ITPEval - Babel-formal Statements: leaderboard
ITPEval measures formal translation between interactive theorem provers: given a source file in Lean 4, Isabelle, Rocq or HOL Light, the model emits the same content in a target prover and a unified verifier checks it there. This board covers statement translation on the controlled Babel-formal suite, which avoids library dependence. The score is the share of directed source-to-target pairs that verify from a single greedy attempt.
Metric: Pass@1 (%; share of the 192 directed Babel-formal (tier A, controlled) statement pairs whose single translated statement verifies; zero-shot direct translation at temperature 0: given a source file in Lean 4, Isabelle, Rocq or HOL Light and a target prover, emit raw target code, which must verify under the target prover). Source: arxiv.org. Saturation forecast: Around February 2027. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 47.4 | #49 |
| 2 | Gemini 3.1 Pro (Preview) | 32.81 | #94 |
| 3 | DeepSeek V4 Pro | 17.19 | #188 |
| 4 | Claude Sonnet 4.6 | 10.94 | #161 |
| 5 | Qwen 3 235B A22B | 9.38 | #623 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=itpeval-babel-formal-statements · How It Works · Data refreshed daily, snapshot 2026-11-13.