AI scientist updates

Reverse-chronological log of changes to the AI scientists tracker. Newest at the top.

Older entries live in COSCIENTIST_CHANGELOG_ARCHIVE.md.

2026-08-08

Added

  • ChatBattery (Lifecycle: Multi-stage) — Mila / Université de Montréal with Oxford, UCL, Ottawa and NRC Canada; expert-guided LLM reasoning over eight stages and seven agents carried three novel cathodes from hypothesis through DFT to synthesis and coin-cell testing, at 174/169/160 mAh/g versus NMC811’s ~135 (arXiv:2507.16110)
  • Claw AI Lab (Lifecycle: Multi-stage, Writing) — NTU Singapore with A*STAR, Moxin, NUIST, Tsinghua and USTC; instantiates a full multi-agent research team from one prompt with dashboard monitoring, artifact inspection and rollback, and a Claw-Code Harness that wires local codebases and checkpoints into runnable experiments (arXiv:2605.22662)
  • MASTER (Lifecycle: Multi-stage) — Los Alamos National Laboratory Theoretical Division with University of Connecticut; hierarchical LLM agents design, execute and interpret DFT in an active-learning loop, cutting required atomistic simulations by up to 90% (arXiv:2512.13930)
  • SR-Scientist (Lifecycle: Multi-stage) — SJTU / Shanghai Innovation Institute / GAIR, ICLR 2026; promotes the LLM from equation proposer to an agent that writes analysis code and iterates on evaluation feedback, beating baselines by an absolute 6–35% on LSR-Synth (arXiv:2510.11661)
  • TourSynbio-Agent (Lifecycle: Multi-stage) — Toursun Synbio Shanghai with CityU Hong Kong, SJTU and Johns Hopkins; protein-specialized LLM routing to ESM-1v / ESMFold / AntiFold agents, with P450 variants at 70% improved 19-hydroxylation selectivity and reductases at 3.7× conversion (arXiv:2411.06029)

Updated

  • AutoResearchClaw — added Claw AI Lab (arXiv:2605.22662) to Other references; it uses AutoResearchClaw as its evaluation baseline.
  • Landscape — ChatBattery added to the chemistry-and-materials exemplars and SR-Scientist to the symbolic-discovery tail; synthesis re-verified against the current grouping and synthesis_reviewed bumped to 2026-08-08.

Flagged

  • None.

Verified (no changes)

  • Synthesis claims re-checked against the regrouped index table; no superlative required correction. The 55-entry link/repo re-verification backlog remains open — Phase B has no web tools, so last_verified was not bumped on those pages.

2026-08-01

Added

  • AIMS (Lifecycle: Multi-stage) — Stanford/SIMES uncertainty-aware AI experimentalist that operates a cryogenic microwave impedance microscope end to end on twisted bilayer MoSe2, cutting sample-navigation time from ~10 h to ~4 h and prioritizing a quantum-fluctuation-renormalized origin for the robust ν = 1/2 crystal (arXiv:2607.16544)
  • AI Sleep Co-Scientist (Lifecycle: Multi-stage) — Stanford expert-guided agent environment over ~124,000 polysomnography recordings, binding each result to executable code; incident Parkinson’s HR 1.48 and Alzheimer’s HR 1.38 from diminished network-level sleep coupling (arXiv:2607.25175)
  • NAIS (NVAITC AI Scientist) (Lifecycle: Multi-stage, Writing) — NVIDIA governed end-to-end research system keeping protected data inside institutional boundaries; agent-orchestrated hypertension GWAS on 286,422 individuals reproduced FGF5, ATP2B1, CNNM2, FTO, and GRB14 (arXiv:2607.11084)
  • OmniQEC (Lifecycle: Multi-stage) — NTU Singapore LLM-orchestrated quantum-error-correction discovery agent optimizing circuit-level logical error rate; beat the BB [[72,12,6]] and [[144,12,12]] codes at complete-implementation budgets of 98 and 240 physical qubits (arXiv:2607.25865)
  • CogEEGAgent (Lifecycle: Analysis) — Tohoku University cognitive-EEG agent separating LLM semantic authority from deterministic scientific authority; 39/40 routing accuracy vs 33 for a matched deterministic router, and held-out confirmation cut the null false-positive rate by 11.6 points to 4.9% (arXiv:2607.25045)

Updated

  • Landscape synthesis — added AIMS to the embodied physical-sciences exemplars and noted that the verifiability turn has spread beyond the ML cluster (CogEEGAgent’s held-out confirmation, NAIS’s governed data broker); synthesis_reviewed bumped to 2026-08-01.

Flagged

  • None.

Verified (no changes)

  • Aging-entry re-verification remains blocked in Phase B (no web tools); the 55-entry link/repo backlog stays logged under Deferred — next-run priority and last_verified was not bumped for those pages.

2026-07-25

No substantive updates — Phase A surfaced no new in-scope systems. Synthesis re-checked against the current grouping (synthesis_reviewed 2026-07-18, within the 30-day cadence; no systems added or removed) — all superlatives still hold, no refresh needed. The link/repo re-verification backlog (55+ entries past the 30-day window) remains deferred to Phase A, which alone has the web/MCP tools to confirm link liveness.

2026-07-18

Added

  • CoDHy (Lifecycle: Hypothesis) — Hannover Medical School (PLRI/CAIMed) with Sanford Burnham Prebys (Younis, Basak, Chavez, Ahmadi; arXiv:2603.00612). An interactive, human-in-the-loop AI co-scientist for biomarker-guided cancer drug-combination hypothesis generation. From a user-specified biomarker, cancer type, and PubMed scope it builds a task-specific Neo4j knowledge graph (structured databases — Reactome, CIViC, TCGA-GDC, ChEMBL, STRING, SynlethDB, DepMap, DrugBank, etc. — plus spaCy-extracted PubMed triples normalized via sentence-transformers), computes Node2Vec embeddings, and runs a hybrid Graph-RAG generation agent, a Llama-3.1-8B validation agent (novelty/plausibility/safety with targeted PubMed grounding → proceed/caution/reject), and a ranking agent (composite graph-evidence + safety score, optional DrugCombDB synergy). Benchmark-only across 7 frozen (biomarker, cancer type) scenarios: Full CoDHy highest exact novelty (35.71%) vs No-Node2Vec (28.57%) and LLM-only (10.71%), diversity 0.89, with a deliberately lower MRR (0.74) reflecting a discovery-over-retrieval bias. Assistive decision support; open source (source).

Updated

  • autonomous-science/summary.md — re-verified the synthesis against the current grouping (CoDHy joins the Biology & medicine hypothesis-generation cluster as a benchmark-validated system and tips no superlative — the “strongest evidence” claim still rests on wet-lab/clinical exemplars) and bumped synthesis_reviewed to 2026-07-18.
  • sources/manifest.json — added the CoDHy preprint (arXiv:2603.00612) with its .txt sidecar.
  • autonomous-science/curator-state.md — added CoDHy to Recently surfaced (trimmed to last 5).

Flagged

None.

Verified (no changes)

  • Synthesis re-reviewed against the freshly grouped table after adding CoDHy; all superlatives (“most loop-closed,” “strongest evidence,” “newest frontier”) still hold. The link/repo re-verification backlog (55+ entries past the 30-day window) remains deferred to Phase A — Phase B has no web/MCP tools to confirm link liveness, so last_verified dates were not bumped.

2026-07-11

Added

  • ARIA (Lifecycle: Hypothesis, Analysis) — Johns Hopkins University (Clancy, Van Durme, Yuille groups; KDD ‘26). A causal-aware LLM framework for trustworthy materials discovery that diagnoses “contextual tunneling” in knowledge-graph-augmented LLMs and gates knowledge use on Process-Structure-Property (PSP) mechanistic completeness, routing each query through a three-tier cascade (direct causal reasoning over complete PSP chains → physics-filtered analogical transfer → parametric fallback). Grounded in a 2,839-relation PSP Causal Knowledge Graph with optional real-time literature enrichment; covers forward property prediction and inverse synthesis-protocol design. Benchmark-only validation on 149 expert-curated 2D-materials tasks: ARIA-FULL +50.6% over baseline on in-domain forward prediction, sustained out-of-domain inverse-design performance, and beats a Self-RAG baseline. Open source (source).

Updated

  • autonomous-science/summary.md — added ARIA’s primary paper to the Sources list; re-verified the synthesis against the current grouping (ARIA fits the existing Chemistry & materials cluster as a benchmark-validated system and tips no superlative) and bumped synthesis_reviewed to 2026-07-11.
  • autonomous-science/curator-state.md — added ARIA to Recently surfaced (trimmed to last 5).

Flagged

None.

Verified (no changes)

  • Synthesis re-reviewed against the freshly grouped table after adding ARIA; all superlatives (“most loop-closed,” “strongest evidence,” “newest frontier”) still hold. The link/repo re-verification backlog (55+ entries past the 30-day window) remains deferred to Phase A — Phase B has no web/MCP tools to confirm link liveness, so last_verified dates were not bumped.

2026-07-04

No new systems surfaced — Phase A’s seed queries (arXiv, bioRxiv, medRxiv via papers MCP, plus WebSearch fallback) returned only out-of-scope work and already-catalogued systems, so the handoff was _None._.

Updated

  • autonomous-science/curator-state.md — refreshed the link/repo re-verification backlog to the 55 entries now past the 30-day window (last_verified 2026-05-20 through 2026-06-02).

Flagged

None.

Verified (no changes)

  • 55 entries have crossed the 30-day re-verification window (last_verified 2026-05-20 through 2026-06-02). Phase B has no web/MCP tools to confirm primary-paper-link or code-repo liveness, so last_verified was intentionally not bumped; the backlog is logged under Deferred — next-run priority for the next Phase A to fetch. summary.md synthesis last reviewed 2026-06-27 (7 days ago) and no systems were added or removed this run, so it remains within the 30-day cadence — no re-verification required.

2026-06-27

Added

  • ARTIQ-MCP (Duke trapped-ion agent) (Lifecycle: Experiment design, Analysis) — Duke Quantum Center (Brown and Linke groups) with University of Maryland. An LLM agent (Claude Opus 4.8) that autonomously writes native ARTIQ control code and runs it on real trapped-ion quantum hardware behind a per-call hardware safety filter: a safety-filter MCP proxy issues a content-bound, single-use authorization token only after an AST denylist check plus isolated dax.sim simulation that traces every operation against preset per-device bounds (block-by-default on unmapped devices), or a human token for sensitive actions; an artiq-mcp server then forwards approved calls to the ARTIQ master over sipyco. Wet-lab validation on a co-trapped 40Ca+/40CaOH+ crystal (full agent-built calibration stack; cross-instrument 60 Hz magnetic-field-stabilization loop) plus interface-level portability on an independent 171Yb+ ARTIQ platform; safety filter red-teamed with 1932 adversarial scripts. Open source (source).

Updated

  • autonomous-science/summary.md — ARTIQ-MCP joins the existing “Physical sciences and embodied systems are the newest frontier” cluster (Qumus, Qiushi, Dr.Sai, BioProVLA, CALMS), tipping no superlative: Biology & medicine remains the largest group (17) and strongest-evidence tier; Chemistry & materials (14) remains the most loop-closed. No synthesis prose was rewritten; synthesis_reviewed bumped to 2026-06-27 after re-checking every superlative against the current 67-system grouping.
  • autonomous-science/curator-state.md — added ARTIQ-MCP at the top of Recently surfaced (trimming Ax-Prover to keep the window at five); refreshed the link/repo re-verification backlog to the 37 entries now past the 30-day window.
  • sources/manifest.json — ARTIQ-MCP (DOI 10.48550/arXiv.2606.27231) entry added by Phase A.

Flagged

None.

Verified (no changes)

  • 37 entries have crossed the 30-day re-verification window (last_verified 2026-05-20 through 2026-05-27). Phase B has no web/MCP tools to confirm primary-paper-link or code-repo liveness, so last_verified was intentionally not bumped; the backlog is logged under Deferred — next-run priority for the next Phase A to fetch.

2026-06-20

Added

  • CALMS (Lifecycle: Experiment design, Analysis) — Argonne National Laboratory (Center for Nanoscale Materials + Advanced Photon Source; Vriza, Prince, Zhou, Chan, Cherukara) AutoGen-based multi-agent framework that operates two real scientific user facilities — a Hard X-ray Nanoprobe (HXN) beamline and an N9 robotic thin-film station. Specialized agents (code writer, code critic, administrator, paper scraper, image explainer, teachability) orchestrate multistep workflows, interpret multimodal nano-diffraction/nano-fluorescence images, and learn on the job by storing human guidance as input–output pairs in a ChromaDB vector store (with a similarity threshold to skip redundant memories). Live demos: natural-language → correct 2D-scan commands and cross-modality scan-region selection at HXN (only o3 reliably reasoned across modalities with high positional precision), and end-to-end fabrication of a defect-free PEDOT:PSS thin film after the literature-scraper agent extracted coating parameters (90 °C, 1 mm/s) from a PDF — teachability markedly improved long-horizon sequential success. Open source (source).

Updated

  • autonomous-science/summary.md — added the CALMS primary-paper link to Sources. CALMS fits the existing “Chemistry and materials are the most loop-closed” / embodied-instrument pattern (alongside MARS, AMASE, MAD, Qumus) and tips no superlative (Biology & medicine remains the largest group and strongest-evidence tier), so no synthesis prose was rewritten; synthesis_reviewed last refreshed 2026-06-11, within the 30-day cadence.
  • autonomous-science/curator-state.md — added CALMS at the top of Recently surfaced (trimming SAGA to keep the window at five); deferred BioMedAgent and logged the link/repo re-verification backlog.
  • sources/manifest.json — CALMS (DOI 10.48550/arXiv.2509.00098) entry added by Phase A.

Deferred

  • BioMedAgent / CAS (Bu et al., Nat. Biomed. Eng., doi:10.1038/s41551-026-01634-6, 2026) — named, benchmark-validated autonomous biomedical-analysis agent (CAS; BioMed-AQA, 327 tasks, ~77% success; generalizes to BixBench), in-scope as an analysis-stage system. Phase A located no openly downloadable PDF (closed access, no preprint), so Phase B could not ground a page; deferred until a citable open source is available (source).

Flagged

None.

Verified (no changes)

  • 19 bootstrap entries crossed the 30-day window today (last_verified 2026-05-20/21). Phase B has no web/MCP tools to confirm link/repo liveness, so last_verified was intentionally not bumped; the backlog is logged under Deferred — next-run priority for the next Phase A to fetch.

2026-06-13

No new systems surfaced — seed queries (arXiv, bioRxiv, medRxiv via papers MCP, plus WebSearch fallback) returned only out-of-scope work and already-catalogued systems.

Deferred

  • EurekAgent (arXiv:2606.13662, Tsinghua + Zhipu AI) — metric-driven autonomous-discovery agent (“environment engineering”; open-sourced) with new SOTA on circle packing, an autocorrelation inequality, a TriMul kernel, and an MLE-Bench subset for ~$11 API cost. Scope-edge: an optimization/discovery substrate on math/kernel/ML benchmarks, not natural-science hypothesis generation, experiment design, or data analysis. PDF archived and logged in the manifest; deferred pending a more science-leaning case (source).

Verified (no changes)

  • All system pages remain inside the 30-day re-verification window (oldest verified 2026-05-20, 24 days ago); no spot-checks required. summary.md synthesis last reviewed 2026-06-11 (2 days ago), within cadence — no re-verification needed.

2026-06-11

Added

  • ATLAS (Lifecycle: Multi-stage) — Google DeepMind (Éltető, Daw, Stachenfeld, Miller; with Princeton/Columbia/UCL) “Active Theory Learning for Automated Science,” an active-learning framework — not an LLM orchestration — that closes the hypothesis-generation ↔ experiment-design loop to discover interpretable mechanistic models of behavior. It iterates a Hypothesis Generator (an ensemble of sparse Disentangled RNNs whose latent-variable interactions form candidate computational graphs), an Experiment Optimizer (hill-climbs binary reward-matrix designs to maximize ensemble disagreement / expected information gain, BALD-style), and an Experiment Runner that executes the chosen design. Validation is in-silico (benchmark-tier): recovering Q-learning and Leaky Actor-Critic agents from bandit behavior, scored on behavioral, structural, and dynamical (bisimulation) similarity. 5–10× sample-efficiency gain over random experimentation, 8/8 correct computational-graph recovery in 100 experiments where baselines needed ~1,000, and matched or surpassed expert-designed experiments. Adds a non-LLM, cognitive-science exemplar to the ML & scientific-computing cluster. No code released (source).

Updated

  • autonomous-science/summary.md — added the ATLAS primary-paper link to Sources; re-verified every synthesis claim against the current grouping and bumped synthesis_reviewed to 2026-06-11. ATLAS fits the existing “ML & scientific computing is a large, benchmark-validated cluster” pattern and tips no superlative (Biology & medicine remains the largest group and strongest-evidence tier), so no synthesis prose was rewritten.
  • autonomous-science/curator-state.md — added ATLAS at the top of Recently surfaced (trimming OriGene to keep the window at five).
  • sources/manifest.json — ATLAS (DOI 10.48550/arXiv.2606.12386) entry added by Phase A.

Flagged

None.

Verified (no changes)

  • All pre-existing system pages remain inside the 30-day re-verification window (oldest additions verified 2026-05-20, 22 days ago); no spot-checks required this run.

2026-06-10

Added

  • DarkAgents (Lifecycle: Multi-stage) — Università di Bologna / INFN language-driven multi-agent system for theoretical astroparticle physics (TAP), and the first end-to-end architecture targeting that domain. An orchestrator interprets a particle-physics model or looser “idea,” selects a pipeline branch, writes an execution plan, and dispatches specialized sub-agents (proposal, librarian, critic, plus deterministic compute stages) that each emit a Markdown report and a fixed-schema JSON handoff it checks before proceeding; it pauses for human audit after each step by default but can run fully autonomous. All physical quantities come from deterministic human-validated code to curb hallucination, and the workflow is LLM-agnostic (Mistral, Anthropic/Claude Code, OpenAI/Codex, local Ollama). The first implementation, DarkAgent-PT, takes a classically scale-invariant model to a PTArcade MCMC fit of the NANOGrav nanohertz gravitational-wave background, then adds a constraint sub-agent and an assumption/prior-auditing sub-agent. Validation: reproduced human Bayesian posteriors across providers (Claude Code Opus 4.8 and Codex GPT-5.5 ran almost autonomously), identified inconsistencies in some published fits and produced novel fits on the dissipative bulk-flow GW template, and correctly rejected the sound-wave template where invalid; a noted failure mode is hallucinated references in the final report. Open source (source).

Updated

  • autonomous-science/summary.md — added the DarkAgents primary-paper link to Sources; re-verified every synthesis claim against the current grouping and bumped synthesis_reviewed to 2026-06-10. DarkAgents fits the existing “Physical sciences are the newest frontier” pattern (alongside Dr.Sai and CMBEvolve/CosmoEvolve as in-silico TAP/cosmology work, not embodied apparatus) and tips no superlative, so no synthesis prose was rewritten.
  • autonomous-science/curator-state.md — added DarkAgents at the top of Recently surfaced (trimming MLEvolve to keep the window at five).
  • sources/manifest.json — DarkAgents (DOI 10.48550/arXiv.2606.11157) entry added by Phase A.

Flagged

None.

Verified (no changes)

  • All pre-existing system pages remain inside the 30-day re-verification window (most recent additions verified 2026-06-05 through 2026-06-09); no spot-checks required this run.

2026-06-09

Added

  • LabOS (Lifecycle: Multi-stage) — Stanford / Princeton AI-XR co-scientist (with Oregon State, U. Washington, NVIDIA) that couples a self-evolving multi-agent digital-lab system with extended-reality smart glasses, a lab-specialized vision-language model, 3D/4D digital twins, and a cobot module to perceive, reason about, and assist in the physical laboratory. The dry-lab module extends the STELLA framework (Manager/Planner, Developer, Critic agents plus a Tool-Creation agent feeding a shared “Tool Ocean”) for hypothesis generation, experiment design, and analysis; the wet-lab module streams egocentric video to LabOS-VLM (Qwen-VL post-trained via SFT + GRPO) for real-time action verification, error detection, and step guidance. Benchmarks: ~32% HLE: Biomedicine, 61% LAB-Bench: DBQA, 65% LAB-Bench: LitQA (up to 8% over next-best), plus a new LabSuperVision (LSV) lab-video benchmark on which LabOS-VLM-235B exceeds 90% error-detection accuracy, beating Claude Opus-4.1, GPT-5, and Gemini 2.5 Pro. Wet-lab: agent-nominated CEACAM6 confirmed as an NK-cell anti-tumor target in a physical killing assay; ITSN1 identified as a cell-fusion regulator. Open source (source).

Updated

  • autonomous-science/summary.md — added the LabOS primary-paper link to Sources. LabOS fits the existing “Biology & medicine carry the strongest evidence” and embodied-systems patterns and tips no superlative; synthesis_reviewed was refreshed yesterday (2026-06-08, within the 30-day window), so no synthesis prose was rewritten.
  • autonomous-science/curator-state.md — added LabOS at the top of Recently surfaced (trimming CatDT to keep the window at five).
  • sources/manifest.json — LabOS (DOI 10.48550/arXiv.2510.14861) entry added by Phase A.

Flagged

None.

Verified (no changes)

  • All pre-existing system pages remain inside the 30-day re-verification window (oldest last_verified is 2026-05-20); no spot-checks required this run.