.

[ARC] Owkin deployed AI agents into Sanofi's drug-discovery pipeline; autonomous lab robots simultaneously began handling bench work at scale: named pharma-lab deployment extending the commercial autonomous-research-agent pattern.

Mathematicians formally settled a 50-year-old open problem using an algebraic-number-theory technique surfaced by an unreleased OpenAI model the prior week: first documented case of an undisclosed model's technique enabling human mathematicians to formally close a named conjecture.

DeepSWE benchmark documents AI agents autonomously resolving end-to-end software engineering tasks at measurable scale; Axiom auto-generated mathematical papers via AI: extends documented capability of autonomous systems closing named problems and generating original technical work beyond single-problem demonstrations.

Alibaba's Qwen3.7-Max ran 35 unattended hours of Triton GPU kernel optimization (432 evaluations, 1,158 tool calls, 10x geometric-mean speedup): the longest documented autonomous agent run and first data point past METR's 12-hour task-horizon ceiling; CuspAI and Kemira simultaneously searched 300 trillion candidate molecular structures, surfacing 5,000+ novel materials for PFAS removal from drinking water.

OpenAI reasoning model disproved Erdős's 1946 conjecture on planar unit distances, producing a small geometric construction human mathematicians had not found in eight decades; construction is human-verifiable, making this the first publicly-deployed reasoning system to overturn a named mainstream-literature conjecture.

TechCrunch reports OpenAI claims to have solved an 80-year-old math problem; Nature feature documents multi-agent AI scientist systems promising to accelerate research, extending documented pattern of AI closing named open problems across mathematics and science.

Google DeepMind UK workforce entered formal Acas mediation over disputed military and defense-related AI work, placing a frontier-lab employment and ethics dispute into the UK's statutory conciliation framework: first use of recognized labor arbitration infrastructure to contest AI lab military-contract commitments.

Nature published Co-Scientist (Google DeepMind) and Robin (FutureHouse) as multi-agent systems closing the full hypothesis-to-wet-lab-validation loop: Co-Scientist identified and validated a drug-repurposing candidate for acute myeloid leukemia; Robin proposed and validated a dry AMD treatment hypothesis, both in roughly a day of machine time with human-confirmed wet-lab results.

ChatGPT closed Erdős Problem 1196, a named open conjecture on covering systems of congruences, in roughly 80 minutes, with Sellke confirming the AI-contributed proof as valid; first documented closure of a named open math problem with AI as primary contributor.

Clark framed AI research as ~99% engineering work now being absorbed by models, with Anthropic’s internal AI training-code optimization benchmark rising from 2.9x to 52x in under a year.

OpenAI chief scientist Pachocki discloses roadmap: "autonomous AI research intern" by September 2026 capable of independent multi-day investigation; full multi-agent research system targeting problems "too large or complex for humans" by 2028; Codex (2M weekly active users) framed as "early version of the AI researcher."