IdeaEval-Bench

The ideation benchmark from IdeaEval. It has 91 cohorts across 40 research topics. Each cohort is a group of real papers that worked on the same research question within the same 31-day window, rewritten as pre-experiment research proposals. An ideation system receives the cohort's research direction and cutoff date, writes one proposal using only literature from before the cutoff, and is compared against every human proposal in the cohort.

📝 Blog post · 💻 Code (construction pipeline: benchmark/docs/PIPELINE.md)

Read this first. We never obtained an evaluator we trust on human-vs-AI comparisons (see the blog post and the IdeaEval-Calibration card). The cohorts and proposals are released as a resource. The win rates in the blog post are not a validated leaderboard.

How the cohorts were built

1. Corpus: 35,505 arXiv AI papers, Sep 2025 – Jan 2026 (IdeaEval-Corpus).
2. Research questions: tree clustering adapted from TaxoAdapt on each paper's title, research problem, task formulation and benchmarks. Levels were split until every leaf held 10–50 papers. The top two levels were deduplicated by hand. A paper could belong to 1–3 nodes.
3. Topics: subtrees holding at most 150 papers. 70 reasonably active topics were hand-picked.
4. Cohorts: each topic was split into 31-day windows. Zero-citation papers were dropped, and papers were audited for topic fit, presence of experiments and paper type. Papers that were public before their arXiv date (journal versions, OpenReview submissions) were removed or re-dated (date_remapped). Cohorts with at least 10 remaining papers were kept: 91 cohorts over 40 topics.
5. Cutoff: the day before the earliest paper in the cohort.
6. Proposals: every paper was converted into a proposal with a fixed structure (Introduction with core insight, falsifiable hypothesis and failure modes; Proposed Approach; Related Work; Experiments; Success Criteria; Impact Statement; References). Experimental design is kept in full, and the proposed method's numbers in the main results table are replaced with TBD. The converter is benchmark/paper2proposal (Claude Opus 4.6).

Files

cohorts.parquet (91 rows)

Field	Description
cohort_id	e.g. l1_001_02_03_w2 (topic id + window index)
topic_id, topic_label, topic_question, topic_description	The research topic
window_start, window_end	The 31-day window
cutoff_date	Literature cutoff for ideation systems
ideation_prompt	The 30–50 word research-direction description given to ideation systems
n_papers, citation_stats	Cohort size and citation distribution
papers	paper_id, title, date, date_remapped, citations, research_problem, task_formulation, application_domain

ideation_inputs.json holds the same cohort_id / ideation_prompt / cutoff_date triples in the format the baseline runner reads.

human_proposals.parquet (1,330 rows)

paper_id (arXiv id), title, cohort_ids, proposal (Markdown).

baseline_proposals.parquet (1,274 rows)

Proposals from the ideation baselines in baselines/, one row per (run, cohort):

run_group	Asset set	Systems	Notes
blog_table	simple	Opus 4.6, DeepSeek-V4-Pro, Kimi-K2.6, GLM-5.1 (Claude Code); GPT-5.4 (Codex)	The runs behind the win-rate table in the blog post
annotation_pool	simple	Opus 4.6, Kimi-K2.6, GLM-5.1 (Claude Code)	Re-runs used as a source of AI proposals for human-vs-AI annotation
hypothesis_generator	scientific-hypothesis-generator	Opus 4.6, Opus 4.8, DeepSeek-V4-Pro, GLM-5.2 (Claude Code); GPT-5.4, GPT-5.5 (Codex)	Later, stronger harness with skills, a reviewer subagent and artifact gates

Other fields: system, harness, model (as sent to the provider), cohort_id, cutoff_date, status, valid (passed the structure check), created_at, total_cost_usd, total_tokens, duration_ms, proposal. Two annotation-pool runs have no proposal.

lit_reviews/<cohort_id>/

The evaluator's literature-review workspace for each cohort, built only from pre-cutoff sources: literature_review.md, key_findings.md, coverage_gaps.md, research_plan.md and paper_summaries/. The reviews were written by GPT-5.4 and the paper summaries by DeepSeek-V4-Flash. The scraped full texts of the reviewed papers are not included.

Known limitations

• Hindsight. Proposals reconstructed from finished papers inherit the paper's structure and choices, so they are more coherent and more conservative than proposals written before the experiments.
• Cohort = same K_t is an approximation. Papers in a cohort share a topic and a time window, not necessarily an identical research question.
• Contamination. The window postdates the GPT-5.4 and Opus 4.6 knowledge cutoffs. Newer models may have seen these papers.
• Converter and judge share a model family. Proposals were produced by Claude Opus 4.6, which is also one of the judge models used in the blog post.

License and takedown

Proposals are LLM-generated rewrites of arXiv papers. They are released under CC BY-NC 4.0 for research use; the underlying papers keep their own licenses and are not redistributed here. If you are an author and want a proposal derived from your paper removed, please open an issue at https://github.com/02hyc/IdeaEval.

Citation

@misc{huang2026ideaeval,
  title  = {IdeaEval: Impact-Oriented Evaluation of Scientific Hypothesis Generation},
  author = {Huang, Yanchen and Niu, Yuchen and Hu, Xiangkun},
  year   = {2026},
  url    = {https://02hyc.github.io/Blog/posts/ideaeval/}
}