Right Answer, Wrong Method - Derivation-Adjusted Accuracy: leaderboard
Right Answer, Wrong Method (arXiv 2608.02442v1, Table 4): 3,528 model solutions to competition (IMO-Bench, PHYBench, SciOlympiad) and HLE math, physics and chemistry problems, audited for shortcut hacking such as formula or answer guessing. Share of problems solved correctly without a flagged hack, percent; the detector under-flags, so values are upper bounds. Higher is better.
Metric: Correct and not flagged as hacked (%; detector-audited). Source: arxiv.org. Saturation forecast: Around February 2027. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Kimi K3 | 41.7 | #35 |
| 2 | Gemini 3.1 Pro (Preview) | 40.4 | #64 |
| 3 | GPT-5.5 | 38 | #33 |
| 4 | Qwen 3.7 Max | 37.5 | #85 |
| 5 | DeepSeek V4 Pro | 34.8 | #112 |
| 6 | Claude Opus 4.7 | 33.1 | #51 |
| 7 | GLM-5.2 | 30.1 | #80 |
| 8 | Claude Opus 4.8 | 26 | #39 |
| 9 | GPT-5.2 | 15.8 | #115 |
| 10 | DeepSeek V3.2 | 11 | #210 |
| 11 | GPT-4.1 | 5.7 | #258 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=right-answer-wrong-method-derivation-adjusted-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-05.