Photonic PCFBench - Closed Loop Delta Proxy: leaderboard
PCFBench v1 Appendix full results; 26 isolated tasks across six categories; temperature0.1, task-specific64–2048token budgets; native fractions/errors. Paired structure+field and Field-only contexts separate. Closed-loop success uses expert sign/magnitude constraints, not physical resimulation; Human Score mean of six dimensions/two experts on120samples. Source-disputed summary/appendix cells and unbound checkpoint variants withheld.
Metric: Delta proxy. Source: arxiv.org. Saturation forecast: Not forecast. 30 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Flash Lite | 0.04 | #432 |
| 2 | GPT-4o | 0.04 | #357 |
| 3 | Qwen 3 VL 235B A22B Instruct | 0.05 | #280 |
| 4 | Qwen 2.5 VL 7B Instruct | 0.05 | #650 |
| 5 | GPT-4o Mini | 0.05 | #598 |
| 6 | Qwen 3.5 122B A10B | 0.06 | #184 |
| 7 | Qwen 2.5 VL 72B Instruct | 0.06 | #383 |
| 8 | GPT-4.1 Mini | 0.06 | #363 |
| 9 | Qwen 3 VL 30B A3B Instruct | 0.06 | #390 |
| 10 | Qwen 3.5 27B | 0.07 | #192 |
| 11 | Qwen 3 VL 8B Instruct | 0.07 | #419 |
| 12 | Qwen 3.5 9B | 0.08 | #391 |
| 13 | Grok 4.1 | 0.08 | #239 |
| 14 | Qwen 2.5 VL 32B Instruct | 0.08 | #463 |
| 15 | Grok 4 | 0.09 | #189 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=photonic-pcfbench-closed-loop-delta-proxy · How It Works · Data refreshed daily, snapshot 2026-11-05.