Right Answer, Wrong Method - Final-Answer 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. Final-answer accuracy before the hacking audit, percent; Kimi-K3 and GLM-5.2 gave no response on 7.2% of items and DeepSeek-V4-Pro on 22.7%. Higher is better.
Metric: Final-answer accuracy (%). Source: arxiv.org. Saturation forecast: Around February 2027. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 48.1 | #33 |
| 2 | Kimi K3 | 46.9 | #35 |
| 3 | DeepSeek V4 Pro | 45.7 | #112 |
| 4 | Qwen 3.7 Max | 45.6 | #85 |
| 5 | Gemini 3.1 Pro (Preview) | 44 | #64 |
| 6 | Claude Opus 4.7 | 42 | #51 |
| 7 | GLM-5.2 | 37.4 | #80 |
| 8 | Claude Opus 4.8 | 34 | #39 |
| 9 | GPT-5.2 | 19.6 | #115 |
| 10 | DeepSeek V3.2 | 15.5 | #210 |
| 11 | GPT-4.1 | 10.1 | #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-final-answer-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-05.