{
 "updated": "2026-10-30",
 "count": 129,
 "topics": [
  {
   "key": "foundations",
   "short": "Foundations",
   "sort": "oldest",
   "label": "Foundations",
   "blurb": "The idea, its first designs, and the arguments about how far it can go."
  },
  {
   "key": "weights",
   "short": "Own output",
   "label": "Learning from its own output",
   "blurb": "The model produces or grades data, then trains its weights on it."
  },
  {
   "key": "scaffold",
   "short": "Prompts & workflows",
   "label": "Rewriting its prompts and workflows",
   "blurb": "The model stays fixed; the instructions, memory, tools and pipeline around it are rewritten."
  },
  {
   "key": "code",
   "short": "Own code",
   "label": "Editing its own code",
   "blurb": "An agent program changes its own source, including the part that makes the changes."
  },
  {
   "key": "research",
   "short": "AI research",
   "label": "Automating AI research",
   "blurb": "Systems that propose, run and score experiments, or discover algorithms that make AI better."
  },
  {
   "key": "measurement",
   "short": "Benchmarks",
   "label": "Measuring AI R&D ability",
   "blurb": "Benchmarks that test how well AI agents do the work of AI researchers and engineers."
  },
  {
   "key": "risk",
   "short": "Limits & risks",
   "label": "Limits, risks and forecasts",
   "blurb": "Where self-improvement fails, what could go wrong, and how fast it might happen."
  },
  {
   "key": "community",
   "short": "Workshops & surveys",
   "label": "Workshops and surveys",
   "blurb": "Where the field meets, and overviews of it."
  }
 ],
 "levels": [
  {
   "level": 0,
   "name": "Fixed",
   "test": "Passes none of the tests below.",
   "fixed": "Everything.",
   "limit": "A model with a fixed prompt answering questions. It is where every system starts."
  },
  {
   "level": 1,
   "name": "Retries",
   "test": "Does it revise its own output during a task, using feedback it gathers?",
   "fixed": "Everything except the current answer. Nothing it learns carries over to the next task.",
   "limit": "Revision works when the feedback comes from outside the model: [Reflexion](#reflexion) reached 91% on HumanEval by running its code, reflecting on the failures and retrying. Without outside feedback, models asked to correct their own reasoning often get worse ([Huang et al.](#cannot-self-correct))."
  },
  {
   "level": 2,
   "name": "Learns",
   "test": "Does a change it makes carry over to later tasks, in its prompts, memory, tools, code or weights?",
   "fixed": "The improver. A procedure written by people proposes and accepts every change, and the loop cannot alter it.",
   "limit": "Gains taper off because the improver never gets better. When the Darwin Gödel Machine was run with its improver held fixed, its \"gains taper off quickly\" ([DGM](#dgm)); repeated self-training \"can reach a point of diminishing returns\" ([Multiagent Finetuning](#multiagent-finetuning))."
  },
  {
   "level": 3,
   "name": "Improves its improver",
   "test": "Can the loop change the part that proposes or accepts changes, such as its editing code, its mutation rules, its judge or its task-setter?",
   "fixed": "The model the improver runs on, and usually the outer rule that picks which version to keep.",
   "limit": "Something always stays fixed so the system stays stable. Self-editing systems \"must leave part of their own editing machinery untouched\", which caps how many levels of self-editing they actually reach at about two ([Metaⁿ](#metan)). Gains so far rest on a frozen model: [DGM](#dgm) went from 20% to 50% on SWE-bench with the same Claude 3.5 Sonnet throughout."
  },
  {
   "level": 4,
   "name": "Improves its own model",
   "test": "Besides its improver, does the loop retrain the weights of the model that does the improving, so the next round runs on the improved model?",
   "fixed": "The training recipe, the outside checks, and the model's design.",
   "limit": "Runs are short: [Self-Rewarding](#self-rewarding) reports three rounds, [MASS](#mass) two. Each round costs a full training run, and a model that grades its own training data can drift toward what its grader likes ([Gao et al.](#reward-overoptimization))."
  },
  {
   "level": 5,
   "name": "Builds its successor",
   "test": "Does the system carry out the research that produces a more capable next generation, which then runs the next round, with no person needed and no slowdown from round to round?",
   "fixed": "Only the goal and the outside world it is measured against.",
   "limit": "Not observed. This is the \"intelligence explosion\" loop of [Good (1965)](#good-1965). Whether it would speed up or stall depends on returns: if each doubling of software efficiency needs more than twice the research effort of the last, progress fizzles; if less, it runs away ([Eth and Davidson](#forethought-sie)).",
   "signals": [
    "[AlphaEvolve](#alphaevolve) sped up a kernel in Gemini's architecture by 23%, cutting Gemini's training time by 1%. Gemini is the model AlphaEvolve runs on, and engineers put the change in place.",
    "As of May 2026, more than 80% of the code merged into Anthropic's codebase was written by Claude ([When AI builds itself](#when-ai-builds-itself)).",
    "[AIDE²](#aide2), an AI research agent, found seven successive improvements to its own code in an 8-day run; the gains held on four held-out benchmarks.",
    "Estimates of today's feedback loops say they are not yet self-sustaining but are strengthening ([The Economics of RSI](#economics-of-rsi))."
   ]
  }
 ],
 "map": [
  {
   "id": "check",
   "short": "Outside check",
   "title": "A change counts only if something outside the loop confirms it.",
   "why": "Models asked to correct their own reasoning with no outside feedback often get worse ([Huang et al.](#cannot-self-correct)). Optimizing hard against a learned grader first raises and then lowers real quality ([Gao et al.](#reward-overoptimization)). The Darwin Gödel Machine kept a change only after the new agent solved real GitHub issues, and went from 20% to 50% ([DGM](#dgm)).",
   "build": [
    {
     "name": "Run the changed system on tasks with real tests",
     "text": "Keep a change only if the new version passes tests it did not write. [Absolute Zero](#absolute-zero) checks the tasks it invents by running the code.",
     "examples": [
      "dgm",
      "sica",
      "absolute-zero"
     ]
    },
    {
     "name": "Choose changes on one set of tasks, judge them on another",
     "text": "[AIDE²](#aide2) keeps the changes that do best on hidden evaluations, then reports results on four held-out benchmarks.",
     "examples": [
      "aide2"
     ]
    },
    {
     "name": "Write the tests first, then freeze them",
     "text": "[HarnessPro](#harnesspro) generates tests from public requirements and freezes them before it starts improving harnesses.",
     "examples": [
      "harnesspro"
     ]
    },
    {
     "name": "When there is no test, compare candidates or evolve the grader against a human anchor",
     "text": "[Harness²](#harness2) contrasts candidate edits on the task itself (Propose–Probe–Compose). The [Red Queen Gödel Machine](#red-queen-godel-machine) evolves its graders together with its agents, anchored to human IMO grades.",
     "examples": [
      "harness2",
      "red-queen-godel-machine"
     ]
    }
   ],
   "test": "Report success on tasks with executable tests, such as SWE-bench Verified and Polyglot ([DGM](#dgm)). Keep the final evaluation hidden from the loop. Measure reward hacking directly: in [AIDE²](#aide2) it fell from 55% to 32% during the run.",
   "open": "How do you check work that has no test, such as open-ended research? When frontier agents got six days to answer the questions of two unpublished NeurIPS submissions, the original authors rejected both results ([case studies](#open-ended-ai-research-case-studies))."
  },
  {
   "id": "compound",
   "short": "Improve the improver",
   "title": "To keep improving, the loop has to improve its improver.",
   "why": "With the part that makes changes held fixed, the Darwin Gödel Machine's \"gains taper off quickly\" ([DGM](#dgm)). Repeated training on a model's own output \"can reach a point of diminishing returns\" ([Multiagent Finetuning](#multiagent-finetuning)). Gains compound when the skill being improved is the skill that does the improving: for a coding agent, a better coder is also a better self-editor.",
   "build": [
    {
     "name": "Let the agent edit its own code, including the part that edits",
     "text": "The agent's source, its tools and its self-editing logic are all open to change.",
     "examples": [
      "dgm",
      "sica",
      "godel-agent",
      "stop"
     ]
    },
    {
     "name": "Make the procedure that makes changes editable too",
     "text": "[Hyperagents](#hyperagents) puts the task agent and the agent that edits it into one editable program; improvements to the editing procedure carried over to other domains.",
     "examples": [
      "hyperagents",
      "metaskill-evolve",
      "meta-context-engineering"
     ]
    },
    {
     "name": "Train the judge or the task-setter, not only the solver",
     "text": "The model that grades answers or invents practice tasks is itself trained by the loop.",
     "examples": [
      "self-rewarding",
      "meta-rewarding",
      "absolute-zero",
      "r-zero"
     ]
    },
    {
     "name": "Learn how to improve from earlier improvements",
     "text": "[Harness²](#harness2) keeps lessons from each task's harness edits and uses them to update the harness that does the editing.",
     "examples": [
      "harness2"
     ]
    }
   ],
   "test": "Run the same loop with the improver frozen and compare ([DGM](#dgm) does this). Then test whether the discovered system is a better self-improver, not just a better solver: for [AIDE²](#aide2), its improvement in that outer role was not decisively established.",
   "open": "How many levels of self-editing are possible? Self-editing systems \"must leave part of their own editing machinery untouched to stay stable\", which caps the depth they reach at about two ([Metaⁿ](#metan))."
  },
  {
   "id": "keep",
   "short": "Stepping stones",
   "title": "Keep the stepping stones.",
   "why": "When the Darwin Gödel Machine kept only its latest agent, one bad edit made every later improvement harder. Its best agent descends from ancestors that twice scored worse than their parents ([DGM](#dgm)).",
   "build": [
    {
     "name": "Keep every version in an archive, and sometimes build on weak ones",
     "text": "Parents are chosen mostly by score, but every agent keeps some chance, and agents with few children are favoured.",
     "examples": [
      "dgm"
     ]
    },
    {
     "name": "Judge a version by how its descendants do",
     "text": "An agent's own score is a poor guide to how good its descendants will be, so the [Huxley-Gödel Machine](#hgm) chooses by its descendants' scores.",
     "examples": [
      "hgm"
     ]
    },
    {
     "name": "Share experience across lineages",
     "text": "[Group-Evolving Agents](#group-evolving-agents) evolve a group of agents as one unit and reached 71.0% on SWE-bench Verified, against 56.7% for tree-shaped self-evolution.",
     "examples": [
      "group-evolving-agents",
      "mendel-godel-machine"
     ]
    },
    {
     "name": "Keep lessons as records that grow and get edited",
     "text": "Instead of rewriting the context from scratch, keep and edit what was learned.",
     "examples": [
      "ace",
      "memevolve",
      "harness2"
     ]
    }
   ],
   "test": "Remove the archive and keep only the latest version, as the [DGM](#dgm) ablation does; compare tree-shaped and group-shaped search on the same budget.",
   "open": "What does a large archive cost to keep and search, and when does that cost outweigh what it finds?"
  },
  {
   "id": "transfer",
   "short": "Unseen tasks",
   "title": "The gain must hold on tasks the loop never saw.",
   "why": "A loop gets better at what it measures, which may not be what was meant. [AIDE²](#aide2) chose its changes on one set of AI research tasks, then matched or beat a human-built research agent on four held-out benchmarks, one of them weather forecasting.",
   "build": [
    {
     "name": "Keep held-out benchmarks for the final judgment",
     "text": "Select on one set of tasks; report on another the loop never touched.",
     "examples": [
      "aide2"
     ]
    },
    {
     "name": "Reject edits that only fit the benchmark",
     "text": "[RRSI](#rrsi) adds a critic that screens out benchmark-specific edits, and limits how many edits one candidate may bundle, with the limit tightening over time.",
     "examples": [
      "rrsi"
     ]
    },
    {
     "name": "Test with other models",
     "text": "The agent the DGM found with Claude 3.5 Sonnet also helped other models; with Claude 3.7 Sonnet it went from 19.0% to 59.5%.",
     "examples": [
      "dgm"
     ]
    }
   ],
   "test": "Held-out benchmarks, a different model, and at least one task unlike the selection tasks.",
   "open": "Leakage is easy: the [Harness²](#harness2) release reports a validation run that downloaded a public JobBench rubric, and it leaves network access unrestricted by default."
  },
  {
   "id": "returns",
   "short": "Returns and cost",
   "title": "Recursive does not mean explosive. Returns and cost decide the speed.",
   "why": "A loop speeds up only if each improvement makes the next one cheaper. If each doubling of software efficiency needs more than twice the research effort of the previous one, progress fizzles; if less, it runs away ([Eth and Davidson](#forethought-sie)). Current estimates say today's loops are not yet self-sustaining but are strengthening ([The Economics of RSI](#economics-of-rsi)).",
   "build": [
    {
     "name": "Count search, execution and human effort separately",
     "text": "[HarnessPro](#harnesspro) treats cost as an objective and returns lite, medium and heavy harnesses; its heavy mode gained 2.0 points at 26.1% lower cost than RHO.",
     "examples": [
      "harnesspro",
      "sol-pi"
     ]
    },
    {
     "name": "Shorten the trip around the loop",
     "text": "The time one round takes is a neglected deciding factor ([Ord](#dynamics-of-intelligence-explosions)).",
     "examples": [
      "dynamics-of-intelligence-explosions"
     ]
    }
   ],
   "test": "Report the cost of every accepted improvement, round by round, and whether later improvements cost less than earlier ones.",
   "open": "Do later improvements cost less than earlier ones in any real system? Measuring it requires per-round cost records."
  },
  {
   "id": "safety",
   "short": "Safety across updates",
   "title": "Safety has to survive every update.",
   "why": "A system that rewrites itself can rewrite away its safeguards. Agents that keep updating their memory and tools drift toward unsafe behaviour: alignment erodes as memory accumulates ([Shao et al.](#misevolve)). It can also go the other way: [AIDE²](#aide2)'s rate of reward hacking fell from 55% to 32% though nothing optimized for it.",
   "build": [
    {
     "name": "Run changes in a sandbox, with people watching",
     "text": "The [DGM](#dgm) ran all experiments with sandboxing and human oversight.",
     "examples": [
      "dgm"
     ]
    },
    {
     "name": "Compare each self-update with a copy that did not update",
     "text": "[SEABench](#seabench) pairs evolving agents with non-evolving ones so unsafe behaviour can be traced to a specific update.",
     "examples": [
      "seabench",
      "evolutionary-safety"
     ]
    }
   ],
   "test": "Measure the same safety tests before and after every update, against a copy that did not update.",
   "open": "A goal-seeking system will want to keep its goals ([Omohundro](#omohundro-2008)), but proving that a self-modified successor is still safe is hard ([Yudkowsky and Herreshoff](#tiling-agents-2013))."
  },
  {
   "id": "data",
   "short": "Fix the data",
   "source": "editor",
   "title": "Fix the data, let the execution change.",
   "why": "The editor's working premise (rsi-design study): define a task by fixed data contracts, meaning the required records, the checkpoints where progress is inspected, and how each is judged. Then let models, prompts, skills and paths change freely to satisfy them. In the editor's words: \"the system must be designed around data, because only the data is fixed in the process and output for any given tasks (forms and ratings, not content)\".",
   "build": [
    {
     "name": "Make checkpoints inspectable records, not fixed steps",
     "text": "A checkpoint says which state must become visible, not which actions produce it.",
     "examples": []
    },
    {
     "name": "Change the representation or the execution, never both at once",
     "text": "Hold one side fixed so a gain can be attributed to the other.",
     "examples": []
    },
    {
     "name": "Rules are append-only",
     "text": "No rule is updated or removed; new rules are added. An output cannot pass by deleting an inconvenient requirement.",
     "examples": []
    },
    {
     "name": "The person chooses the task and owns the final judgment",
     "text": "The system pauses only for judgments that the task's owner has to make.",
     "examples": []
    },
    {
     "name": "Fill missing data first, then improve what limits quality",
     "text": "Produce required missing records before polishing ones that exist but are rough.",
     "examples": []
    }
   ],
   "test": "Compare a representation change and an execution change on the same fixed inputs, each with the other held fixed, under the same rules and the same judges.",
   "open": "The evidence is thin so far. [AFlow](#aflow) searches explicit agent structures and [ACE](#ace) keeps incremental knowledge records, but neither isolates changes to the task's data structure as the main source of improvement. No reviewed experiment directly tests the missing-first rule."
  }
 ],
 "entries": [
  {
   "id": "meta-agents-workshop-2026",
   "title": "Managing Agents that Manage Agents: Workshop on Responsible Use of Meta-Agents",
   "authors": "NeurIPS 2026 workshop",
   "year": 2026,
   "month": 12,
   "topic": "community",
   "listed": "2026-10-30",
   "kind": "event",
   "url": "https://meta-agents-workshop.github.io/",
   "note": "Workshop on agents that design, train, supervise or improve other agents, 11 December 2026, Sydney.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "arrsi-2026",
   "title": "Workshop on Autonomous Research and Recursive Self-Improvement (ARRSI)",
   "authors": "AACL-IJCNLP 2026 workshop",
   "year": 2026,
   "month": 11,
   "topic": "community",
   "listed": "2026-10-30",
   "kind": "event",
   "url": "https://ar-rsi.github.io/",
   "note": "First workshop on autonomous research and recursive self-improvement, 10 November 2026, Hengqin, China.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "harness2",
   "title": "Harness²: Recursive Agent Harnessing for an Open World",
   "authors": "Ruiyao Xu et al.",
   "year": 2026,
   "month": 10,
   "topic": "scaffold",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 3,
   "recursive": true,
   "url": "https://arxiv.org/abs/2610.14224",
   "code": "https://github.com/google-research/harness2",
   "note": "Edits the agent's harness for each task, then learns from those edits to update the harness that does the editing: up to +11.5 points on WorkBuddy-Bench.",
   "step": "analyzed",
   "record": "https://rsi.org.ai/records/harness2/",
   "map": [
    "check",
    "compound",
    "keep",
    "transfer"
   ]
  },
  {
   "id": "harnesspro",
   "title": "HarnessPro: Recursive Harness Self-Improvement with Pareto Optimization for Performance-Cost Trade-offs",
   "authors": "Zhuoshi Pan et al.",
   "year": 2026,
   "month": 10,
   "topic": "scaffold",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 2,
   "url": "https://arxiv.org/abs/2610.26770",
   "code": "https://github.com/pzs19/HarnessPro",
   "note": "Agents diagnose runs and rewrite the harness, scored by model-written tests; heavy mode: +2.0 points pass rate at 26.1% lower cost than RHO.",
   "step": "analyzed",
   "record": "https://rsi.org.ai/records/harnesspro/",
   "map": [
    "check",
    "returns"
   ]
  },
  {
   "id": "mass",
   "name": "MASS",
   "title": "Recursive Self-Improvement through Multi-Agent Self-Supervision",
   "authors": "Hyunin Lee et al.",
   "year": 2026,
   "month": 10,
   "topic": "weights",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 4,
   "recursive": true,
   "url": "https://arxiv.org/abs/2610.12176",
   "note": "The model designs, runs and grades multi-agent workflows, then is fine-tuned on the results, so the next round's designer and grader is the improved model.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "terminal-rsi",
   "title": "Terminal-RSI: Recursive Self-Improving Terminal Agents via Recursive Self-Rewrite and Policy-Refine",
   "authors": "Junyao Yang et al.",
   "year": 2026,
   "month": 10,
   "topic": "weights",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 2,
   "url": "https://arxiv.org/abs/2610.22192",
   "code": "https://github.com/jyyang26/Terminal-RSI",
   "note": "Alternates training on the model's own verified terminal sessions with RL refinement, twice: strict task completion 0/46 → 5/46 on LHTB.",
   "step": "analyzed",
   "record": "https://rsi.org.ai/records/terminal-rsi/",
   "map": []
  },
  {
   "id": "evolutionary-safety",
   "title": "Evolutionary Safety of Recursive Self-Improving AI: Taxonomy, Risk Discovery, and Evaluation",
   "authors": "Chang Gong et al.",
   "year": 2026,
   "month": 9,
   "topic": "risk",
   "listed": "2026-10-30",
   "kind": "paper",
   "url": "https://arxiv.org/abs/2609.31186",
   "note": "Asks whether safety properties persist or degrade across self-updates; gives a taxonomy of risks and ways to evaluate them.",
   "step": "listed",
   "record": null,
   "map": [
    "safety"
   ]
  },
  {
   "id": "generalized-agent-iteration",
   "title": "Generalized Agent Iteration: One Formal Framework for Iterative Policy Improvement and Recursive Self-Improvement",
   "authors": "Hongyao Tang et al.",
   "year": 2026,
   "month": 9,
   "topic": "foundations",
   "listed": "2026-10-30",
   "kind": "paper",
   "url": "https://arxiv.org/abs/2609.13406",
   "note": "One formal framework for \"evaluate, then improve\" loops; ordinary policy improvement and RSI differ only in whether the improving mechanism is part of the agent.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "metarsi",
   "name": "MetaRSI",
   "title": "MetaRSI / RSI2: A Meta-Recursive Self-Improving System for Recursive Self-Improving Systems Themselves",
   "authors": "Zihan Tan et al.",
   "year": 2026,
   "month": 9,
   "topic": "weights",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 4,
   "recursive": true,
   "url": "https://arxiv.org/abs/2609.06396",
   "note": "Combines three kinds of change, to training data, to a five-slot harness and to the weights; a meta-level policy chooses their order and revises that schedule.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "r2flow",
   "title": "R² Flow: Recursive Self-Improvement via Recursive Skill Evolution",
   "authors": "Mingda Zhang et al.",
   "year": 2026,
   "month": 9,
   "topic": "scaffold",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 3,
   "recursive": true,
   "url": "https://arxiv.org/abs/2609.33867",
   "code": "https://github.com/beita6969/r2flow",
   "note": "Alternates training the policy that chains skills, checking proposed edits with an independent verifier, and committing versioned skill-library changes.",
   "step": "analyzed",
   "record": "https://rsi.org.ai/records/r2flow/",
   "map": []
  },
  {
   "id": "aide2",
   "name": "AIDE²",
   "title": "Recursive self-improvement of AI research agents",
   "authors": "Dhruv Srikanth et al.",
   "year": 2026,
   "month": 9,
   "topic": "code",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 3,
   "recursive": true,
   "url": "https://arxiv.org/abs/2609.26457",
   "note": "An AI research agent tests modified versions of its own code on AI R&D tasks and keeps the best: seven successive improvements in an 8-day run.",
   "step": "analyzed",
   "record": "https://rsi.org.ai/records/aide2/",
   "map": [
    "check",
    "compound",
    "transfer",
    "safety"
   ]
  },
  {
   "id": "rrsi",
   "title": "RRSI: Regularized Recursive Self-Improvement of Agent Harnesses",
   "authors": "Peng Xia et al.",
   "year": 2026,
   "month": 9,
   "topic": "scaffold",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 2,
   "url": "https://arxiv.org/abs/2609.24972",
   "code": "https://github.com/google-research/rrsi",
   "note": "Harness self-improvement with a shrinking edit budget and a critic that rejects benchmark-specific edits, aimed at gains that hold on unseen benchmarks.",
   "step": "listed",
   "record": null,
   "map": [
    "transfer"
   ]
  },
  {
   "id": "seabench",
   "title": "SEABench: Benchmarking Endogenous Misalignment In Self-Evolving Agents",
   "authors": "Saswat Das et al.",
   "year": 2026,
   "month": 9,
   "topic": "risk",
   "listed": "2026-10-30",
   "kind": "benchmark",
   "url": "https://arxiv.org/abs/2609.35596",
   "note": "48 long task sequences compare self-evolving agents with matched non-evolving ones, so unsafe behavior can be traced to earlier self-updates.",
   "step": "listed",
   "record": null,
   "map": [
    "safety"
   ]
  },
  {
   "id": "sol-pi",
   "title": "SoL-Pi: Recursively Scaling Auto-Research Loops for Efficient Agent Harness",
   "authors": "Haozhe Liu et al.",
   "year": 2026,
   "month": 9,
   "topic": "scaffold",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 2,
   "url": "https://arxiv.org/abs/2609.20519",
   "code": "https://github.com/NVlabs/SoL-Pi",
   "note": "Runs harness-improvement loops across many environments and keeps the four changes that survive: same EdgeBench results with 44.7–49.0% less token traffic.",
   "step": "listed",
   "record": null,
   "map": [
    "returns"
   ]
  },
  {
   "id": "economics-of-rsi",
   "title": "The Economics of Recursive Self-Improvement",
   "authors": "Tom Cunningham et al.",
   "year": 2026,
   "month": 9,
   "topic": "risk",
   "listed": "2026-10-30",
   "kind": "paper",
   "url": "https://arxiv.org/abs/2609.15802",
   "note": "Models AI progress as feedback loops whose speed-up depends on the product of elasticities; calibration says the loops are not yet self-sustaining but strengthening.",
   "step": "listed",
   "record": null,
   "map": [
    "returns"
   ]
  },
  {
   "id": "ai-rnd-intelligence-explosion",
   "title": "What if automating AI R&D triggers an intelligence explosion?",
   "authors": "Alan Chan et al.",
   "year": 2026,
   "month": 9,
   "topic": "risk",
   "listed": "2026-10-30",
   "kind": "paper",
   "url": "https://arxiv.org/abs/2609.36054",
   "note": "Weighs the evidence that automating AI R&D could compress years of progress into months, the loss-of-control risks, and policies for visibility and steering.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "mendel-godel-machine",
   "title": "Mendel Gödel Machine: Recursive Self-Improving Coding Agents via Comparative Evolution",
   "authors": "Changzhi Liu et al.",
   "year": 2026,
   "month": 8,
   "topic": "code",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 3,
   "recursive": true,
   "url": "https://arxiv.org/abs/2608.07645",
   "code": "https://github.com/RealLcz/MGM",
   "note": "A self-rewriting coding agent that derives edits from comparisons across several tasks and across lineages in its archive, not from one failure at a time.",
   "step": "listed",
   "record": null,
   "map": [
    "keep"
   ]
  },
  {
   "id": "metan",
   "name": "Metaⁿ",
   "title": "Metaⁿ: Recursive Self-Improvement through Emergent Depth",
   "authors": "Zae Myung Kim et al.",
   "year": 2026,
   "month": 8,
   "topic": "scaffold",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 2,
   "url": "https://arxiv.org/abs/2608.24735",
   "code": "https://github.com/minnesotanlp/meta-n",
   "note": "A fixed operation reads the solver stack's code and traces and writes a new layer on top, again and again; depth stops when gains converge.",
   "step": "listed",
   "record": null,
   "map": [
    "compound"
   ]
  },
  {
   "id": "dynamics-of-intelligence-explosions",
   "title": "The Dynamics of Intelligence Explosions",
   "authors": "Toby Ord",
   "year": 2026,
   "month": 8,
   "topic": "risk",
   "listed": "2026-10-30",
   "kind": "paper",
   "url": "https://arxiv.org/abs/2608.14426",
   "note": "Shows that runaway growth toward a vertical asymptote is harder to reach than recent economic models suggest, and that the time to go once around the loop is a neglected key parameter.",
   "step": "listed",
   "record": null,
   "map": [
    "returns"
   ]
  },
  {
   "id": "open-ended-ai-research-case-studies",
   "title": "Can AI agents conduct open-ended AI research? Early evidence from two case studies",
   "authors": "Peter Kirgis et al.",
   "year": 2026,
   "month": 7,
   "topic": "measurement",
   "listed": "2026-10-30",
   "kind": "paper",
   "url": "https://arxiv.org/abs/2607.27191",
   "note": "Frontier agents got six days and large compute to answer the questions of two unpublished NeurIPS submissions; the original authors graded and rejected both results.",
   "step": "listed",
   "record": null,
   "map": [
    "check"
   ]
  },
  {
   "id": "frontis-ma1",
   "title": "Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering",
   "authors": "Junlin Yang et al.",
   "year": 2026,
   "month": 7,
   "topic": "research",
   "listed": "2026-10-30",
   "kind": "paper",
   "url": "https://arxiv.org/abs/2607.28568",
   "code": "https://github.com/FrontisAI/OpenRSI",
   "note": "Trains a 35B model on four program-evolution operators for ML engineering, then uses it in long search: MLE-Bench Lite medal rate 39.4% → 60.6%.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "metaskill-evolve",
   "title": "MetaSkill-Evolve: Recursive Self-Improvement of LLM Agents via Two-Timescale Meta-Skill Evolution",
   "authors": "Zefeng Wang et al.",
   "year": 2026,
   "month": 7,
   "topic": "scaffold",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 3,
   "recursive": true,
   "url": "https://arxiv.org/abs/2607.05297",
   "note": "A fast loop evolves task skill files; a slow loop evolves the meta-skill that configures how those skill files are improved.",
   "step": "listed",
   "record": null,
   "map": [
    "compound"
   ]
  },
  {
   "id": "rsi-survey-2026",
   "title": "Recursive Self-Improvement in AI: From Bounded Self-Refinement to Autonomous Research Loops",
   "authors": "Mingguang Chen, Licheng Wang and Bo Qu",
   "year": 2026,
   "month": 7,
   "topic": "community",
   "listed": "2026-10-30",
   "kind": "paper",
   "url": "https://arxiv.org/abs/2607.07663",
   "note": "Survey of RSI work from 2024–2026, sorted by what the system improves and how closed the loop is, from bounded self-refinement to open-ended research loops.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "rsibench-data",
   "title": "RSIBench-Data: Benchmarking Data-Centric Research for Recursive Self-Improvement",
   "authors": "Fanqing Meng et al.",
   "year": 2026,
   "month": 7,
   "topic": "measurement",
   "listed": "2026-10-30",
   "kind": "benchmark",
   "url": "https://arxiv.org/abs/2607.25886",
   "code": "https://github.com/evolvent-ai/RSIBench-Data",
   "note": "Agents repeatedly revise training-data strategies for a fixed model and budget; they beat their own first valid attempt in 58% of settings.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "red-queen-godel-machine",
   "title": "The Red Queen Gödel Machine: Co-Evolving Agents and Their Evaluators",
   "authors": "Alex Iacob et al.",
   "year": 2026,
   "month": 6,
   "topic": "code",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 3,
   "recursive": true,
   "url": "https://arxiv.org/abs/2606.26294",
   "note": "Evolves agents together with the evaluators that grade them, so the scoring criterion improves too: 82.1% vs 75.0% on held-out DeepSWE tasks.",
   "step": "listed",
   "record": null,
   "map": [
    "check"
   ]
  },
  {
   "id": "mls-bench",
   "title": "MLS-Bench: A Holistic and Rigorous Assessment of AI Systems on Building Better AI",
   "authors": "Bohan Lyu et al.",
   "year": 2026,
   "month": 5,
   "topic": "measurement",
   "listed": "2026-10-30",
   "kind": "benchmark",
   "url": "https://arxiv.org/abs/2605.08678",
   "code": "https://github.com/Imbernoulli/MLS-Bench",
   "note": "140 tasks across 12 ML areas: improve one component of a method and show the gain holds across controlled settings and scales.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "sia",
   "title": "SIA: Self Improving AI with Harness & Weight Updates",
   "authors": "Prannay Hebbar et al.",
   "year": 2026,
   "month": 5,
   "topic": "weights",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 2,
   "url": "https://arxiv.org/abs/2605.27276",
   "code": "https://github.com/hexo-ai/sia",
   "note": "A feedback agent updates both the harness and the weights of a task agent; combining both beat harness edits alone in all three test domains.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "ahe",
   "name": "Agentic Harness Engineering",
   "title": "Agentic Harness Engineering: Observability-Driven Automatic Evolution of Coding-Agent Harnesses",
   "authors": "Jiahang Lin et al.",
   "year": 2026,
   "month": 4,
   "topic": "scaffold",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 2,
   "url": "https://arxiv.org/abs/2604.25850",
   "code": "https://github.com/china-qijizhifeng/agentic-harness-engineering",
   "note": "Each edit to a coding agent's harness comes with a predicted effect that is checked next round: Terminal-Bench 2 pass@1 69.7% → 77.0% in ten rounds.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "iclr-2026-rsi-workshop",
   "title": "Workshop on AI with Recursive Self-Improvement (RSI 2026)",
   "authors": "ICLR 2026 workshop",
   "year": 2026,
   "month": 4,
   "topic": "community",
   "listed": "2026-10-30",
   "kind": "event",
   "url": "https://recursive-workshop.github.io/",
   "note": "A workshop dedicated to recursive self-improvement, held 26 April 2026 in Rio de Janeiro; the site lists its accepted papers.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "autoresearch",
   "title": "autoresearch",
   "authors": "Andrej Karpathy",
   "year": 2026,
   "month": 3,
   "topic": "research",
   "listed": "2026-10-30",
   "kind": "project",
   "url": "https://github.com/karpathy/autoresearch",
   "code": "https://github.com/karpathy/autoresearch",
   "note": "An agent edits a small GPT training script, trains for five minutes, and keeps the change only if validation loss improves; then repeats.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "bilevel-autoresearch",
   "title": "Bilevel Autoresearch: Meta-Autoresearching Itself",
   "authors": "Yaonan Qu and Meng Lu",
   "year": 2026,
   "month": 3,
   "topic": "research",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 3,
   "recursive": true,
   "url": "https://arxiv.org/abs/2603.23420",
   "code": "https://github.com/EdwardOptimization/Bilevel-Autoresearch",
   "note": "An outer loop reads an autoresearch loop's code and traces and adds new search mechanisms to it; the same model works at both levels.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "rl-algorithm-discovery",
   "title": "From AI Assistant to AI Scientist: Autonomous Discovery of LLM-RL Algorithms with LLM Agents",
   "authors": "Sirui Xia et al.",
   "year": 2026,
   "month": 3,
   "topic": "research",
   "listed": "2026-10-30",
   "kind": "paper",
   "url": "https://arxiv.org/abs/2603.23951",
   "note": "An agent loop proposes, implements and tests new RL algorithms for LLMs starting from GRPO; the best of 64 raised AIME25 pass@32 from 26.7% to 43.3%.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "hyperagents",
   "title": "Hyperagents",
   "authors": "Jenny Zhang et al.",
   "year": 2026,
   "month": 3,
   "topic": "code",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 3,
   "recursive": true,
   "url": "https://arxiv.org/abs/2603.19461",
   "code": "https://github.com/facebookresearch/HyperAgents",
   "note": "Puts the task agent and the agent that edits it into one editable program, so the editing procedure improves too; extends DGM beyond coding.",
   "step": "analyzed",
   "record": "https://rsi.org.ai/records/hyperagents/",
   "map": [
    "compound"
   ]
  },
  {
   "id": "measuring-ai-rnd-automation",
   "title": "Measuring AI R&D Automation",
   "authors": "Alan Chan et al.",
   "year": 2026,
   "month": 3,
   "topic": "measurement",
   "listed": "2026-10-30",
   "kind": "paper",
   "url": "https://arxiv.org/abs/2603.03992",
   "note": "Proposes indicators companies and outsiders could report to track AI R&D automation, such as the AI share of R&D spending, how researchers spend their time, and incidents of AI subversion.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "posttrainbench",
   "title": "PostTrainBench: Can LLM Agents Automate LLM Post-Training?",
   "authors": "Ben Rank et al.",
   "year": 2026,
   "month": 3,
   "topic": "measurement",
   "listed": "2026-10-30",
   "kind": "benchmark",
   "url": "https://arxiv.org/abs/2603.08640",
   "code": "https://github.com/aisa-group/PostTrainBench",
   "note": "An agent gets a base model, one H100 and 10 hours to post-train it; the best agent reached 23.2% versus 51.1% for official instruction-tuned models.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "airs-bench",
   "title": "AIRS-Bench: a Suite of Tasks for Frontier AI Research Science Agents",
   "authors": "Alisia Lupidi et al.",
   "year": 2026,
   "month": 2,
   "topic": "measurement",
   "listed": "2026-10-30",
   "kind": "benchmark",
   "url": "https://arxiv.org/abs/2602.06855",
   "code": "https://github.com/facebookresearch/airs-bench",
   "note": "20 tasks taken from recent ML papers across language modeling, math, bioinformatics and time series, scored against published human results.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "group-evolving-agents",
   "name": "Group-Evolving Agents",
   "title": "Group-Evolving Agents: Open-Ended Self-Improvement via Experience Sharing",
   "authors": "Zhaotian Weng et al.",
   "year": 2026,
   "month": 2,
   "topic": "code",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 3,
   "recursive": true,
   "url": "https://arxiv.org/abs/2602.04837",
   "code": "https://github.com/UCSB-AI/GEA",
   "note": "A group of self-modifying coding agents evolves as one unit and shares experience: 71.0% vs 56.7% on SWE-bench Verified over tree-style self-evolution.",
   "step": "listed",
   "record": null,
   "map": [
    "keep"
   ]
  },
  {
   "id": "meta-context-engineering",
   "name": "Meta Context Engineering",
   "title": "Meta Context Engineering via Agentic Skill Evolution",
   "authors": "Haoran Ye et al.",
   "year": 2026,
   "month": 1,
   "topic": "scaffold",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 3,
   "recursive": true,
   "url": "https://arxiv.org/abs/2601.21557",
   "code": "https://github.com/metaevo-ai/meta-context-engineering",
   "note": "A meta-level agent evolves the skills that a base agent uses to build its own context, instead of keeping those context-building rules fixed.",
   "step": "listed",
   "record": null,
   "map": [
    "compound"
   ]
  },
  {
   "id": "execution-grounded-research",
   "title": "Towards Execution-Grounded Automated AI Research",
   "authors": "Chenglei Si et al.",
   "year": 2026,
   "month": 1,
   "topic": "research",
   "listed": "2026-10-30",
   "kind": "paper",
   "url": "https://arxiv.org/abs/2601.14525",
   "note": "Runs model-generated research ideas as real GPU experiments on LLM training; execution-guided search found a post-training method scoring 69.4% vs GRPO's 48.0%.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "automated-alignment-researchers",
   "title": "Automated Researchers Can Mitigate Well-Characterized Alignment Failures",
   "authors": "Chen Yueh-Han, Jiaxin Wen and Jan Hendrik Kirchner",
   "year": 2026,
   "topic": "research",
   "listed": "2026-10-30",
   "kind": "paper",
   "url": "https://alignment.anthropic.com/2026/automated-alignment-researchers/",
   "note": "Claude agents invent post-training fixes for ten alignment failures such as sycophancy and deception, with held-out benchmarks deciding whether a fix works.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "automated-w2s-researcher",
   "title": "Automated Weak-to-Strong Researcher",
   "authors": "Jiaxin Wen et al.",
   "year": 2026,
   "topic": "research",
   "listed": "2026-10-30",
   "kind": "paper",
   "url": "https://alignment.anthropic.com/2026/automated-w2s-researcher/",
   "note": "Parallel Claude agents ran weak-to-strong supervision research: performance gap recovered 0.23 (best human-tuned) → 0.97 in 5 days, for about $18,000.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "when-ai-builds-itself",
   "title": "When AI builds itself",
   "authors": "Marina Favaro and Jack Clark (Anthropic)",
   "year": 2026,
   "topic": "risk",
   "listed": "2026-10-30",
   "kind": "essay",
   "url": "https://www.anthropic.com/institute/recursive-self-improvement",
   "note": "As of May 2026, more than 80% of the code merged into Anthropic's codebase was written by Claude; lays out scenarios and coordination options for AI building its successors.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "memevolve",
   "name": "MemEvolve",
   "title": "MemEvolve: Meta-Evolution of Agent Memory Systems",
   "authors": "Guibin Zhang et al.",
   "year": 2025,
   "month": 12,
   "topic": "scaffold",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 3,
   "recursive": true,
   "url": "https://arxiv.org/abs/2512.18746",
   "code": "https://github.com/bingreeky/MemEvolve",
   "note": "Evolves both what the agent remembers and the memory system itself (how experience is encoded, stored and retrieved).",
   "step": "listed",
   "record": null,
   "map": [
    "keep"
   ]
  },
  {
   "id": "self-play-swe-rl",
   "name": "Self-Play SWE-RL",
   "title": "Toward Training Superintelligent Software Agents through Self-Play SWE-RL",
   "authors": "Yuxiang Wei et al.",
   "year": 2025,
   "month": 12,
   "topic": "weights",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 4,
   "recursive": true,
   "url": "https://arxiv.org/abs/2512.18552",
   "note": "One agent learns by RL both to plant bugs in real repositories (each defined by a test) and to fix them: +10.4 points on SWE-bench Verified with no human issues.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "agent0",
   "title": "Agent0: Unleashing Self-Evolving Agents from Zero Data via Tool-Integrated Reasoning",
   "authors": "Peng Xia et al.",
   "year": 2025,
   "month": 11,
   "topic": "weights",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 4,
   "recursive": true,
   "url": "https://arxiv.org/abs/2511.16043",
   "code": "https://github.com/aiming-lab/Agent0",
   "note": "A curriculum agent proposes harder tool-using tasks while an executor agent learns to solve them; both start from the same model and co-evolve without outside data.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "live-swe-agent",
   "title": "Live-SWE-agent: Can Software Engineering Agents Self-Evolve on the Fly?",
   "authors": "Chunqiu Steven Xia et al.",
   "year": 2025,
   "month": 11,
   "topic": "code",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 1,
   "url": "https://arxiv.org/abs/2511.13646",
   "code": "https://github.com/OpenAutoCoder/live-swe-agent",
   "note": "Starts from a bash-only scaffold and writes new tools into its own scaffold while solving each issue, with no offline training.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "ace",
   "title": "Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models",
   "authors": "Qizheng Zhang et al.",
   "year": 2025,
   "month": 10,
   "topic": "scaffold",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 2,
   "url": "https://arxiv.org/abs/2510.04618",
   "code": "https://github.com/ace-agent/ace",
   "note": "Treats the context as a playbook that the model grows and edits from its own results without collapsing it: +10.6% on agent benchmarks.",
   "step": "analyzed",
   "record": "https://rsi.org.ai/records/ace/",
   "map": [
    "keep",
    "data"
   ]
  },
  {
   "id": "hgm",
   "title": "Huxley-Gödel Machine: Human-Level Coding Agent Development by an Approximation of the Optimal Self-Improving Machine",
   "authors": "Wenyi Wang et al.",
   "year": 2025,
   "month": 10,
   "topic": "code",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 3,
   "recursive": true,
   "url": "https://arxiv.org/abs/2510.21614",
   "code": "https://github.com/metauto-ai/HGM",
   "note": "Chooses which self-modification to expand by how well its descendants score, not its own score; reached human-level results on SWE-bench Lite.",
   "step": "listed",
   "record": null,
   "map": [
    "keep"
   ]
  },
  {
   "id": "shinkaevolve",
   "title": "ShinkaEvolve: Towards Open-Ended And Sample-Efficient Program Evolution",
   "authors": "Robert Tjarko Lange, Yuki Imajuku and Edoardo Cetin",
   "year": 2025,
   "month": 9,
   "topic": "research",
   "listed": "2026-10-30",
   "kind": "paper",
   "url": "https://arxiv.org/abs/2509.19349",
   "code": "https://github.com/SakanaAI/ShinkaEvolve",
   "note": "Open-source program evolution tuned to need few samples: a new best circle-packing solution from only 150 samples.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "misevolve",
   "title": "Your Agent May Misevolve: Emergent Risks in Self-evolving LLM Agents",
   "authors": "Shuai Shao et al.",
   "year": 2025,
   "month": 9,
   "topic": "risk",
   "listed": "2026-10-30",
   "kind": "paper",
   "url": "https://arxiv.org/abs/2509.26354",
   "code": "https://github.com/ShaoShuai0605/Misevolution",
   "note": "Self-evolving agents can drift unsafe as they update memory, tools or workflows, for example alignment eroding as memory accumulates.",
   "step": "listed",
   "record": null,
   "map": [
    "safety"
   ]
  },
  {
   "id": "r-zero",
   "title": "R-Zero: Self-Evolving Reasoning LLM from Zero Data",
   "authors": "Chengsong Huang et al.",
   "year": 2025,
   "month": 8,
   "topic": "weights",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 4,
   "recursive": true,
   "url": "https://arxiv.org/abs/2508.05004",
   "code": "https://github.com/Chengsong-Huang/R-Zero",
   "note": "Two copies of one model: a Challenger rewarded for questions at the edge of the Solver's ability, and a Solver trained on them. No human tasks or labels.",
   "step": "listed",
   "record": null,
   "map": [
    "compound"
   ]
  },
  {
   "id": "gepa",
   "title": "GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning",
   "authors": "Lakshya A Agrawal et al.",
   "year": 2025,
   "month": 7,
   "topic": "scaffold",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 2,
   "url": "https://arxiv.org/abs/2507.19457",
   "code": "https://github.com/gepa-ai/gepa",
   "note": "The model reads its own execution traces and writes better prompts: beat GRPO reinforcement learning by 6% on average with up to 35× fewer rollouts.",
   "step": "analyzed",
   "record": "https://rsi.org.ai/records/gepa/",
   "map": []
  },
  {
   "id": "asi-arch",
   "title": "Neural Architecture Discovery via Autonomous Evolution",
   "authors": "Weixian Xu et al.",
   "year": 2025,
   "month": 7,
   "topic": "research",
   "listed": "2026-10-30",
   "kind": "paper",
   "url": "https://arxiv.org/abs/2507.18074",
   "code": "https://github.com/GAIR-NLP/ASI-Arch",
   "note": "Agents propose, code, train and analyze linear-attention architectures in a closed loop: 1,773 experiments produced 105 designs the authors report as state of the art.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "alphaevolve",
   "title": "AlphaEvolve: A coding agent for scientific and algorithmic discovery",
   "authors": "Alexander Novikov et al.",
   "year": 2025,
   "month": 6,
   "topic": "research",
   "listed": "2026-10-30",
   "kind": "paper",
   "url": "https://arxiv.org/abs/2506.13131",
   "code": "https://github.com/google-deepmind/alphaevolve_results",
   "note": "Models propose code edits, automatic evaluators score them, the best are kept; found a 48-multiplication 4×4 complex matrix product and sped up training of the model it runs on.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "seal",
   "name": "SEAL",
   "title": "Self-Adapting Language Models",
   "authors": "Adam Zweiger et al.",
   "year": 2025,
   "month": 6,
   "topic": "weights",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 4,
   "recursive": true,
   "url": "https://arxiv.org/abs/2506.10943",
   "code": "https://github.com/Continual-Intelligence/SEAL",
   "note": "The model writes its own fine-tuning data and settings (\"self-edits\"); reinforcement learning rewards self-edits that make it better after the update.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "llm-speedrunning",
   "title": "The Automated LLM Speedrunning Benchmark: Reproducing NanoGPT Improvements",
   "authors": "Bingchen Zhao et al.",
   "year": 2025,
   "month": 6,
   "topic": "measurement",
   "listed": "2026-10-30",
   "kind": "benchmark",
   "url": "https://arxiv.org/abs/2506.22419",
   "code": "https://github.com/facebookresearch/llm-speedrunner",
   "note": "19 tasks: reproduce each successive NanoGPT speedrun record from the previous record's code, scored by how much of the speedup the agent recovers.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "absolute-zero",
   "title": "Absolute Zero: Reinforced Self-play Reasoning with Zero Data",
   "authors": "Andrew Zhao et al.",
   "year": 2025,
   "month": 5,
   "topic": "weights",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 4,
   "recursive": true,
   "url": "https://arxiv.org/abs/2505.03335",
   "code": "https://github.com/LeapLabTHU/Absolute-Zero-Reasoner",
   "note": "One model invents coding tasks and solves them, checked by running code; it is also rewarded for inventing tasks it can learn from. No external data.",
   "step": "listed",
   "record": null,
   "map": [
    "check",
    "compound"
   ]
  },
  {
   "id": "dgm",
   "title": "Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents",
   "authors": "Jenny Zhang et al.",
   "year": 2025,
   "month": 5,
   "topic": "code",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 3,
   "recursive": true,
   "url": "https://arxiv.org/abs/2505.22954",
   "code": "https://github.com/jennyzzt/dgm",
   "note": "Keeps an archive of agent versions; each new version is an agent rewriting its own code, kept if it scores well: SWE-bench 20% → 50%.",
   "step": "analyzed",
   "record": "https://rsi.org.ai/records/dgm/",
   "map": [
    "check",
    "compound",
    "keep",
    "transfer",
    "safety"
   ]
  },
  {
   "id": "sica",
   "name": "SICA",
   "title": "A Self-Improving Coding Agent",
   "authors": "Maxime Robeyns, Martin Szummer and Laurence Aitchison",
   "year": 2025,
   "month": 4,
   "topic": "code",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 3,
   "recursive": true,
   "url": "https://arxiv.org/abs/2504.15228",
   "code": "https://github.com/MaximeRobeyns/self_improving_coding_agent",
   "note": "A coding agent edits its own codebase and keeps changes that score better: 17% → 53% on a SWE-bench Verified subset.",
   "step": "listed",
   "record": null,
   "map": [
    "check",
    "compound"
   ]
  },
  {
   "id": "ai-2027",
   "title": "AI 2027",
   "authors": "Daniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland and Romeo Dean",
   "year": 2025,
   "month": 4,
   "topic": "risk",
   "listed": "2026-10-30",
   "kind": "essay",
   "url": "https://ai-2027.com/",
   "note": "A month-by-month scenario in which AI agents automate AI research and progress speeds up sharply, written with two different endings.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "paperbench",
   "title": "PaperBench: Evaluating AI's Ability to Replicate AI Research",
   "authors": "Giulio Starace et al.",
   "year": 2025,
   "month": 4,
   "topic": "measurement",
   "listed": "2026-10-30",
   "kind": "benchmark",
   "url": "https://arxiv.org/abs/2504.01848",
   "code": "https://github.com/openai/frontier-evals",
   "note": "Agents replicate 20 ICML 2024 papers from scratch, graded on 8,316 rubric items; the best agent averaged 21.0%.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "iclr-2025-self-improving-workshop",
   "title": "Self-Improving Foundation Models Without Human Supervision",
   "authors": "ICLR 2025 workshop",
   "year": 2025,
   "month": 4,
   "topic": "community",
   "listed": "2026-10-30",
   "kind": "event",
   "url": "https://iclr.cc/virtual/2025/workshop/23971",
   "note": "Workshop on models that generate and curate their own training data: synthetic data, weak-to-strong generalization and self-supervision at inference time.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "ai-scientist-v2",
   "title": "The AI Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree Search",
   "authors": "Yutaro Yamada et al.",
   "year": 2025,
   "month": 4,
   "topic": "research",
   "listed": "2026-10-30",
   "kind": "paper",
   "url": "https://arxiv.org/abs/2504.08066",
   "code": "https://github.com/SakanaAI/AI-Scientist-v2",
   "note": "Drops human code templates and uses agentic tree search; produced the first fully AI-written paper accepted at a peer-reviewed workshop.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "forethought-sie",
   "title": "Will AI R&D Automation Cause a Software Intelligence Explosion?",
   "authors": "Daniel Eth and Tom Davidson",
   "year": 2025,
   "month": 3,
   "topic": "risk",
   "listed": "2026-10-30",
   "kind": "essay",
   "url": "https://www.forethought.org/research/will-ai-r-and-d-automation-cause-a-software-intelligence-explosion",
   "note": "Models whether AI that automates AI research would speed progress up without limit, given that ideas get harder to find; finds it plausible.",
   "step": "listed",
   "record": null,
   "map": [
    "returns"
   ]
  },
  {
   "id": "deepmind-fsf",
   "title": "Frontier Safety Framework (2025 update)",
   "authors": "Google DeepMind",
   "year": 2025,
   "month": 2,
   "topic": "risk",
   "listed": "2026-10-30",
   "kind": "policy",
   "url": "https://deepmind.google/discover/blog/updating-the-frontier-safety-framework/",
   "note": "Defines critical capability levels for machine-learning R&D, where a model could speed up AI development, and the safeguards each one triggers.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "self-improving-transformers",
   "name": "Self-Improving Transformers",
   "title": "Self-Improving Transformers Overcome Easy-to-Hard and Length Generalization Challenges",
   "authors": "Nayoung Lee et al.",
   "year": 2025,
   "month": 2,
   "topic": "weights",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 2,
   "url": "https://arxiv.org/abs/2502.01612",
   "note": "The model labels problems slightly harder than its training, trains on them, and repeats: from 10-digit to 100-digit addition.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "agent-laboratory",
   "title": "Agent Laboratory: Using LLM Agents as Research Assistants",
   "authors": "Samuel Schmidgall et al.",
   "year": 2025,
   "month": 1,
   "topic": "research",
   "listed": "2026-10-30",
   "kind": "paper",
   "url": "https://arxiv.org/abs/2501.04227",
   "code": "https://github.com/SamuelSchmidgall/AgentLaboratory",
   "note": "Agents carry out literature review, experiments and report writing, with optional human feedback: 84% lower research cost.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "multiagent-finetuning",
   "name": "Multiagent Finetuning",
   "title": "Multiagent Finetuning: Self Improvement with Diverse Reasoning Chains",
   "authors": "Vighnesh Subramaniam et al.",
   "year": 2025,
   "month": 1,
   "topic": "weights",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 2,
   "url": "https://arxiv.org/abs/2501.05707",
   "note": "Copies of one model are fine-tuned on different roles from their own debates, so self-training keeps improving instead of stalling.",
   "step": "listed",
   "record": null,
   "map": [
    "compound"
   ]
  },
  {
   "id": "openevolve",
   "title": "OpenEvolve",
   "authors": "Asankhaya Sharma and contributors",
   "year": 2025,
   "topic": "research",
   "listed": "2026-10-30",
   "kind": "project",
   "url": "https://github.com/algorithmicsuperintelligence/openevolve",
   "code": "https://github.com/algorithmicsuperintelligence/openevolve",
   "note": "Open-source implementation of AlphaEvolve's propose–evaluate–keep loop for evolving code with language models.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "re-bench",
   "title": "RE-Bench: Evaluating frontier AI R&D capabilities of language model agents against human experts",
   "authors": "Hjalmar Wijk et al.",
   "year": 2024,
   "month": 11,
   "topic": "measurement",
   "listed": "2026-10-30",
   "kind": "benchmark",
   "url": "https://arxiv.org/abs/2411.15114",
   "code": "https://github.com/METR/RE-Bench",
   "note": "7 research-engineering tasks with 71 eight-hour expert attempts. At 2 hours agents scored 4× higher than experts; at 32 hours experts scored 2× higher.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "aflow",
   "title": "AFlow: Automating Agentic Workflow Generation",
   "authors": "Jiayi Zhang et al.",
   "year": 2024,
   "month": 10,
   "topic": "scaffold",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 2,
   "url": "https://arxiv.org/abs/2410.10762",
   "code": "https://github.com/FoundationAgents/AFlow",
   "note": "Monte Carlo tree search over workflows written as code, using execution results to choose edits: 5.7% average gain over baselines.",
   "step": "analyzed",
   "record": "https://rsi.org.ai/records/aflow/",
   "map": [
    "data"
   ]
  },
  {
   "id": "godel-agent",
   "title": "Gödel Agent: A Self-Referential Agent Framework for Recursive Self-Improvement",
   "authors": "Xunjian Yin et al.",
   "year": 2024,
   "month": 10,
   "topic": "code",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 3,
   "recursive": true,
   "url": "https://arxiv.org/abs/2410.04444",
   "code": "https://github.com/Arvid-pku/Godel_Agent",
   "note": "The agent reads and rewrites its own running code, including its improvement logic, guided only by a high-level goal.",
   "step": "listed",
   "record": null,
   "map": [
    "compound"
   ]
  },
  {
   "id": "mle-bench",
   "title": "MLE-bench: Evaluating Machine Learning Agents on Machine Learning Engineering",
   "authors": "Jun Shern Chan et al.",
   "year": 2024,
   "month": 10,
   "topic": "measurement",
   "listed": "2026-10-30",
   "kind": "benchmark",
   "url": "https://arxiv.org/abs/2410.07095",
   "code": "https://github.com/openai/mle-bench",
   "note": "75 Kaggle competitions; the best setup (o1-preview with AIDE) reached bronze-medal level in 16.9% of them.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "agent-workflow-memory",
   "title": "Agent Workflow Memory",
   "authors": "Zora Zhiruo Wang et al.",
   "year": 2024,
   "month": 9,
   "topic": "scaffold",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 2,
   "url": "https://arxiv.org/abs/2409.07429",
   "note": "The agent extracts reusable workflows from past web tasks into its memory: +51.1% relative success on WebArena.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "score",
   "name": "SCoRe",
   "title": "Training Language Models to Self-Correct via Reinforcement Learning",
   "authors": "Aviral Kumar et al.",
   "year": 2024,
   "month": 9,
   "topic": "weights",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 2,
   "url": "https://arxiv.org/abs/2409.12917",
   "note": "Reinforcement learning on the model's own correction attempts teaches it to fix its first answer: self-correction +15.6% on MATH.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "adas",
   "name": "ADAS",
   "title": "Automated Design of Agentic Systems",
   "authors": "Shengran Hu, Cong Lu and Jeff Clune",
   "year": 2024,
   "month": 8,
   "topic": "scaffold",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 2,
   "url": "https://arxiv.org/abs/2408.08435",
   "code": "https://github.com/ShengranHu/ADAS",
   "note": "A fixed meta agent writes new agent designs as code, tests them, and archives the good ones for the next round.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "ai-scientist",
   "title": "The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery",
   "authors": "Chris Lu et al.",
   "year": 2024,
   "month": 8,
   "topic": "research",
   "listed": "2026-10-30",
   "kind": "paper",
   "url": "https://arxiv.org/abs/2408.06292",
   "code": "https://github.com/SakanaAI/AI-Scientist",
   "note": "Generates ideas, runs experiments, writes papers and reviews them, for under $15 per paper.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "meta-rewarding",
   "name": "Meta-Rewarding",
   "title": "Meta-Rewarding Language Models: Self-Improving Alignment with LLM-as-a-Meta-Judge",
   "authors": "Tianhao Wu et al.",
   "year": 2024,
   "month": 7,
   "topic": "weights",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 4,
   "recursive": true,
   "url": "https://arxiv.org/abs/2407.19594",
   "note": "The model also judges its own judgments, so its grading skill is trained too: AlpacaEval win rate 22.9% → 39.4% for Llama-3-8B-Instruct.",
   "step": "listed",
   "record": null,
   "map": [
    "compound"
   ]
  },
  {
   "id": "discopop",
   "title": "Discovering Preference Optimization Algorithms with and for Large Language Models",
   "authors": "Chris Lu et al.",
   "year": 2024,
   "month": 6,
   "topic": "research",
   "listed": "2026-10-30",
   "kind": "paper",
   "url": "https://arxiv.org/abs/2406.08414",
   "code": "https://github.com/SakanaAI/DiscoPOP",
   "note": "A model writes new preference-optimization loss functions, each is trained and scored, and the scores feed back; found DiscoPOP.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "situational-awareness",
   "title": "Situational Awareness: The Decade Ahead",
   "authors": "Leopold Aschenbrenner",
   "year": 2024,
   "month": 6,
   "topic": "risk",
   "listed": "2026-10-30",
   "kind": "essay",
   "url": "https://situational-awareness.ai/",
   "note": "Argues that automating AI research could compress a decade of algorithmic progress into a year or less.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "agent-symbolic-learning",
   "name": "Agent Symbolic Learning",
   "title": "Symbolic Learning Enables Self-Evolving Agents",
   "authors": "Wangchunshu Zhou et al.",
   "year": 2024,
   "month": 6,
   "topic": "scaffold",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 2,
   "url": "https://arxiv.org/abs/2406.18532",
   "code": "https://github.com/aiwaves-cn/agents",
   "note": "Agents update their own prompts, tools and pipeline from language \"losses\" and \"gradients\" written by a model.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "textgrad",
   "title": "TextGrad: Automatic \"Differentiation\" via Text",
   "authors": "Mert Yuksekgonul et al.",
   "year": 2024,
   "month": 6,
   "topic": "scaffold",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 2,
   "url": "https://arxiv.org/abs/2406.07496",
   "code": "https://github.com/zou-group/textgrad",
   "note": "Uses written feedback from a model as a \"gradient\" to edit prompts, code or answers: GPQA 51% → 55% for GPT-4o.",
   "step": "analyzed",
   "record": "https://rsi.org.ai/records/textgrad/",
   "map": []
  },
  {
   "id": "alphallm",
   "name": "AlphaLLM",
   "title": "Toward Self-Improvement of LLMs via Imagination, Searching, and Criticizing",
   "authors": "Ye Tian et al.",
   "year": 2024,
   "month": 4,
   "topic": "weights",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 2,
   "url": "https://arxiv.org/abs/2404.12253",
   "note": "Combines tree search, self-generated prompts and self-critique to produce its own training data, aiming at improvement without new labels.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "quiet-star",
   "title": "Quiet-STaR: Language Models Can Teach Themselves to Think Before Speaking",
   "authors": "Eric Zelikman et al.",
   "year": 2024,
   "month": 3,
   "topic": "weights",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 2,
   "url": "https://arxiv.org/abs/2403.09629",
   "note": "Learns to produce hidden reasoning before each token, rewarded when that reasoning makes the following text more predictable.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "gptswarm",
   "name": "GPTSwarm",
   "title": "Language Agents as Optimizable Graphs",
   "authors": "Mingchen Zhuge et al.",
   "year": 2024,
   "month": 2,
   "topic": "scaffold",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 2,
   "url": "https://arxiv.org/abs/2402.16823",
   "code": "https://github.com/metauto-ai/GPTSwarm",
   "note": "Treats a multi-agent system as a graph and optimizes both the prompt at each node and which nodes connect.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "v-star",
   "title": "V-STaR: Training Verifiers for Self-Taught Reasoners",
   "authors": "Arian Hosseini et al.",
   "year": 2024,
   "month": 2,
   "topic": "weights",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 2,
   "url": "https://arxiv.org/abs/2402.06457",
   "note": "Also trains a verifier on its own correct and incorrect solutions, then uses it to pick answers: 4–17% better on code and math.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "spin",
   "name": "SPIN",
   "title": "Self-Play Fine-Tuning Converts Weak Language Models to Strong Language Models",
   "authors": "Zixiang Chen et al.",
   "year": 2024,
   "month": 1,
   "topic": "weights",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 2,
   "url": "https://arxiv.org/abs/2401.01335",
   "note": "The model learns to tell human-written answers from its own previous round's answers, which pushes it toward the human ones without new labels.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "self-rewarding",
   "name": "Self-Rewarding",
   "title": "Self-Rewarding Language Models",
   "authors": "Weizhe Yuan et al.",
   "year": 2024,
   "month": 1,
   "topic": "weights",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 4,
   "recursive": true,
   "url": "https://arxiv.org/abs/2401.10020",
   "note": "The model grades its own answers and trains on those grades; over three rounds both its answers and its grading improved.",
   "step": "listed",
   "record": null,
   "map": [
    "compound"
   ]
  },
  {
   "id": "rest-em",
   "name": "ReST-EM",
   "title": "Beyond Human Data: Scaling Self-Training for Problem-Solving with Language Models (ReST-EM)",
   "authors": "Avi Singh et al.",
   "year": 2023,
   "month": 12,
   "topic": "weights",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 2,
   "url": "https://arxiv.org/abs/2312.06585",
   "note": "Generate solutions, keep the ones a checker marks correct, fine-tune, repeat. On MATH and APPS this beat fine-tuning on human-written solutions.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "funsearch",
   "title": "Mathematical discoveries from program search with large language models",
   "authors": "Bernardino Romera-Paredes et al.",
   "year": 2023,
   "month": 12,
   "topic": "research",
   "listed": "2026-10-30",
   "kind": "paper",
   "url": "https://doi.org/10.1038/s41586-023-06924-6",
   "code": "https://github.com/google-deepmind/funsearch",
   "note": "A model proposes programs, an evaluator scores them, and the best seed the next prompts; found new constructions for the cap set problem.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "weak-to-strong",
   "title": "Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision",
   "authors": "Collin Burns et al.",
   "year": 2023,
   "month": 12,
   "topic": "risk",
   "listed": "2026-10-30",
   "kind": "paper",
   "url": "https://arxiv.org/abs/2312.09390",
   "note": "Can a weaker supervisor train a stronger model? GPT-4 supervised by a GPT-2-level model recovers close to GPT-3.5-level performance.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "dspy",
   "title": "DSPy: Compiling Declarative Language Model Calls into Self-Improving Pipelines",
   "authors": "Omar Khattab et al.",
   "year": 2023,
   "month": 10,
   "topic": "scaffold",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 2,
   "url": "https://arxiv.org/abs/2310.03714",
   "code": "https://github.com/stanfordnlp/dspy",
   "note": "Compiles pipelines of model calls: generates and selects its own examples and instructions for each step against a metric.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "cannot-self-correct",
   "title": "Large Language Models Cannot Self-Correct Reasoning Yet",
   "authors": "Jie Huang et al.",
   "year": 2023,
   "month": 10,
   "topic": "risk",
   "listed": "2026-10-30",
   "kind": "paper",
   "url": "https://arxiv.org/abs/2310.01798",
   "note": "Without outside feedback, models asked to correct their own reasoning often get worse rather than better.",
   "step": "analyzed",
   "record": "https://rsi.org.ai/records/cannot-self-correct/",
   "map": [
    "check"
   ]
  },
  {
   "id": "stop",
   "name": "STOP",
   "title": "Self-Taught Optimizer (STOP): Recursively Self-Improving Code Generation",
   "authors": "Eric Zelikman et al.",
   "year": 2023,
   "month": 10,
   "topic": "code",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 3,
   "recursive": true,
   "url": "https://arxiv.org/abs/2310.02304",
   "code": "https://github.com/microsoft/stop",
   "note": "A scaffolding program that uses a language model to improve programs is run on its own source code to improve itself.",
   "step": "listed",
   "record": null,
   "map": [
    "compound"
   ]
  },
  {
   "id": "evoprompt",
   "title": "EvoPrompt: Connecting LLMs with Evolutionary Algorithms Yields Powerful Prompt Optimizers",
   "authors": "Qingyan Guo et al.",
   "year": 2023,
   "month": 9,
   "topic": "scaffold",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 2,
   "url": "https://arxiv.org/abs/2309.08532",
   "note": "Evolutionary search in which the model itself performs crossover and mutation on prompts: up to 25% gain on Big-Bench Hard.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "opro",
   "name": "OPRO",
   "title": "Large Language Models as Optimizers",
   "authors": "Chengrun Yang et al.",
   "year": 2023,
   "month": 9,
   "topic": "scaffold",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 2,
   "url": "https://arxiv.org/abs/2309.03409",
   "note": "The model sees past prompts with their scores and writes a better one: up to 8% over human prompts on GSM8K and 50% on Big-Bench Hard.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "promptbreeder",
   "title": "Promptbreeder: Self-Referential Self-Improvement Via Prompt Evolution",
   "authors": "Chrisantha Fernando et al.",
   "year": 2023,
   "month": 9,
   "topic": "scaffold",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 3,
   "recursive": true,
   "url": "https://arxiv.org/abs/2309.16797",
   "note": "Evolves task prompts and also the mutation prompts that rewrite them, so the way it improves prompts improves too.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "anthropic-rsp",
   "title": "Responsible Scaling Policy",
   "authors": "Anthropic",
   "year": 2023,
   "month": 9,
   "topic": "risk",
   "listed": "2026-10-30",
   "kind": "policy",
   "url": "https://www.anthropic.com/responsible-scaling-policy",
   "note": "Capability thresholds that trigger stronger safeguards, including a threshold for models that can automate AI research and development.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "expel",
   "name": "ExpeL",
   "title": "ExpeL: LLM Agents Are Experiential Learners",
   "authors": "Andrew Zhao et al.",
   "year": 2023,
   "month": 8,
   "topic": "scaffold",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 2,
   "url": "https://arxiv.org/abs/2308.10144",
   "note": "The agent collects experience across tasks and distills it into written insights it reads before the next task; no weight updates.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "rest",
   "name": "ReST",
   "title": "Reinforced Self-Training (ReST) for Language Modeling",
   "authors": "Caglar Gulcehre et al.",
   "year": 2023,
   "month": 8,
   "topic": "weights",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 2,
   "url": "https://arxiv.org/abs/2308.08998",
   "note": "Samples from the current model, scores samples with a reward model, fine-tunes on the best, and repeats with a rising threshold.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "alphadev",
   "title": "Faster sorting algorithms discovered using deep reinforcement learning",
   "authors": "Daniel J. Mankowitz et al.",
   "year": 2023,
   "month": 6,
   "topic": "research",
   "listed": "2026-10-30",
   "kind": "paper",
   "url": "https://doi.org/10.1038/s41586-023-06004-9",
   "note": "A reinforcement-learning agent searched assembly instructions and found faster sorting routines, now in the LLVM C++ standard library.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "voyager",
   "title": "Voyager: An Open-Ended Embodied Agent with Large Language Models",
   "authors": "Guanzhi Wang et al.",
   "year": 2023,
   "month": 5,
   "topic": "scaffold",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 2,
   "url": "https://arxiv.org/abs/2305.16291",
   "code": "https://github.com/MineDojo/Voyager",
   "note": "A Minecraft agent writes code skills, checks them in the game, and stores the ones that work in a growing skill library.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "reflexion",
   "title": "Reflexion: Language Agents with Verbal Reinforcement Learning",
   "authors": "Noah Shinn et al.",
   "year": 2023,
   "month": 3,
   "topic": "scaffold",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 1,
   "url": "https://arxiv.org/abs/2303.11366",
   "code": "https://github.com/noahshinn/reflexion",
   "note": "After a failed attempt the agent writes a reflection into memory and retries: 91% on HumanEval versus 80% for GPT-4 alone.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "self-refine",
   "title": "Self-Refine: Iterative Refinement with Self-Feedback",
   "authors": "Aman Madaan et al.",
   "year": 2023,
   "month": 3,
   "topic": "scaffold",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 1,
   "url": "https://arxiv.org/abs/2303.17651",
   "code": "https://github.com/madaan/self-refine",
   "note": "One model drafts, critiques and revises its own answer at test time: about 20% better on average across 7 tasks.",
   "step": "analyzed",
   "record": "https://rsi.org.ai/records/self-refine/",
   "map": []
  },
  {
   "id": "lion",
   "title": "Symbolic Discovery of Optimization Algorithms",
   "authors": "Xiangning Chen et al.",
   "year": 2023,
   "month": 2,
   "topic": "research",
   "listed": "2026-10-30",
   "kind": "paper",
   "url": "https://arxiv.org/abs/2302.06675",
   "note": "Program search discovered Lion, an optimizer simpler than Adam that saved up to 5× pre-training compute on JFT.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "constitutional-ai",
   "name": "Constitutional AI",
   "title": "Constitutional AI: Harmlessness from AI Feedback",
   "authors": "Yuntao Bai et al.",
   "year": 2022,
   "month": 12,
   "topic": "weights",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 2,
   "url": "https://arxiv.org/abs/2212.08073",
   "note": "The model critiques and revises its own outputs against a written list of principles, then trains on AI preference labels instead of human ones.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "self-instruct",
   "title": "Self-Instruct: Aligning Language Models with Self-Generated Instructions",
   "authors": "Yizhong Wang et al.",
   "year": 2022,
   "month": 12,
   "topic": "weights",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 2,
   "url": "https://arxiv.org/abs/2212.10560",
   "note": "The model writes its own instructions and examples, filters them, and fine-tunes on them: +33 points for GPT-3 on Super-NaturalInstructions.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "ape",
   "name": "APE",
   "title": "Large Language Models Are Human-Level Prompt Engineers",
   "authors": "Yongchao Zhou et al.",
   "year": 2022,
   "month": 11,
   "topic": "scaffold",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 2,
   "url": "https://arxiv.org/abs/2211.01910",
   "note": "The model proposes instructions for a task and keeps the best-scoring ones; matched or beat human prompts on 19 of 24 tasks.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "velo",
   "title": "VeLO: Training Versatile Learned Optimizers by Scaling Up",
   "authors": "Luke Metz et al.",
   "year": 2022,
   "month": 11,
   "topic": "research",
   "listed": "2026-10-30",
   "kind": "paper",
   "url": "https://arxiv.org/abs/2211.09760",
   "note": "A small neural network trained to be a general-purpose optimizer for deep learning, meta-trained with about 4,000 TPU-months of compute.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "lmsi",
   "name": "LMSI",
   "title": "Large Language Models Can Self-Improve",
   "authors": "Jiaxin Huang et al.",
   "year": 2022,
   "month": 10,
   "topic": "weights",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 2,
   "url": "https://arxiv.org/abs/2210.11610",
   "note": "Answers unlabeled questions many times and trains on its own majority-vote answers: GSM8K 74.4% → 82.1% with no labels.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "reward-overoptimization",
   "title": "Scaling Laws for Reward Model Overoptimization",
   "authors": "Leo Gao, John Schulman and Jacob Hilton",
   "year": 2022,
   "month": 10,
   "topic": "risk",
   "listed": "2026-10-30",
   "kind": "paper",
   "url": "https://arxiv.org/abs/2210.10760",
   "note": "Optimizing against a learned reward model first raises and then lowers true quality; measures how the turning point scales with reward-model size.",
   "step": "analyzed",
   "record": "https://rsi.org.ai/records/reward-overoptimization/",
   "map": [
    "check"
   ]
  },
  {
   "id": "elm",
   "title": "Evolution through Large Models",
   "authors": "Joel Lehman et al.",
   "year": 2022,
   "month": 6,
   "topic": "research",
   "listed": "2026-10-30",
   "kind": "paper",
   "url": "https://arxiv.org/abs/2206.08896",
   "note": "Uses a code model as the mutation step of evolutionary search; produced hundreds of thousands of working programs for walking robots.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "star",
   "title": "STaR: Bootstrapping Reasoning With Reasoning",
   "authors": "Eric Zelikman et al.",
   "year": 2022,
   "month": 3,
   "topic": "weights",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 2,
   "url": "https://arxiv.org/abs/2203.14465",
   "note": "The model writes step-by-step reasoning; only reasoning that reaches the correct answer is kept and trained on; repeat.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "automl-zero",
   "title": "AutoML-Zero: Evolving Machine Learning Algorithms From Scratch",
   "authors": "Esteban Real et al.",
   "year": 2020,
   "month": 3,
   "topic": "research",
   "listed": "2026-10-30",
   "kind": "paper",
   "url": "https://arxiv.org/abs/2003.03384",
   "note": "Evolves whole learning algorithms from basic math operations; rediscovered gradient descent and techniques such as weight averaging.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "alphazero",
   "name": "AlphaZero",
   "title": "Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm",
   "authors": "David Silver et al.",
   "year": 2017,
   "month": 12,
   "topic": "weights",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 2,
   "url": "https://arxiv.org/abs/1712.01815",
   "note": "Learns chess, shogi and Go only by playing itself; within 24 hours it beat a world-champion program in each game.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "exit",
   "name": "Expert Iteration",
   "title": "Thinking Fast and Slow with Deep Learning and Tree Search",
   "authors": "Thomas Anthony, Zheng Tian and David Barber",
   "year": 2017,
   "month": 5,
   "topic": "weights",
   "listed": "2026-10-30",
   "kind": "paper",
   "level": 2,
   "url": "https://arxiv.org/abs/1705.08439",
   "note": "Tree search finds better moves than the network; the network trains to imitate them; the better network makes the search stronger.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "nas",
   "title": "Neural Architecture Search with Reinforcement Learning",
   "authors": "Barret Zoph and Quoc V. Le",
   "year": 2016,
   "month": 11,
   "topic": "research",
   "listed": "2026-10-30",
   "kind": "paper",
   "url": "https://arxiv.org/abs/1611.01578",
   "note": "A controller network proposes architectures and learns from their accuracy; it designed a CIFAR-10 network that rivaled the best human ones.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "l2l",
   "title": "Learning to learn by gradient descent by gradient descent",
   "authors": "Marcin Andrychowicz et al.",
   "year": 2016,
   "month": 6,
   "topic": "research",
   "listed": "2026-10-30",
   "kind": "paper",
   "url": "https://arxiv.org/abs/1606.04474",
   "note": "Trains a neural network to act as the optimizer for other networks, replacing a hand-designed update rule.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "everitt-2016",
   "title": "Self-Modification of Policy and Utility Function in Rational Agents",
   "authors": "Tom Everitt et al.",
   "year": 2016,
   "month": 5,
   "topic": "foundations",
   "listed": "2026-10-30",
   "kind": "paper",
   "url": "https://arxiv.org/abs/1605.03142",
   "note": "Formal analysis of when a rational agent able to edit its own policy or utility function would choose to keep its goals.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "yampolskiy-2015",
   "title": "From Seed AI to Technological Singularity via Recursively Self-Improving Software",
   "authors": "Roman V. Yampolskiy",
   "year": 2015,
   "month": 2,
   "topic": "foundations",
   "listed": "2026-10-30",
   "kind": "paper",
   "url": "https://arxiv.org/abs/1502.06512",
   "note": "Classifies kinds of self-improving software and lists limits on how far recursive self-improvement can go.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "bostrom-2014",
   "title": "Superintelligence: Paths, Dangers, Strategies",
   "authors": "Nick Bostrom",
   "year": 2014,
   "topic": "foundations",
   "listed": "2026-10-30",
   "kind": "book",
   "url": "https://global.oup.com/academic/product/superintelligence-9780199678112",
   "note": "Analyzes how fast a self-improving system could take off, and why keeping control of it is hard.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "bounded-rsi",
   "title": "Bounded Recursive Self-Improvement",
   "authors": "E. Nivel et al.",
   "year": 2013,
   "month": 12,
   "topic": "foundations",
   "listed": "2026-10-30",
   "kind": "paper",
   "recursive": true,
   "url": "https://arxiv.org/abs/1312.6764",
   "note": "Describes an implemented architecture that keeps building and revising its own models after it leaves the lab, within bounds its designers set.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "iem-2013",
   "title": "Intelligence Explosion Microeconomics",
   "authors": "Eliezer Yudkowsky",
   "year": 2013,
   "topic": "foundations",
   "listed": "2026-10-30",
   "kind": "paper",
   "url": "https://intelligence.org/files/IEM.pdf",
   "note": "Frames RSI as a question of returns: does each unit of added intelligence buy more or less than one unit of further improvement?",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "tiling-agents-2013",
   "title": "Tiling Agents for Self-Modifying AI, and the Löbian Obstacle",
   "authors": "Eliezer Yudkowsky and Marcello Herreshoff",
   "year": 2013,
   "topic": "foundations",
   "listed": "2026-10-30",
   "kind": "paper",
   "url": "https://intelligence.org/files/TilingAgentsDraft.pdf",
   "note": "Shows why an agent that acts only on proofs cannot easily prove that a successor using the same logic is safe to build (the Löbian obstacle).",
   "step": "listed",
   "record": null,
   "map": [
    "safety"
   ]
  },
  {
   "id": "chalmers-2010",
   "title": "The Singularity: A Philosophical Analysis",
   "authors": "David J. Chalmers",
   "year": 2010,
   "topic": "foundations",
   "listed": "2026-10-30",
   "kind": "paper",
   "url": "https://consc.net/papers/singularity.pdf",
   "note": "Sets out the \"AI+ builds AI++\" argument as explicit premises and examines which of them could fail.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "omohundro-2008",
   "title": "The Basic AI Drives",
   "authors": "Stephen M. Omohundro",
   "year": 2008,
   "topic": "foundations",
   "listed": "2026-10-30",
   "kind": "paper",
   "url": "https://selfawaresystems.com/wp-content/uploads/2008/01/ai_drives_final.pdf",
   "note": "Argues that almost any goal-seeking system will want to improve itself, and to protect its goals while doing so.",
   "step": "listed",
   "record": null,
   "map": [
    "safety"
   ]
  },
  {
   "id": "godel-machine",
   "title": "Goedel Machines: Self-Referential Universal Problem Solvers Making Provably Optimal Self-Improvements",
   "authors": "Juergen Schmidhuber",
   "year": 2003,
   "month": 9,
   "topic": "foundations",
   "listed": "2026-10-30",
   "kind": "paper",
   "recursive": true,
   "url": "https://arxiv.org/abs/cs/0309048",
   "note": "A program that rewrites any part of its own code, including the proof searcher that decides on rewrites, once it has proved the rewrite raises expected utility.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "schmidhuber-1987",
   "title": "Evolutionary Principles in Self-Referential Learning (diploma thesis)",
   "authors": "Jürgen Schmidhuber",
   "year": 1987,
   "topic": "foundations",
   "listed": "2026-10-30",
   "kind": "paper",
   "recursive": true,
   "url": "https://people.idsia.ch/~juergen/diploma1987ocr.pdf",
   "note": "Proposes learners that also learn how they learn, with meta-levels stacked without a fixed top: an early design for self-referential improvement.",
   "step": "listed",
   "record": null,
   "map": []
  },
  {
   "id": "good-1965",
   "title": "Speculations Concerning the First Ultraintelligent Machine",
   "authors": "I. J. Good",
   "year": 1965,
   "topic": "foundations",
   "listed": "2026-10-30",
   "kind": "paper",
   "url": "https://doi.org/10.1016/S0065-2458(08)60418-0",
   "note": "Argues that a machine better than people at designing machines would design a better one, setting off an \"intelligence explosion\".",
   "step": "listed",
   "record": null,
   "map": []
  }
 ]
}