TIDE (Evolving Documents, Retrieval): leaderboard

Metric: Macro-averaged strict accuracy (%; mean over the eight task types; TIDE, 3,050 expert-verified questions over 644 Bangladesh customs instruments (1969-2025); free-form answers are judged by a council of Claude Opus 4.6, Gemini 3.5 Flash and GPT-5.1 by majority vote with a hard date gate, deterministic tasks by exact match; strict accuracy without the partial-rubric credit; thinking enabled, temperature 0; retrieval: passages retrieved from the instrument corpus are supplied). Source: arxiv.org. Saturation forecast: Around 2029. 9 models tracked.

Top models

#ModelScore
1Claude Sonnet 4.5 (Thinking)58.05
2GPT-5.2 (Thinking)57.06
3Kimi K2.5 (Thinking)54.46
4Gemini 3.5 Flash52.87
5GPT-OSS-120B52.69
6Qwen 3.6 35B A3B52.13
7Gemma 4 31B (Reasoning)47.69
8GLM-4.645.43

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