FormalTCS - Theorem Elicitation: leaderboard
FormalTCS (2026) benchmarks LLMs on frontier theoretical computer science research over 143 instances from 2025-2026 conference papers, each annotated in natural language and in Lean 4. LLM-rubric score for writing a natural-language theorem from the core claim, in percent; higher is better.
Metric: LLM-rubric score (0-100) from a Qwen3.8-Max judge in QoderCLI, a model outside the evaluated families, weighted 0.4 logical validity, 0.3 completeness, 0.2 correctness and 0.1 clarity; one deterministic generation; natural-language theorem written from the core claim; 143 instances from theoretical computer science papers accepted at STOC, FOCS, SODA, COLT and similar venues in 2025-2026, each stage given the human-annotated input; each model runs in its own vendor harness. Source: arxiv.org. Saturation forecast: Around June 2027. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Sol | 67.4 | #29 |
| 2 | Claude Opus 5 | 66.9 | #20 |
| 3 | Claude Sonnet 5 | 63 | #108 |
| 4 | GPT-5.6 Terra | 60.7 | #60 |
| 5 | GPT-5.6 Luna | 56.4 | #117 |
| 6 | Claude Haiku 4.5 | 48.7 | #568 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=formaltcs-theorem-elicitation · How It Works · Data refreshed daily, snapshot 2026-11-12.