RECON (Long-Context Reasoning) - Counterfactual Reasoning: leaderboard
Metric: Score (-1 to 1; how alternative timelines would have unfolded; full 50K-100K-token case file in context; 1,414 questions after a closed-book contamination filter; abstention allowed, MCQ-single wrong answers -0.2, MCQ-multiple option-weighted in [-1, 1], ordering by Kendall tau, free-form by two cross-family LLM judges; temperature 0). Source: arxiv.org. Saturation forecast: Around 2029. 8 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 2.5 Flash | 0.31 |
| 2 | Gemini 2.5 Pro | 0.27 |
| 3 | GPT-4.1 Mini | 0.23 |
| 4 | Kimi K2.5 | 0.21 |
| 5 | GPT-5.1 | 0.21 |
| 6 | GPT-4o Mini (2024-07-18) | 0.19 |
| 7 | Llama 3.3 70B | 0.14 |
Interactive version: theaggregate.ai/benchmark?slug=recon-long-context-reasoning-counterfactual-reasoning · How It Works · Data refreshed daily, snapshot 2026-09-29.