QCalEval - Parameter Extraction: leaderboard
Metric: Score (0-100) for family-specific physical parameters extracted as JSON, per-field tolerance scoring, failed parses scored 0, on the 243 quantum-calibration plots of QCalEval (22 experiment families, superconducting qubits and neutral atoms), zero-shot with the family background text, greedy decoding; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 18 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 71.5 |
| 2 | Gemma 4 31B (IT) | 68.3 |
| 3 | Claude Opus 4.6 | 64.7 |
| 4 | GPT-5.4 | 64.3 |
| 5 | Gemini 3.1 Flash Lite | 63.8 |
| 6 | GPT-5.4 Mini | 62.6 |
| 7 | Qwen 3.5 397B A17B | 62.5 |
| 8 | Qwen 3.5 122B A10B | 61.2 |
| 9 | Claude Sonnet 4.6 | 60.4 |
| 10 | Qwen 3.5 27B | 58.7 |
| 11 | Qwen 3.5 35B A3B | 57.8 |
| 12 | Qwen 3.5 9B | 57.1 |
| 13 | InternVL3-78B | 52.9 |
| 14 | Claude Haiku 4.5 | 51 |
| 15 | InternVL3-38B | 49.2 |
Interactive version: theaggregate.ai/benchmark?slug=qcaleval-parameter-extraction · How It Works · Data refreshed daily, snapshot 2026-10-07.