Reconstruction (Research Idea Recovery): leaderboard
Reconstruction (AI-Professor Project, 2026) is a blind idea-recovery benchmark: from only the pre-publication bibliography of a published paper, a model proposes five hypotheses, and an LLM judge decides which match the withheld paper idea. Match rate in percent, the unweighted mean of six domain scores over 643 papers, each scored by the six other evaluated models as leave-one-out judges; higher is better.
Metric: Match rate (%; unweighted mean of the six domain scores over 643 seed papers; share of a model's five hypotheses, proposed in one call from a published paper's pre-publication bibliography only (titles and abstracts under anonymous IDs, the paper itself and later literature withheld), that an LLM judge matches to the held-out paper's idea; mean over the six other evaluated models as leave-one-out judges). Source: arxiv.org. Saturation forecast: Around 2029. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.8 | 13.3 | #56 |
| 2 | GPT-5.6 Pro Sol | 12.8 | #25 |
| 3 | Kimi K3 | 10 | #53 |
| 4 | GLM-5.2 | 9.3 | #137 |
| 5 | Gemini 3.1 Pro (Preview) | 8.9 | #91 |
| 6 | DeepSeek V4 Pro | 6.3 | #181 |
| 7 | Qwen 3.7 Max | 5.9 | #142 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=reconstruction-research-idea-recovery · How It Works · Data refreshed daily, snapshot 2026-11-12.