Skip to content

Benchmark

ExplorationBench · AlienLogic · M4 · Best@3 held-out accuracy

Scores are percentages. Higher is better.

Models and systems measured
10
Scores recorded
10
Sources
1
IQ weight
0.51

Score distribution

74.3%
Mean
76.2%
Median
5.8%
Std dev
69%
10th percentile
79.5%
90th percentile
59.5%
Min
59.5%81%
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 0.51 for chat models, measured over 6 models. The weight is how much it counts toward llminfo IQ. We measure it from the data, not from the raw scores.

Leaderboard

#Model or systemScore
1Claude Opus 5official 2026-10-13 leaderboard snapshot · 70 held-out tasks in this world · 4 exploration rounds · 3 answers/task · max reasoning · 3 independent autonomous trajectories· Reported by Fudan University / Tencent Hy / Tsinghua University, Tencent Hy / ExplorationBench authors· via BenchBoard docs79.5%
2GPT-5.6 Solofficial 2026-10-13 leaderboard snapshot · 70 held-out tasks in this world · 4 exploration rounds · 3 answers/task · max reasoning · 3 independent autonomous trajectories· Reported by Fudan University / Tencent Hy / Tsinghua University, Tencent Hy / ExplorationBench authors· via BenchBoard docs76.7%
3Hy4 previewofficial 2026-10-13 leaderboard snapshot · 70 held-out tasks in this world · 4 exploration rounds · 3 answers/task · high reasoning · 3 independent autonomous trajectories· Reported by Fudan University / Tencent Hy / Tsinghua University, Tencent Hy / ExplorationBench authors· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.76.2%
4Kimi K3official 2026-10-13 leaderboard snapshot · 70 held-out tasks in this world · 4 exploration rounds · 3 answers/task · max reasoning · 3 independent autonomous trajectories· Reported by Fudan University / Tencent Hy / Tsinghua University, Tencent Hy / ExplorationBench authors· via BenchBoard docs74.3%
5DeepSeek V4.1 Flashofficial 2026-10-13 leaderboard snapshot · 70 held-out tasks in this world · 4 exploration rounds · 3 answers/task · max reasoning · 3 independent autonomous trajectories· Reported by Fudan University / Tencent Hy / Tsinghua University, Tencent Hy / ExplorationBench authors· via BenchBoard docs73.8%
6Gemini 3.8 Flashofficial 2026-10-13 leaderboard snapshot · 70 held-out tasks in this world · 4 exploration rounds · 3 answers/task · high reasoning · 3 independent autonomous trajectories· Reported by Fudan University / Tencent Hy / Tsinghua University, Tencent Hy / ExplorationBench authors· via BenchBoard docs59.5%

Not ranked

The source marks these results as descriptive, not capability scores, so we list them by name instead of ranking them.

Model or systemScore
DeepSeek V4 Pro (version unspecified)official 2026-10-13 leaderboard snapshot · 70 held-out tasks in this world · 4 exploration rounds · 3 answers/task · max reasoning · 3 independent autonomous trajectories· Reported by Fudan University / Tencent Hy / Tsinghua University, Tencent Hy / ExplorationBench authors· via BenchBoard docs75.2%
Grok 4.6official 2026-10-13 leaderboard snapshot · 70 held-out tasks in this world · 4 exploration rounds · 3 answers/task · xhigh reasoning · 3 independent autonomous trajectories· Reported by Fudan University / Tencent Hy / Tsinghua University, Tencent Hy / ExplorationBench authors· via BenchBoard docs81%
Qwen3.8 Max 0902official 2026-10-13 leaderboard snapshot · 70 held-out tasks in this world · 4 exploration rounds · 3 answers/task · max reasoning · 3 independent autonomous trajectories· Reported by Fudan University / Tencent Hy / Tsinghua University, Tencent Hy / ExplorationBench authors· via BenchBoard docs77.6%
Seed2.1 Pro 0915official 2026-10-13 leaderboard snapshot · 70 held-out tasks in this world · 4 exploration rounds · 3 answers/task · high reasoning · 3 independent autonomous trajectories· Reported by Fudan University / Tencent Hy / Tsinghua University, Tencent Hy / ExplorationBench authors· via BenchBoard docs69%

Data credits: BenchBoard by SYuan03 (github.com/SYuan03/benchboard). MIT covers its code and data organization; no results licence is inferred. Numeric relays retain the original stating-source links.