{
  "schemaVersion": 1,
  "kind": "leviathan-editorial-claim-revision",
  "scope": "The claim and its cited source cards at capture. External papers and linked project pages are not frozen by this record. A digest identifies bytes, not truth or endorsement.",
  "claim": {
    "id": "C12",
    "title": "How could Leviathan improve its own ways of learning?",
    "statement": "Leviathan should be able to generate and test new questions, concepts and methods, then use what it learns to revise how future learning happens.",
    "kind": "Discovery and self-improvement hypothesis",
    "status": "Research ambition with testable parts",
    "summary": "We want participants to find questions, concepts, and useful methods that were not written into the system in advance. A connection between distant fields might make a new experiment possible; a failed explanation might reveal a missing measurement. Testing and sharing those results could improve both practical work and the process that produced it.",
    "definitions": [
      {
        "term": "A research shadow",
        "meaning": "A retained gap, contradiction or unexplained pattern that may guide further inquiry. This research use is distinct from other meanings of shadow in Leviathan’s documents."
      },
      {
        "term": "Method learning",
        "meaning": "Acquiring or developing a reusable way to ask, measure, compare, build or decide, with evidence about where it works."
      },
      {
        "term": "Recursive self-improvement",
        "meaning": "Improvement of capabilities or processes that themselves help produce later improvements. The ambition includes this feedback; each claimed improvement still needs its own evaluation."
      }
    ],
    "values": [
      "Make room for original questions and unexpected connections.",
      "Preserve failed attempts and unresolved gaps when they can teach something.",
      "Let other participants examine improvements and choose whether to adopt them."
    ],
    "reasoning": [
      "A growing archive becomes more useful when it helps generate a question, prediction, or method that was previously missing. Connections should be judged partly by what they let us discover or do.",
      "A shadow can direct attention to cases our current representation treats alike even though they may differ. New observations, instruments, or concepts could make the distinction learnable.",
      "A method from another field may provide a way forward. Creative analogy proposes the connection; domain knowledge and testing determine which assumptions travel with it. The analogy should yield something to examine, not just similar words.",
      "A useful result or a concrete failure could change a concept, tool, or working procedure. In our local review exchange, an omitted decision led to an explicit requirement to include the exact decision in later review packets. This records a procedure revision; whether it improves later work remains to be tested.",
      "The ambition extends to improving the methods that produce further learning. Another Leviathan could find that a revision works only locally, improve it further, or reject it. Transfer, disagreement, and continued inquiry are part of that research.",
      "A useful candidate-generation process may still narrow the range of ideas considered. We propose measuring both the quality of an individual result and the diversity of approaches kept available."
    ],
    "limits": "The proposed loop is broader than the mechanisms demonstrated in any cited study. Better records, changed agent software and altered model representations are different changes. A new version or a higher score alone does not show improved discovery.",
    "strongestObjection": "The system could reward its own preferred questions and make its evaluations easier to satisfy. Apparent progress may come from extra compute, hidden expert work or memorized cases. Shared revisions can also spread a blind spot more efficiently.",
    "whatWouldChangeOurView": "If gains vanish on unfamiliar tasks, under an unchanged evaluation or after counting total resources, we should withdraw the improvement claim. Methods that other participants can use successfully under their own conditions would provide stronger support.",
    "nextQuestion": "Can a participant discover one useful method and teach it to another, then improve the way that method was discovered without weakening the criteria used to judge it?",
    "evidence": [
      {
        "sourceId": "R-LRN-01",
        "relation": "motivates",
        "reason": "BootLoops offers a first-person account of scientific methods crossing fields with expert guidance. It motivates creative transfer; it does not establish autonomous general discovery."
      },
      {
        "sourceId": "R-LRN-02",
        "relation": "supports-part",
        "reason": "GOLLuM adapts representations using observed outcomes and guides candidate selection through uncertainty. This is a concrete learning mechanism within a narrower optimization setting."
      },
      {
        "sourceId": "R-LRN-04",
        "relation": "context",
        "reason": "Predictive state representations describe state through predictions of future observations under actions. Relating this to research shadows and new distinctions is our proposed connection."
      },
      {
        "sourceId": "R-AG-04",
        "relation": "supports-part",
        "reason": "Bounded self-modification of agent software offers one example of evaluating revisions. Fixed underlying weights and specific coding tasks limit the inference to broader recursive improvement."
      },
      {
        "sourceId": "R-AG-05",
        "relation": "limits",
        "reason": "Comparing optimization at equal expenditure highlights the need to count compute and human effort when judging whether a learning method improved."
      },
      {
        "sourceId": "R-AG-09",
        "relation": "supports-part",
        "reason": "Evolving candidate programs with an evaluator is one implemented discovery mechanism. Choosing reliable evaluators and worthwhile questions remains necessary."
      },
      {
        "sourceId": "R-LRN-06",
        "relation": "limits",
        "reason": "The story-writing experiment separates individual benefit from collective diversity. Whether a similar tradeoff occurs in our method exchange remains a question."
      },
      {
        "sourceId": "R-AG-10",
        "relation": "supports-part",
        "reason": "Separating explicit evidence state from evolving conclusions helps propagate later contradictions across long-horizon reasoning tasks (reviewed via indexed digest)."
      },
      {
        "sourceId": "R-AG-11",
        "relation": "supports-part",
        "reason": "Pairing disagreement resolution with an explicit consensus challenger improves long-horizon verification when unchallenged rollouts share errors (reviewed via indexed digest)."
      },
      {
        "sourceId": "R-AG-12",
        "relation": "supports-part",
        "reason": "Backward revision of stored reflections based on downstream task outcomes offers a testable mechanism for improving non-parametric memory over time (reviewed via indexed digest)."
      },
      {
        "sourceId": "R-LRN-07",
        "relation": "limits",
        "reason": "Accumulating every past attempt in active working context reduced discovery performance by causing agents to anchor on local edits, whereas periodic context resets improved results by 15%–39% (reviewed via indexed digest)."
      }
    ],
    "history": [
      {
        "date": "2026-10-03",
        "note": "Editorial synthesis by Codex, following Mimar’s founding direction; offered for public criticism, with no community adoption implied. Added discovery and revision of learning methods as a foundational question. Connected creative transfer and research shadows to scoped scientific examples, while retaining the earlier empirical note on self-modification."
      },
      {
        "date": "2026-10-04",
        "note": "Connected creative transfer and method revision to the actual missing-context development case. Distinguished a recorded working-procedure change from demonstrated improvement, autonomous discovery, or durable learning. Preserved the wider self-improvement ambition and all source evidence."
      },
      {
        "date": "2026-10-04",
        "note": "Reviewed selected primary-source sections proposed in the external assessment and added scoped connections, access limits and research questions. Interlat retains its existing source ID; its reviewed preprint is distinguished from the final conference text. No experiment was reproduced, claim status promoted or governance rule adopted."
      },
      {
        "date": "2026-10-21",
        "note": "Added LongHarness (R-AG-10), VeriHarness (R-AG-11), ReHarness (R-AG-12), and Scaling Discovery Requires Forgetting (R-LRN-07), all reviewed via the curated DAIR.AI 2026 technical digest without independent reproduction. Distinguished archival failure retention from active working-context resets during search."
      }
    ],
    "projectLinks": [
      {
        "label": "How the parts connect",
        "href": "/what-is"
      },
      {
        "label": "A revised local working procedure",
        "href": "/research#what-changed"
      },
      {
        "label": "Evidence about bounded self-improvement",
        "href": "/library/claims/C04"
      },
      {
        "label": "Language and learning work",
        "href": "/work"
      }
    ]
  },
  "sources": [
    {
      "id": "R-LRN-01",
      "title": "Claude-shaped science",
      "url": "https://www.anthropic.com/research/claude-shaped-science",
      "authors": "Matthew Schwartz; guest research account published by Anthropic",
      "publishedAt": "2026-10-01",
      "version": "First-person account published 1 October 2026; describes work over the preceding summer and approximately three months",
      "type": "Research account",
      "reviewLevel": "Selected sections: cross-field methods, expert redirection, workflow and limitations",
      "finding": "Schwartz describes building reusable computational tools with Claude and applying methods across scientific fields. Domain experts redirected technically successful calculations toward questions they considered scientifically valuable.",
      "limitations": "This is a participant's account, not an independent replication of every reported result. The workflow required substantial human direction and resources. Its advantages do not establish general autonomous scientific judgment.",
      "changes": "Motivates testing whether a method learned in one setting becomes useful elsewhere, and separating computational correctness from the importance of the question being answered.",
      "checkedAt": "2026-10-03",
      "verificationStatus": "Selected primary-source sections reviewed; underlying projects not reproduced",
      "relatedUrls": []
    },
    {
      "id": "R-LRN-02",
      "title": "Large language models as uncertainty-calibrated optimizers for experimental discovery",
      "url": "https://www.nature.com/articles/s42256-026-01283-z",
      "authors": "Bojana Ranković, Ryan-Rhys Griffiths and Philippe Schwaller",
      "publishedAt": "2026-08-28",
      "version": "Journal publication: 28 August 2026; preprint first posted 8 April 2025, revised to v3 on 7 November 2025",
      "type": "Journal article",
      "reviewLevel": "Abstract and selected main-text methods and benchmark sections; preprint version history",
      "finding": "GOLLuM couples a language encoder with a Gaussian process. Training on observed outcomes adapts representations, while uncertainty guides the next candidate selection. The authors report improved search performance across chemistry and materials benchmarks.",
      "limitations": "The evidence concerns benchmark optimization under specified budgets and candidate spaces. It is not a new autonomous wet-lab campaign or evidence of open-ended self-improvement. Supplementary methods and code were not audited here.",
      "changes": "Provides a concrete example of outcome-driven representation learning. This motivates a test of whether changing a representation improves subsequent predictions or choices; it does not show that adding contextual prose has the same effect.",
      "checkedAt": "2026-10-03",
      "verificationStatus": "Selected primary-source sections reviewed; results not reproduced",
      "relatedUrls": [
        "https://arxiv.org/abs/2504.06265"
      ]
    },
    {
      "id": "R-LRN-04",
      "title": "Predictive Representations of State",
      "url": "https://papers.neurips.cc/paper/1983-predictive-representations-of-state.pdf",
      "authors": "Michael L. Littman, Richard S. Sutton and Satinder Singh",
      "publishedAt": null,
      "version": "NIPS 2001, Advances in Neural Information Processing Systems 14; exact publication day not established",
      "type": "Conference paper",
      "reviewLevel": "Abstract and introductory formulation; proceedings metadata",
      "finding": "The paper represents the state of a controlled dynamical system using predictions of future observations conditioned on action sequences. It develops a linear predictive formulation and compares its representation capacity with other state models.",
      "limitations": "This is foundational representation theory, not an LLM or value-learning study. Its theoretical assumptions do not establish how to identify every relevant observation or represent a moral disagreement.",
      "changes": "Offers a conceptual connection for asking whether a new observation distinguishes situations that an existing representation treats alike. Applying that idea to Leviathan's proposed shadows remains a design hypothesis.",
      "checkedAt": "2026-10-03",
      "verificationStatus": "Primary abstract and introductory formulation reviewed; proofs not independently checked",
      "relatedUrls": [
        "https://proceedings.neurips.cc/paper_files/paper/2001/hash/1e4d36177d71bbb3558e43af9577d70e-Abstract.html"
      ]
    },
    {
      "id": "R-AG-04",
      "title": "Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents",
      "url": "https://arxiv.org/html/2505.22954v3",
      "authors": "Jenny Zhang, Shengran Hu, Cong Lu, Robert Lange and Jeff Clune",
      "publishedAt": "2025-05-29",
      "version": "arXiv v3: 12 March 2026",
      "type": "Preprint",
      "reviewLevel": "Selected sections",
      "finding": "An agent modifies its own software and selects changes using coding evaluations, improving on two benchmarks. An archive of different past solutions helps the search.",
      "limitations": "The underlying model weights stay fixed. Bounded coding experiments do not establish open-ended improvement of model training, unlimited recursive improvement, or an AGI timetable.",
      "changes": "Read alongside the METR cost framework: measured gains, total expenditure and independent evaluation answer different questions.",
      "checkedAt": "2026-09-30",
      "verificationStatus": "Primary source reviewed",
      "relatedUrls": [
        "https://arxiv.org/abs/2505.22954",
        "https://sakana.ai/dgm/"
      ]
    },
    {
      "id": "R-AG-05",
      "title": "Expenditure Horizon: Measuring Optimization Ability, with an Application to NanoGPT",
      "url": "https://metr.org/blog/2026-07-21-expenditure-horizon/",
      "authors": "Tom Cunningham, Manish Shetty, Vincent Cheng and Nate Rush; METR",
      "publishedAt": "2026-07-21",
      "version": "Research report",
      "type": "Research report",
      "reviewLevel": "Selected sections",
      "finding": "The report proposes comparing human and agent optimization at equal expenditure, illustrated with NanoGPT. It counts experimental compute and human effort alongside model usage.",
      "limitations": "The human comparison is estimated, the agent results are preliminary, and the study covers one optimization problem. It does not directly measure the returns from human–AI collaboration.",
      "changes": "More generated code or a higher benchmark score alone does not establish faster research or economic advantage.",
      "checkedAt": "2026-09-30",
      "verificationStatus": "Primary source reviewed",
      "relatedUrls": []
    },
    {
      "id": "R-AG-09",
      "title": "AlphaEvolve: A coding agent for scientific and algorithmic discovery",
      "url": "https://arxiv.org/abs/2506.13131v1",
      "authors": "Alexander Novikov and co-authors",
      "publishedAt": "2025-06-16",
      "version": "arXiv:2506.13131v1; white paper submitted 16 June 2025",
      "type": "Research white paper",
      "reviewLevel": "Selected sections",
      "finding": "An evolutionary pipeline combines LLM-generated code with supplied evaluators, including evolving search algorithms. Reported results include a 48-multiplication algorithm for two 4-by-4 complex matrices and optimizations to Google's computational infrastructure.",
      "limitations": "The main boundary is availability of automated evaluators. The paper describes self-improvement feedback as modest and occurring over months. It does not demonstrate unrestricted recursive improvement, ethical goal selection or Leviathan's architecture. Results and deployments were not independently reproduced here.",
      "changes": "Read v1 task definition, code search, evaluation, selected results and discussion. Reported code and deployments were not reproduced; supplied evaluators constrain the demonstrated discovery process.",
      "checkedAt": "2026-10-04",
      "verificationStatus": "Primary source reviewed; no independent reproduction",
      "relatedUrls": [
        "https://arxiv.org/html/2506.13131v1",
        "https://doi.org/10.48550/arXiv.2506.13131"
      ]
    },
    {
      "id": "R-LRN-06",
      "title": "Generative AI enhances individual creativity but reduces the collective diversity of novel content",
      "url": "https://doi.org/10.1126/sciadv.adn5290",
      "authors": "Anil R. Doshi and Oliver P. Hauser",
      "publishedAt": "2024-07-12",
      "version": "Science Advances 10(28), eadn5290; published version, 12 July 2024",
      "type": "Peer-reviewed journal article",
      "reviewLevel": "Selected sections",
      "finding": "In a randomized study of 293 short-story writers, access to GPT-4 ideas improved assessed novelty and usefulness, especially for lower-scoring writers. AI-assisted stories were more similar to other stories within their condition under an embedding-based measure.",
      "limitations": "The task used eight-sentence stories, fixed prompts and no interactive dialogue. It did not study professional writers, scientific collaboration or independent Leviathans. Story similarity is not a direct measure of diversity in scientific explanations; transfer is a research hypothesis. Data and code were not rerun.",
      "changes": "Read the UCL published-version PDF, pages 1–6. Publisher and PMC direct access failed; supplementary analyses, data and code were not reviewed or rerun.",
      "checkedAt": "2026-10-04",
      "verificationStatus": "Published primary text reviewed through UCL repository; no independent reproduction",
      "relatedUrls": [
        "https://discovery.ucl.ac.uk/id/eprint/10195027/",
        "https://discovery.ucl.ac.uk/id/eprint/10195027/1/Generative%20AI%20enhances%20individual%20creativity%20but%20reduces%20the%20collective%20diversity%20of%20novel%20content.pdf",
        "https://pubmed.ncbi.nlm.nih.gov/38996021/",
        "https://pmc.ncbi.nlm.nih.gov/articles/PMC11244532/"
      ]
    },
    {
      "id": "R-AG-10",
      "title": "LongHarness: Advancing Agentic Reasoning over 100M-Token Corpora via Stateful Evidence-Conclusion Coevolution",
      "url": "https://academy.dair.ai/papers/longharness-advancing-agentic-reasoning-over-100m-token-corpora-via-stateful-evi-2610.13075",
      "authors": "Shanghai AI Laboratory, Fudan University, Zhejiang University, and Shanghai Jiao Tong University researchers (as indexed by DAIR.AI)",
      "publishedAt": "2026-10",
      "version": "arXiv:2610.13075; indexed in DAIR.AI Papers of the Week (Oct 12–Oct 18, 2026)",
      "type": "Preprint",
      "reviewLevel": "Indexed technical digest (DAIR.AI Papers of the Week 2026)",
      "finding": "LongHarness separates retrieved evidence and evolving conclusions into explicit persistent state rather than leaving them mixed in interaction history. The authors identify evidence-state drift—where earlier evidence or later contradicting discoveries fail to update current conclusions—and evaluate Evidence-Conclusion Coevolution on corpora up to 100M tokens.",
      "limitations": "This review is based on the curated DAIR.AI technical summary rather than a full-text audit or benchmark reproduction. Stateful evidence tracking for single-agent document synthesis does not by itself validate Leviathan’s multi-participant meaning kernels or governance.",
      "changes": "Motivates separating explicit evidence state and contradiction propagation from raw conversation history when conclusions are revised over time.",
      "checkedAt": "2026-10-21",
      "verificationStatus": "Indexed technical summary reviewed (DAIR.AI 2026); primary full text and experiments not independently reproduced",
      "relatedUrls": [
        "https://github.com/dair-ai/AI-Papers-of-the-Week/blob/main/years/2026.md"
      ]
    },
    {
      "id": "R-AG-11",
      "title": "VeriHarness: Scaling Agentic Verification for Long-Horizon Tasks",
      "url": "https://academy.dair.ai/papers/veriharness-scaling-agentic-verification-for-long-horizon-tasks-2610.00972",
      "authors": "Google researchers (as indexed by DAIR.AI)",
      "publishedAt": "2026-10",
      "version": "arXiv:2610.00972; indexed in DAIR.AI Papers of the Week (Sep 28–Oct 4, 2026)",
      "type": "Preprint",
      "reviewLevel": "Indexed technical digest (DAIR.AI Papers of the Week 2026)",
      "finding": "Across parallel long-horizon agent rollouts, consensus agreement can mask shared systematic errors, whereas disagreement often contains the correct alternative. VeriHarness combines a disagreement resolver with an explicit consensus challenger that searches for failure modes even when candidate outputs agree.",
      "limitations": "Reviewed via the curated DAIR.AI technical summary; primary benchmark runs were not reproduced. Verification gains on bounded coding, math, and web tasks do not establish ethical legitimacy or prove that a challenge mechanism prevents all shared blind spots.",
      "changes": "Supports separating agreement from correctness and motivates explicit consensus-challenging procedures in multi-agent and plural review workflows.",
      "checkedAt": "2026-10-21",
      "verificationStatus": "Indexed technical summary reviewed (DAIR.AI 2026); primary full text and experiments not independently reproduced",
      "relatedUrls": [
        "https://github.com/dair-ai/AI-Papers-of-the-Week/blob/main/years/2026.md"
      ]
    },
    {
      "id": "R-AG-12",
      "title": "ReHarness: Recurrent Harness for Agentic Continual Learning",
      "url": "https://academy.dair.ai/papers/recurrent-harness-for-agentic-continual-learning-2610.12668",
      "authors": "MIT, Meta, and Yale University researchers (as indexed by DAIR.AI)",
      "publishedAt": "2026-10",
      "version": "arXiv:2610.12668; indexed in DAIR.AI Papers of the Week (Oct 12–Oct 18, 2026)",
      "type": "Preprint",
      "reviewLevel": "Indexed technical digest (DAIR.AI Papers of the Week 2026)",
      "finding": "Standard reflection memory is forward-only: once a lesson is written, later tasks rarely test whether that memory actually helped. ReHarness introduces a backward pass over non-parametric memory that attributes downstream task outcomes to retrieved experiences and rewrites stale or misleading entries with source-grounded revisions.",
      "limitations": "Reviewed via the curated DAIR.AI technical summary without full-text audit or code reproduction. Continual-learning benchmark gains do not establish that rewritten memories preserve contested human values or work across independent communities.",
      "changes": "Motivates tracking which stored notes or revisions were used in later tasks and updating or retiring entries when downstream evidence shows they misled.",
      "checkedAt": "2026-10-21",
      "verificationStatus": "Indexed technical summary reviewed (DAIR.AI 2026); primary full text and experiments not independently reproduced",
      "relatedUrls": [
        "https://github.com/dair-ai/AI-Papers-of-the-Week/blob/main/years/2026.md"
      ]
    },
    {
      "id": "R-LRN-07",
      "title": "Scaling Discovery Requires Forgetting",
      "url": "https://academy.dair.ai/papers/scaling-discovery-requires-forgetting-2610.20563",
      "authors": "UC Berkeley, Stanford University, Bespoke Labs, and University of Texas at Austin researchers (as indexed by DAIR.AI)",
      "publishedAt": "2026-10",
      "version": "arXiv:2610.20563; indexed in DAIR.AI Papers of the Week (Oct 19–Oct 25, 2026)",
      "type": "Preprint",
      "reviewLevel": "Indexed technical digest (DAIR.AI Papers of the Week 2026)",
      "finding": "In iterative test-time scientific and algorithmic discovery, accumulating full interaction and attempt history in context causes agents to anchor on minor edits and plateau. Periodically discarding session history and restarting from only the best discovered candidate improves performance across discovery tasks by 15% to 39%.",
      "limitations": "Reviewed via the curated DAIR.AI technical summary; experiments were not rerun. The benefit of clearing working-context history during search does not mean archival provenance or negative results should be deleted from a shared research library.",
      "changes": "Clarifies the distinction between an archival record of failed attempts and an agent’s active working context during search.",
      "checkedAt": "2026-10-21",
      "verificationStatus": "Indexed technical summary reviewed (DAIR.AI 2026); primary full text and experiments not independently reproduced",
      "relatedUrls": [
        "https://github.com/dair-ai/AI-Papers-of-the-Week/blob/main/years/2026.md"
      ]
    }
  ]
}
