TCSExam-Lite - Theorem Proving: leaderboard
TCSExam (Moonshot AI) evaluates model-agent systems on verifiable theoretical computer science research: Lean proofs of results from recent TCS papers, reusable solvers judged on hidden instances, explicit constructions verified in Lean, and open conjectures. Unlike PutnamBench or miniF2F, which formalize competition problems, its theorems come from research papers and its solver and construction tracks are scored against published best-known results. Agent harness, version and reasoning setting per model follow the paper's Table 9 (arXiv 2611.07962v1). Question access: task statements and checkers are public on GitHub; solver judging uses hidden instances. Lite suite: 20 of the Full tasks, five per track; Offline Solver protocol with public self-tests during development and one final hidden-instance evaluation. Share of theorem-proving attempts that produce a complete Lean proof accepted by the kernel and axiom checks (%).
Metric: Complete Lean proof success (%). Source: moonshotai.github.io. Saturation forecast: Rough model projection: around 2026. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5.5 (Benchmark Sweep Run: source=TCSExam; suite=lite; harness=Claude Code 2.1.283; 1M-token context option; effort=max) | 80 | |
| 2 | GPT-6 Astra (Benchmark Sweep Run: source=TCSExam; suite=lite; harness=Codex 0.154.0; effort=max) | 60 | |
| 3 | Kimi K3.1 Pro (Benchmark Sweep Run: source=TCSExam; suite=lite; harness=Kimi Code 0.43.0; effort=max) | 40 | |
| 4 | Kimi K3 (Benchmark Sweep Run: source=TCSExam; suite=lite; harness=Kimi Code 0.43.0; effort=max) | 33.33 | |
| 5 | GLM-5.3 (Benchmark Sweep Run: source=TCSExam; suite=lite; harness=Claude Code 2.1.257; effort=max) | 20 | |
| 6 | Qwen 3.8 Max (0902) (Benchmark Sweep Run: source=TCSExam; suite=lite; harness=Claude Code 2.1.257; effort=max) | 20 | |
| 7 | Grok 4.6 (Benchmark Sweep Run: source=TCSExam; suite=lite; harness=Grok Build 1.0.41; thinking on; effort=xhigh) | 20 | |
| 8 | DeepSeek V4.1 Flash (1017) (Benchmark Sweep Run: source=TCSExam; suite=lite; harness=Claude Code 2.1.257; effort=max) | 6.67 |
Interactive version: theaggregate.ai/benchmark?slug=tcsexam-lite-theorem-proving · How It Works · Data refreshed daily, snapshot 2026-11-12.