FIRE-Bench - Cross-Domain: leaderboard
The cross-domain extension of FIRE-Bench (Wang et al., 2026): rediscovering the findings of five computer-vision and vision-language papers and five neural-network-analysis papers end to end. Claim-level F1 in percent against the source findings with a GPT-5.2 entailment judge, averaged over the ten tasks and three runs; higher is better.
Metric: Claim-level F1 (%; atomic-claim entailment against the source findings with a GPT-5.2 judge; mean over the 10 cross-domain tasks (five CV/VLM papers, five neural-network-analysis papers) and three runs). Source: arxiv.org. Saturation forecast: Around January 2028. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 (Medium) | 30.2 | #163 (GPT-5) |
| 2 | O4 Mini | 23.4 | #354 |
| 3 | GPT-5 | 13.3 | #163 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=fire-bench-cross-domain · How It Works · Data refreshed daily, snapshot 2026-11-12.