FireWorldBench - Counterfactual Intervention Reasoning (Structured Records): leaderboard

Metric: Completion accuracy (%; P5, counterfactual intervention reasoning items of the controlled-simulation subset; structured-observation track (textual sensor traces and physical records); each cell is the mean of the matched choice-format score (Jaccard overlap of selected and gold option sets) and open-report score (exact agreement of required structured slots); deterministic evaluator, no LLM judge; failed or unparseable responses score 0). Source: arxiv.org. Saturation forecast: Around 2029. 15 models tracked.

Top models

#ModelScore
1GPT-5.6 Luna48.07
2Gemma 3 27B (IT)46.85
3Qwen 3.5 Plus (2026-02-15)44.92
4DeepSeek V3.244.45
5Gemini 2.5 Flash40.41
6Qwen 3 32B39.61
7Kimi K238
8Qwen 3 235B A22B36.47
9Qwen 3 Next 80B A3B Instruct33.95
10Qwen 3 8B31.58
11Claude Haiku 4.531.45
12GLM 4.5 Air29.56
13Llama 3.1 8B Instruct29.38
14Gemma 3 12B (IT)27.89
15Llama 3.3 70B Instruct27.8

Interactive version: theaggregate.ai/benchmark?slug=fireworldbench-counterfactual-intervention-reasoning-structured-records · How It Works · Data refreshed daily, snapshot 2026-09-26.