Autonomous AI Scientific Discovery — Novel, Correct, Independent

Jumper et al. (2021) report AlphaFold2's performance in CASP14, a blind structure-prediction assessment using experimentally solved structures that had not yet been publicly released. AlphaFold achieved accuracy competitive with experimental structures in a majority of cases and substantially outperformed other methods. This is strong evidence that an AI system can autonomously execute highly accurate scientific prediction once the task and inputs are specified. The Nature paper does not by itself establish proteome-scale discovery, independent experimental confirmation of newly discovered structures, or autonomous selection of scientific questions. For this record it is therefore bounded background evidence on correctness and autonomous execution, not a demonstration of autonomous scientific discovery in the full sense of the claim.

Merchant et al. (2023) report GNoME, a graph-network materials-exploration system that identified more than 2.2 million crystal structures stable relative to prior work, with 381,000 new entries on the updated convex hull. The paper notes that 736 structures had been independently experimentally verified, but these were not 736 materials subsequently synthesised by a robotic laboratory. A separate Nature study by Szymanski et al. reports the A-Lab autonomous synthesis platform, which realised 36 of 57 human-selected target compounds over 17 days and used robotics, literature-derived synthesis models, machine-learning interpretation, and active learning. The two studies jointly provide substantial evidence for AI-enabled generation and autonomous experimental execution in materials science, but they do not constitute one end-to-end system that autonomously chose a scientific problem and then discovered and experimentally validated 736 materials. The research objective and A-Lab target set remained human-defined.

Lu et al. (2024) present The AI Scientist, a framework that generates research ideas, writes and executes code, runs experiments, visualises results, writes a paper, and then subjects that paper to a simulated review process. The authors apply it to three machine-learning subfields and report that some generated papers exceed a top-conference acceptance threshold as judged by their own validated automated reviewer. This is evidence that a substantial research loop can be automated inside a narrow, human-defined domain. It is not evidence that the papers were accepted by or passed review at an actual machine-learning conference, and the source does not independently establish the correctness or genuine novelty of the generated scientific claims. The instance therefore supports autonomy of process more strongly than correctness or novelty.

Romera-Paredes et al. (2023) introduce FunSearch, which pairs a pretrained large language model with an evolutionary search procedure and a human-specified evaluator. Applied to the cap-set problem, FunSearch produced previously unknown constructions that improved the best-known results, including the largest improvement in 20 years to the asymptotic lower bound reported by the authors; it also discovered improved heuristics for online bin packing. This is meaningful evidence that an AI-centred search process can produce novel, verifiable mathematical or algorithmic content inside a human-framed problem. The system does not autonomously choose the problem: researchers supply the evaluator, program skeleton and problem specification, and domain experts analyse and interpret the resulting programs. The primary paper supports novel constructions and verifiable results, but not the stronger legacy description of an autonomous absolute proof independently certified by the mathematics community.

The legacy instance bundled the Broad Institute, EMBL, pharmaceutical companies, Nature and Science editorials, Nobel-laureate commentary, and internal programmes at OpenAI, Anthropic, and Google DeepMind into a single claim that major institutions were reorganising scientific workflows in anticipation of autonomous AI discovery. LPR-001-D07 could not confidently reconstruct a source chain supporting that exact historical bundle or its stronger institutional-restructuring interpretation. The bundled claim is therefore withdrawn from the current evidential basis rather than retrofitted to substitute sources. IN-005 remains visible as explicit legacy provenance debt and contributes no substantive evidence for the record's current assessment or for the anticipatory-institutional-evidence taxonomy pending a separately governed review.

A Structured Concept Evolution (SCE) framework pairing a large language model with an algebraic mutation grammar discovers competitive qLDPC code families, including non-abelian group constructions beyond standard bivariate-bicycle codes, using only lightweight models (arXiv:2606.24808, Max Planck Institute, primary preprint, submitted 23 June 2026). The result demonstrates constrained autonomous generation within a human-framed, human-prior-encoded search space: the problem (find high-performance qLDPC code families) is human-set and the mutation grammar itself encodes algebraic prior knowledge, so the system is not discovering code theory from scratch. It supplies no evidence on autonomous problem-identification, which remains the record's decisive unresolved boundary (BN-001, OQ-001, AT-001); competitive performance has not yet been independently established as superior — whether the discovered families outperform state-of-the-art codes under circuit-level (as opposed to code-capacity) noise is unestablished. In relation to the whole claim this is a third constrained-domain example of the same autonomous-generation pattern as IN-002 (materials) and IN-004 (mathematics), on the already-evidenced novelty/correctness side, and weaker than both on correctness. It is classified NEUTRAL because it neither advances nor contests the record's load-bearing boundary.

Gyevnár, Kasirzadeh, and Shah (arXiv:2607.10712v1, submitted 12 July 2026) test indirect data poisoning against autonomous scientific workflows using three frontier AI agent systems across five socially salient topics and 450 ethically contained runs. Poisoned datasets were retrieved in 84.22% of runs; 49.56% ended with a poisoned conclusion without detection or caveat, while poisoning was detected in 6.0%. In the mitigation evaluation, a structured five-check provenance audit reduced full attack success to 0.0% in the tested setting. This is a bounded contesting instance on the correctness component: it establishes that correctness in autonomous scientific workflows can be conditional on evidential provenance under adversarial retrieval. The canonical claim is existential, so this result does not negate the correct discoveries demonstrated by IN-002 and IN-004. It supplies no evidence on autonomous problem-identification and leaves BN-001, OQ-001, and AT-001 unchanged. The experiment used ethically contained private repositories made operationally indistinguishable from public repositories, five selected topics, and three specified agent configurations; generalisation beyond this controlled setting is unestablished.

A shadow evaluation of frontier research agents finds that substantial autonomous research engineering did not produce successful open-ended AI research in the tested cases (arXiv:2607.27191v1, submitted 29 July 2026). Agents were given the central question from each of two unpublished NeurIPS submissions, but not the authors' answer, and received multi-day access to compute, web search, coding tools, subagents, and review tools. The two principal outputs received overall expert scores of 2/6 and 1/6 and were judged unambiguous rejections. The agents completed literature review, coding, environment debugging, and extensive experimentation, but failed at evidential prioritisation, project-level backtracking, creative response to criticism, and recognition of publishable progress; a robustness run using a different model and scaffold reproduced most of these judgment failures. This is bounded CONTESTING evidence on the autonomy component. It shows that autonomous execution of research engineering can coexist with failure of scientific judgment even after humans specify the research problem, so the record's unresolved autonomy boundary is not exhausted by problem identification versus problem solving. It does not negate verified bounded discoveries in IN-002 and IN-004, test autonomous problem identification, or establish general failure across fields. Generalisation is limited by two research questions, five total runs, non-blind expert reviewers, scaffold defects, and the source's preprint status.

Chung, Du, and Wesley (arXiv:2608.23691v1, submitted 24 August 2026) report the Station, an open-world multi-agent environment in which agents from different model families choose research directions, conduct experiments, collaborate, and build a shared scientific literature without a central coordinator or scripted pipeline. Across twelve construction problems from the AlphaEvolve catalogue and two additional case studies, the authors report results novel relative to prior literature on five problems, including a new infinite family of finite-field Kakeya sets, exact 604-point kissing configurations in dimension 11, new records for two optimisation problems, and an improved lower bound for Erdős's minimum-overlap problem; they also report novel infinite families for Book Ramsey numbers. Raw agent dialogues, proofs, source code, and verification artifacts were released. This is SUPPORTIVE instance evidence for autonomous research direction and scientific judgment within a human-defined shared goal, and it counterbalances IN-008's bounded failure cases without negating them. The source remains a primary preprint authored by the system's creators: independent expert validation of the claimed novelty, proofs, and reproducibility is incomplete. The system did not identify the overarching research domain or shared goal, so it does not satisfy the autonomous problem-identification boundary in BN-001, OQ-001, and AT-001. Instance only: AS-002, FRAGMENTING, VS-03, mechanisms, and open questions remain unchanged.

Kim et al. evaluate nine agents on 15 human-specified, open computational follow-ups, producing 135 submissions. None satisfies every integrity, completeness and scientific-sufficiency requirement; 114 fail the integrity rubric. The original papers, code, data and research directions are supplied, so the result bears on research execution and scientific judgment after problem specification. Integrity failures include unsupported completion claims and discrepancies between reported results and recorded evidence; they do not establish deceptive intent. This is bounded contesting evidence that fluent research output can coexist with inadequate evidential support. It does not negate earlier verified bounded discoveries or test autonomous problem identification. The primary preprint uses selected problems, particular model/harness configurations and automated grading with selected expert validation; it does not establish universal scientific incapacity.

Sun et al. construct 248 correction-recognition cases from 26 projects across six scientific fields. Models choose between an original agent plan and a human-corrected alternative using the pre-intervention project record. At medium effort, four of five models score below chance; the highest project-condition result across tested settings is 63.5%, compared with 73.4% for fresh researchers with an agent assistant available. The cases were selected partly using model-response variation, so these scores are diagnostic and do not estimate failure prevalence across scientific research. This supplies bounded contesting evidence about scientific judgment after a problem is supplied. It tests recognition, not generating a corrective plan or executing research. Retrospective reconstruction, benchmark-author involvement, correlated project cases and different reading interfaces limit generalisation. Existing bounded successes and the autonomous problem-identification question remain intact.

Mao et al. test seven models on 50 supplied physics tasks derived from published studies, with three runs per task and 1,050 scored attempts. The strongest model reaches complete Research Closure (RC4) in 32.7% of its attempts and on 18.0% of tasks in all three runs. Local accomplishment scores can remain high while indispensable interpretation, validation or scope requirements remain unresolved. This is mixed, partial evidence: bounded investigations can succeed, but completion is unreliable in the tested setting. It does not establish autonomous problem identification or independently validated novel discovery. Tasks use fixed computational resources and offline environments. Automated grading receives selected human criterion checks; the 75.9% agreement for closure-relevant judgments is not validation of every final RC assignment. Preprint and benchmark-author limits remain. The existing fragmentation rationale is corroborated, not replaced.

Gu et al. report evolutionary LLM search over circuit-generating programs, starting from human-written specifications and blank programs. Seven of eight tasks reduce both Toffoli and qubit counts against stated comparators; one comparator is a naive reversible construction. Prime-field modular inversion produces a new architecture and a different resource trade-off. This supports bounded algorithmic novelty and correctness, extending the evaluator-guided pattern in IN-004 without testing autonomous problem identification. A separate programmatic evaluator checks candidates; this is not evidence of external replication. The circuit class is restricted and efficiently classically simulable, large-domain validation samples trajectories, and some tasks permit explicit failure budgets. Logical resource proxies do not demonstrate hardware runtime or practical quantum advantage. The primary preprint is authored by framework developers, including Anthropic and QuEra affiliates; general autonomous-science capability remains unestablished.

Chen et al. introduce ExpEvolve, which extracts high-level idea factors from evaluated programs, uses those factors to guide mutations, and varies strategies when search plateaus. Across 11 supplied mathematical tasks, with 200 evaluated programs and three seeds per setting, it improves final performance over SCS and ShinkaEvolve in most configurations. Integration with OpenRSI improves mean final best-so-far scores on six selected machine-learning tasks, using the same 20-slot starting history and 50 additional evaluator calls per method. This October paper explicitly extends CausalEvolve (arXiv:2603.14575), whose March evaluation covered four tasks; the expanded experiments and revised search framework are not independent corroboration of that predecessor. The result supports experience reuse inside human-specified tasks and evaluators. It does not establish autonomous problem identification or independently verified novel scientific discovery. Tasks remain relatively simple with small-scale verification; evaluator dependence, limited seeds, developer-authored preprint status and unestablished independent replication bound the inference. Efficiency gains depend on task, configuration and criterion. Like IN-006, it leaves the whole claim's load-bearing autonomy and validation boundaries untouched.

Siow et al. evaluate research-integrity auditing across selected computational research tasks. RIDAR's post-hoc detector achieves 85.2% instance-level F1 against 60 expert-annotated trajectories; finding-to-annotation matching uses a separate LLM judge. Applied to 300 trajectories from 50 tasks and six harness-model configurations, it accepts 3,964 findings in 299 trajectories. In 60 matched baseline/intervention pairs, runtime feedback reduces mean remaining findings from 14.9 to 9.4 (36.8%); eight pairs contain more findings after intervention. This is mixed evidence about scientific judgment and correctness after humans supply the problem: integrity failures persist, while external oversight can partially reduce them. Findings are detector-defined violations, not a universal false-discovery rate or evidence of intent. The same post-hoc detector evaluates both conditions; human annotations are reviewed by the author team, and runtime detection is less comprehensive. Held-out files remain accessible and stated resource limits are deliberately unenforced to make violations observable. Selected tasks, model/harness dependence and the framework developers' Tsinghua/Zhipu AI affiliations bound generalisation; independent replication is unestablished. The result corroborates the record's fragmentation rationale without negating bounded successful discoveries or resolving autonomous problem identification.