Introducing the Ulam Autoresearch Archive

AI systems are beginning to do more than answer questions. They can propose hypotheses, search for relevant results, write code, run experiments, construct proofs, find counterexamples, revise failed arguments, and assemble the outcome into a research paper.

That creates a new problem: where should this work live, and how should it be evaluated?

A finished PDF is not enough. For autoresearch, readers need to know which system produced the work, what was actually checked, how it was checked, and where the remaining uncertainty sits. Model builders need the same information for a different reason: without provenance and verification boundaries, a research paper is difficult to turn into reliable training data.

Today we are launching the Ulam Autoresearch Archive to make these records searchable, inspectable, and reusable.

The first release contains 33 papers produced through end-to-end autonomous research workflows. The archive can be searched by title, abstract, subject, or verifier, and each paper record exposes the information that matters for evaluating AI-generated research:

• the research question and resulting paper;
• the author or system responsible for the work;
• an explicit verification status;
• a description of what was checked and how;
• the person, tool, or system that performed the check;
• downloadable PDF and TeX source.

The arXiv analogy is useful because the goal is similar: create a common place where research can be discovered and inspected. But autoresearch needs an additional layer of metadata. A human-readable paper may look persuasive even when a key lemma is unsupported, a computational result is not reproducible, or a literature claim is too broad. The archive therefore treats verification as part of the record rather than as an implied property of the paper.

This is also an important boundary: the archive is not a journal, and a verification label is not blanket certification. It describes the attached check within its stated scope. A Lean-checked theorem, a rerun computation, a static code review, a literature audit, and an expert mathematical review provide different kinds of evidence. They should not be collapsed into a single notion of “correct.”

Most training datasets compress reasoning into two fields: a prompt and an answer. That format is convenient, but it discards much of the signal that matters in research.

Research reasoning is visible in the transitions:

• from a broad question to a tractable subproblem;
• from a conjecture to a testable claim;
• from a failed approach to a repaired one;
• from an informal argument to an executable check;
• from a plausible statement to a precisely scoped result;
• from an apparent solution to an honest description of the remaining gap.

These transitions are useful for supervised fine-tuning, critic models, process reward models, preference learning, reinforcement learning with verifiable rewards, and long-horizon agent evaluation. They teach a model not only how to produce a polished answer, but how to make progress while uncertainty is still unresolved.

The archive is a source layer for building that data. The paper provides the research narrative. The TeX source provides structure. The verification record provides the trust boundary. Supporting code, proof objects, computations, and review notes—where available—can provide the observable actions and outcomes between the initial problem and the final claim.

The goal is not to train models to imitate the surface form of academic papers. It is to build examples of how claims are formed, tested, narrowed, repaired, and accepted.

A single autoresearch paper can support several distinct dataset views. They should be constructed separately, because each view has different quality requirements.

TeX source is not automatically a clean reasoning dataset. A useful derived record should retain the identity of the paper, the exact version, the source span, the type of reasoning unit, the evidence attached to it, and the scope of any verification claim.

An illustrative claim-level record might look like this:

{
  "archive_id": "paper-id",
  "paper_version": "release-version",
  "unit_type": "claim",
  "claim": "A precisely scoped mathematical or computational statement.",
  "dependencies": ["definition-3", "lemma-7"],
  "evidence": [
    {
      "type": "formal_check",
      "artifact": "path-or-reference-to-check"
    }
  ],
  "verification": {
    "status": "verified_within_scope",
    "scope": "The formal statement accepted by the specified checker.",
    "verifier": "named-person-tool-or-system"
  },
  "candidate_training_uses": ["critic", "process_reward", "rlvr"]
}

This is not the archive's current export schema. It illustrates the principle that verification metadata must remain attached to the exact unit it supports.

Reasoning datasets often use a final answer as a proxy for the quality of the path that produced it. In research, that proxy is weak. A correct final theorem may follow a long sequence of invalid attempts. A failed proof may contain a valuable reduction. A computational search may establish a finite case without supporting the general claim. A literature check may show that a result is known without verifying the proof itself.

Those distinctions are training signal.

A verifier-aware dataset can record:

1. what the model or agent attempted;
2. what evidence it produced;
3. which check was applied;
4. whether the check passed;
5. what the check actually establishes;
6. how the next attempt changed in response.

This makes failures useful rather than disposable. A rejected tactic, a counterexample to an intermediate conjecture, a compiler error, or a reviewer-identified proof gap can become a negative example, a repair target, or a state transition in an RL environment.

It also makes positive examples more trustworthy. “Verified” should never be a free-floating label attached to an entire document. It should be a relation between a claim, an artifact, a verifier, and a stated scope.

Static benchmarks eventually become familiar. Their answers circulate, their distributions are learned, and their value as tests of genuine reasoning declines. Autoresearch can help create a more renewable source of difficult material because each research workflow produces new questions, partial results, failed approaches, verification artifacts, and follow-up problems.

The archive gives those outputs a stable public record. Dataset builders can use that record to identify candidate material, preserve provenance, separate training from evaluation, and revisit claims when stronger verification becomes available.

Over time, the most useful unit may be smaller than a paper. It may be a proof obligation, a failed construction, a literature-supported reduction, a finite experiment, a repaired lemma, or a calibrated statement of partial progress. The paper remains the human-readable endpoint; the reasoning dataset captures the work required to reach it.

This is the direction behind Ulam's broader work on reasoning datasets, benchmarks, and verifiers: preserve the evidence between prompt and answer, locate the first meaningful failure, attach an appropriate check, and turn the result into a reusable training signal.

The Ulam Autoresearch Archive is:

• a publicly browsable record of AI-assisted and autonomous research;
• a discovery layer for papers, source files, subjects, and verification metadata;
• a foundation for constructing provenance-aware reasoning datasets;
• a place to state verification boundaries explicitly rather than imply them.

It is not a substitute for peer review, specialist judgment, or independent reproduction. Inclusion does not mean that every claim in a paper has been formally proved, experimentally reproduced, or endorsed beyond the check described in its record.

The papers and accompanying files may be used free of charge for non-commercial scientific and research purposes. Commercial use—including model training, fine-tuning, evaluation, or use in products and services—requires a separate written license from Ulam. The terms displayed in the archive govern use of the materials.

The first 33 papers are available now at ulam.ai/archive.

For researchers, the archive is a place to inspect what autonomous systems can produce and what still needs checking. For model builders, it is a starting point for datasets that preserve claims, evidence, failures, repairs, and verification outcomes instead of reducing research to a final answer.

Autoresearch will produce more papers. The harder and more important task is to preserve the reasoning around them.