SeekerGym - Wikipedia: leaderboard

SeekerGym (2026) measures how completely an information-seeking agent recovers the information goals of a hidden source document through search queries over its passages. This board is completeness in percent over 120 long Wikipedia articles, counting deduplicated retrieved passages, over three seeds; higher is better.

Metric: Information-gathering completeness (%): share of a document's information goals recovered by the agent's deduplicated query trajectory, 120 long Wikipedia articles, 3 seeds; higher is better. Source: arxiv.org. Saturation forecast: Around 2030. 6 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1GPT-OSS-20B42.5#1183
2Qwen 3 235B A22B 2507 (Thinking)41#466
3Nemotron 3 Nano40.5#2025
4GPT-OSS-120B38.5#704
5Gemini 2.0 Flash37.9#723
6GPT-4o Mini28#1667

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=seekergym-wikipedia · How It Works · Data refreshed daily, snapshot 2026-11-21.