LexKairos: leaderboard

Metric: Mean score (%; average of the nine sub-task scores, accuracy or F1, over 3,800 Chinese legal temporal items from real judicial cases and statutes; vanilla setting, the model answers directly without extra reasoning steps (thinking disabled where the model allows it); temperature 0, 8,000 output tokens). Source: arxiv.org. Saturation forecast: Around December 2026. 8 models tracked.

Top models

#ModelScore
1DeepSeek R179.78
2GPT-5.4 (Non-reasoning)77.47
3Gemini 3 Flash (Minimal)76.71
4DeepSeek V4 Flash (Non-reasoning)67.07
5Qwen 3.5 9B (Non-reasoning)55.45
6Llama 3.1 8B28.11

Interactive version: theaggregate.ai/benchmark?slug=lexkairos · How It Works · Data refreshed daily, snapshot 2026-09-29.