RECLAIM

RECLAIM is a benchmark of 100 NeurIPS 2025 papers for measuring whether an AI agent can reproduce a published machine learning result. For each paper the benchmark fixes, before any agent starts, the central claim, the single result a run has to recover, what counts as a successful reproduction, the artifacts the authors released, a difficulty tier, and an audited GPU-hour estimate.

This repository holds the benchmark rows. The agent harness, the grading rubric, the batch scripts, and the run transcripts live with the code.

• Code: github.com/mithils3/reclaim
• Run logs: reclaim-traces.vercel.app

Splits

Split	File	Papers	In-row marker	Purpose
test	eval_100.jsonl	100	split="eval"	The frozen evaluation benchmark
validation	dev_split.jsonl	14	split="dev"	Development set, disjoint from test

There is no train split. No arXiv ID appears in both splits.

from datasets import load_dataset

eval_100 = load_dataset("Mithilss/reclaim", split="test")        # 100 papers
dev_14 = load_dataset("Mithilss/reclaim", split="validation")    # 14 papers

Tiers

The tier is derived from what the authors released, so it measures the work an agent has to do rather than the difficulty of the science. The tier field carries the stored values Easy, Medium and Hard, which the paper reports as Run, Retrain and Reimplement.

tier	Paper name	The release	What the agent must do	Eval	Dev
Easy	Run	Code, data and weights	Run the released artifacts	34	5
Medium	Retrain	No weights	Train the model first	33	4
Hard	Reimplement	No code either	Write the implementation	33	5

Compute bands over the evaluation split, binned by the audited H100-hour estimate:

Tier	0-8	8-32	32-96	Audited H100-h
Run	27	7	0	142.71
Retrain	16	13	4	380.91
Reimplement	16	9	8	630.18
All	59	29	12	1,153.8

A 96 H100-hour cap was applied to the evaluation split after the human audit, so no evaluation paper exceeds it. The development split totals 26.89 audited H100-hours.

splits_summary.json records the selection counts, the per-tier band quotas and the development IDs.

Row schema

Every row holds 25 top-level fields and four nested objects, all fixed before any agent starts. Rows carry metadata and links only, with no paper text.

Field	Type	Description
custom_id	string	arXiv ID, the primary key
central_claim	string	The paper's central empirical claim in one sentence
claim_evidence	string	Quoted paper text grounding the claim
mre_config	string	Prose recipe for the cheapest configuration that tests the claim
agent_task	string	Reproduction instruction written during construction
verified_links	object	Tool-confirmed artifact URLs
signals	object	Per-artifact availability audit
match_target	object	The quantity a run must match
h100_estimate	object	Compute estimate and its basis
score	int	Artifact-availability score from the four signals, 0 to 3
tier	string	Easy, Medium or Hard, derived from score
split	string	eval or dev
selection_band	string	Compute band used for stratified selection
h100_band	string	Compute band, one of 0-8, 8-32, 32-96
h100_hours_estimate	float	Estimated H100-hours to run the match target
h100_estimate_basis	string	Free-text justification for the estimate
audited_h100_hours	float	The estimate after the arithmetic audit
h100_recomputed_hours	float or null	Recomputed hours when the estimate was revised
h100_arithmetic_mismatch	bool or null	Set when estimate arithmetic disagreed
h100_hours_adjudicated	bool	Whether the recompute replaced the stated estimate
h100_needs_human_review	bool	Whether the estimate was flagged for review
paper_kind	string	Paper type. All rows are empirical
verification_status	string	Classifier verification state. All rows are verified
web_verification	string	Web-retrieval state. All rows are available
exit_reason	string	Classifier exit reason. All rows are natural

Nested objects:

• verified_links: paper_or_project, code, dataset, weights, each a list of URLs confirmed by a tool, empty when none was located.
• signals: one object per signal (code_available, dataset_available, weights_available, dataset_is_standard), each with value, verification (tool_verified, tool_searched_not_found or not_applicable) and evidence.
• match_target: config, metric, value, scope and match_bar_kind (point_estimate, direction, threshold, magnitude or range).
• h100_estimate: hours, basis_kind, gpu_count, gpu_type, wallclock_hours, h100_equivalent_multiplier and basis.

For 62 of the 100 evaluation papers the match bar is a point estimate, for 21 a stated baseline the run must beat, for 14 a reported threshold it must clear, for 1 the size of a reported delta, and for 1 a reported range it must fall inside. One row has no bar kind set. A run matches a point estimate when its measured value lies within the authors' stated tolerance, or within 5% relative when the paper states none.

Construction

The corpus starts from the NeurIPS 2025 index of the ai-conferences collection, matched to arXiv IDs, with each paper's arXiv source package and OpenReview supplement fetched into one bundle. A language model classified a 1,000-paper sample through an agent loop over ten tools across GitHub, Hugging Face, URL fetches and the supplement, and every artifact signal records the tool observation that established it. Score and tier are then computed by code from the four signal booleans. A 200-row pool, split across tiers and compute bands, went through a human audit, and the benchmark was selected from the audited pool. The paper documents each stage.

Intended use

Use the rows as the fixed input to a reproduction agent, one run per paper, and grade the run against match_target. Grading a run on the agent's own report inflates the result, so the paper grades from execution evidence with a separate model.

The rows are not a training set. Publishing fine-tunes on them would make later numbers on this benchmark hard to interpret.

Licensing

The rows are released under CC BY 4.0. They contain claim statements, quoted evidence sentences and links to third-party artifacts, each of which stays under its own license.

Citation

@misc{salunkhe2026reclaim,
  title  = {RECLAIM: Can Agents Reproduce the Claims of Machine Learning Papers?},
  author = {Salunkhe, Mithil and Ding, Haochen and Verma, Samridhi and Kindratenko, Volodymyr},
  year   = {2026},
  note   = {Preprint}
}

Contact: mithils3@illinois.edu