MemCalib - Exact Calibration: leaderboard
The exact-calibration view of MemCalib (2026), a memory-use benchmark for LLMs. Exact calibration is the share of responses with zero memory over-use and zero under-use across all atoms, judged by a DeepSeek-V4-Pro rubric over three non-thinking seeds, in percent, higher is better.
Metric: Exact calibration (%; share of responses with zero memory over-use and zero under-use across all atoms, DeepSeek-V4-Pro rubric judge, mean over 3 seeds; non-thinking mode). Source: arxiv.org. Saturation forecast: Around 2029. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Sol (Non-reasoning) | 28.4 | #29 (GPT-5.6 Sol) |
| 2 | Kimi K2.6 (Non-reasoning) | 20 | #173 (Kimi K2.6) |
| 3 | Claude Sonnet 4.6 | 19.09 | #158 |
| 4 | GLM-5.2 (Non-reasoning) | 18.2 | #137 (GLM-5.2) |
| 5 | DeepSeek V4 Flash (Non-reasoning) | 16.96 | #254 (DeepSeek V4 Flash) |
| 6 | Qwen 3 8B (Non-reasoning) | 15.29 | #1838 (Qwen 3 8B) |
| 7 | Qwen 3.5 35B A3B (Non-reasoning) | 12.24 | #480 (Qwen 3.5 35B A3B) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=memcalib-exact-calibration · How It Works · Data refreshed daily, snapshot 2026-11-12.