SLR-Bench - Hard: leaderboard
The Hard tier of SLR-Bench (AIML lab, TU Darmstadt) — the most complex tier of its scalable, symbolically verified first-order rule-induction puzzles; accuracy (%).
Metric: Accuracy (%). Source: huggingface.co. Saturation forecast: Around January 2027. 56 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash (Preview) | 54 | #82 |
| 2 | GPT-5 | 50 | #98 |
| 3 | GPT-5 Mini (High) | 44 | #186 (GPT-5 Mini) |
| 4 | GPT-5.4 Mini | 28 | #212 |
| 5 | GPT-5 Mini (Medium) | 23 | #186 (GPT-5 Mini) |
| 6 | GLM-4.7 FP8 | 23 | #209 |
| 7 | Gemini 2.5 Flash | 18 | #245 |
| 8 | Gemini 3.1 Flash Lite (High) | 18 | #191 (Gemini 3.1 Flash Lite) |
| 9 | Gemini 2.5 Pro | 16 | #152 |
| 10 | GPT-OSS-120B (High) | 15 | #341 (GPT-OSS-120B) |
| 11 | O4 Mini | 14 | #180 |
| 12 | GPT-5 Chat | 14 | #241 |
| 13 | Gemma 4 31B | 12 | #200 |
| 14 | Qwen 3.5 35B A3B | 10 | #256 |
| 15 | O3 Mini | 10 | #272 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=slr-bench-hard · How It Works · Data refreshed daily, snapshot 2026-10-20.