ACR Human Audit - Soft Correctness: leaderboard
The ACR study uses human reviewers to assess the correctness and variety of mathematical solutions from two unmodified models, with four candidate solutions per problem. Per-trajectory soft-correctness (%).
Metric: Per-trajectory soft-correctness (%). Source: arxiv.org. Saturation forecast: Around January 2027. 2 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemma 4 26B A4B | 100 | #353 |
| 2 | Qwen 3.6 35B A3B (Non-reasoning) | 98.3 | #233 (Qwen 3.6 35B A3B) |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=acr-human-audit-soft-correctness · How It Works · Data refreshed daily, snapshot 2026-10-29.