ExplorationBench: Measuring AI Systems' Exploration in Verifiable Alien Worlds

              

📄 Paper  |   💻 Code  |   🖥️ Website & Leaderboard  |   📝 Blog

---

This dataset holds the two public demo worlds of ExplorationBench. They are built with the same procedure as the evaluation worlds and use the same language, manual, proof format, protocol and task families, but their rules are new and differ from every rule of the evaluation worlds. The evaluation worlds themselves are kept private, because a world whose rules are public measures recall instead of exploration. The evaluation code, which runs these demo worlds end to end, is on GitHub.

🔥 News

• 2026.10.28: 🤗 The two public demo worlds are released as a dataset on Hugging Face. [Dataset]
• 2026.10.20: 📰 Synced (机器之心) covers ExplorationBench. [Article]
• 2026.10.13: 📣 Tencent AI features ExplorationBench on X. [Post]
• 2026.10.08: 💻 This repository goes public with the evaluation code and the two public demo worlds.
• 2026.09.30: 📣 Tencent Hy announces ExplorationBench on X. [Post]
• 2026.09.24: 🎉 The paper is released on arXiv — two verifiable Alien Worlds with 55 discovery targets and 140 held-out tasks, and results for 10 frontier AI systems. [Paper] [Website] [Blog]

📦 What Is in This Dataset

Config	File	Rows	Content
aliencode_heldout_tasks	aliencode/heldout_tasks.jsonl	5	Held-out tasks with expected outputs and reference solutions
aliencode_worked_examples	aliencode/worked_examples.jsonl	4	The fixed worked examples a system reads first, with their real outputs and what the manual predicts
aliencode_discovery_targets	aliencode/discovery_targets.jsonl	5	The hidden rule changes: what the manual says, what actually happens, and the S-expression answer key
alienlogic_heldout_theorems	alienlogic/heldout_theorems.jsonl	5	Held-out theorems, including an unprovable one, with reference proofs and the standard-calculus proofs this world refuses
alienlogic_worked_examples	alienlogic/worked_examples.jsonl	5	The fixed worked proofs, accepted and refused, with the proof checker's verdicts
alienlogic_side_conditions	alienlogic/side_conditions.jsonl	3	The patched side conditions of the demo proof system

aliencode/manual.md is the flawed manual a system starts from, and examples/ shows both demo worlds as a system sees them. The answer keys (discovery_targets, reference solutions and proofs, side conditions) are for checking results and are never shown to a system during evaluation. Task prompts are in Chinese, the language of the evaluation prompts.

To evaluate a model on the demo worlds, use the harness in the GitHub repository. Demo-world scores are for development only and are not comparable to the leaderboard. To have a model evaluated on the private evaluation worlds, see Evaluate Your Model.

📄 Overview

Scientific discovery begins where known problems end. There, AI systems must explore: frame hypotheses, design experiments, and iterate on the results. ExplorationBench measures this ability in verifiable Alien Worlds. Their rules are executable, so every answer is checked exactly, and they conflict with familiar knowledge, so recall alone cannot solve the tasks.

A system starts from a flawed manual and a fixed set of worked examples. Over four rounds it probes the world through a single tool, and the world answers with deterministic feedback and no explanation. After every round, a copy of the conversation with tools disabled reports the rules it believes hold and answers every held-out task three times. The copy is then discarded, so test items never leak into exploration.

The evaluation protocol. Every system starts from the same flawed manual and the same fixed worked examples and chooses its own probes over four rounds. At every milestone M_t, a tool-disabled copy reports the rules it believes hold and answers the held-out tasks.

Held-out accuracy over autonomous exploration. Top: AlienCode; bottom: AlienLogic. Each cell highlights one system's M₀–M₄ curve; the other nine are grey. Each curve is the system's Best@3 trajectory, and each milestone is the mean of three answers per held-out task.

Stats: 2 sandboxes · 55 discovery targets · 140 held-out tasks · 10 frontier systems · 3 independent trajectories per system and sandbox · every verdict from an interpreter or a proof checker, with no LLM judge

Sandboxes:

• AlienCode — program synthesis in a small calculation language whose familiar-looking operators follow hidden semantics: SHATTER multiplies, EMIT(100) prints 127, and no primitive adds. 31 discovery targets and 70 held-out tasks; each program is graded by an interpreter on five private inputs.
• AlienLogic — formal proof in a first-order natural-deduction system with 24 patched inference rules. 70 held-out theorems, 23 of them unprovable; a proof checker verifies every proof, and an unprovable theorem scores only when the system declines to prove it.

🏆 Leaderboard

Held-out accuracy (%) after four rounds of autonomous exploration (M₄). Best@3 is the best of three independent trajectories and Mean@3 their mean; M₀ is the Best@3 trajectory's accuracy after the worked examples alone. Every score is the mean of three answers per task, and every system runs at its highest reasoning setting. More columns, the control conditions, and interactive trajectories are on the website.

AlienCode

Rank	System	M0	Best@3	Mean@3
1	Claude Opus 5	3.8	89.0	73.5
2	GPT-5.6 Sol	11.0	88.1	79.5
3	Gemini 3.8 Flash	1.4	79.0	39.7
4	Kimi K3	2.4	79.0	38.7
5	Grok 4.6	4.8	69.0	53.2
6	Qwen3.8-Max	2.4	65.7	64.0
7	Hy4 preview	0.5	63.3	30.2
8	DeepSeek-V4.1-Flash	1.4	59.5	46.3
9	Seed2.1 Pro	2.4	40.0	24.8
10	DeepSeek-V4-Pro	1.9	12.9	7.3

AlienLogic

Rank	System	M0	Best@3	Mean@3
1	Grok 4.6	42.9	81.0	72.1
2	Claude Opus 5	43.3	79.5	75.9
3	Qwen3.8-Max	45.2	77.6	74.3
4	GPT-5.6 Sol	51.0	76.7	75.1
5	Hy4 preview	35.2	76.2	69.2
6	DeepSeek-V4-Pro	33.8	75.2	66.3
7	Kimi K3	45.2	74.3	69.4
8	DeepSeek-V4.1-Flash	51.4	73.8	60.6
9	Seed2.1 Pro	37.6	69.0	59.8
10	Gemini 3.8 Flash	45.7	59.5	59.0

💡 Key Findings

• Exploration, not recall or thinking alone, produces the knowledge the tasks require. No AlienCode trajectory exceeds 15.7% after the worked examples; after four rounds the best reaches 89.0%. The same number of rounds without environment feedback stays at 0.5–11.0%.
• Exploration works best when the system designs its own experiments. Replaying a system's own best probes to it gives it exactly the same evidence, yet in AlienCode autonomous exploration still does better for 9 of 10 systems, by a median of 17.4 points.
• The two worlds get stuck in different places. In AlienLogic, being told the rules yields 96–100%, above every system's best trajectory, so discovery is the bottleneck. In AlienCode, the best trajectory beats being told the rules up front for 7 of 10 systems, so using the rules is.
• Knowing a rule does not guarantee using it correctly. Tasks whose required rules the final rule report states correctly are still solved only 73.4% of the time.
• A single score hides a lot. One system's trajectories under the same budget can end tens of points apart (Kimi K3: 5.7–79.0% in AlienCode), a system's rank in one world barely predicts its rank in the other (Spearman 0.35), and gains arrive in leaps that later rounds can undo.

📜 License

The data and code are licensed under the Apache-2.0 License.

📝 Citation

@article{zhang2026explorationbench,
  title   = {ExplorationBench: Measuring AI Systems' Exploration
             in Verifiable Alien Worlds},
  author  = {Zhang, Ming and Xiang, Zhenghao and
             Gao, Peizhong and Shen, Yujiong and Wang, Yuhui and
             Yue, Zhonghan and Dou, Shihan and Yin, Zhangyue and
             Ye, Junjie and Liu, Shichun and Zheng, Weihuang and
             Chen, Jiahao and Chen, Jiayi and Liu, Hongzhang and
             Shao, Jiaqi and Gui, Tao and Zhang, Qi and
             Huang, Xuanjing and Zheng, Suncong and Pan, Maxm},
  journal = {arXiv preprint arXiv:2609.30199},
  year    = {2026}
}

_{Synced from GitHub commit 8892fd0.}