We introduce Hero's Journey, a benchmark for rule induction in goal-directed episodic tasks, where an agent must infer hidden rules from demonstrations and act on them through multi-step execution. Across eight tasks in two families, each with four structural rule forms, controllable lexical grounding, and identifiability conditions, the agent must:

• Induce the hidden rule connecting an entity's attributes to the requirement it imposes (e.g., which item to buy, which action to take at what order).
• Execute the inferred rule as an ordered sequence of dependent actions.

We evaluate models along two dimensions:

• ECSR (efficiency-calibrated success rate): whether the model succeeds efficiently
• RV (rule verbalization): whether the model can explicitly articulate the rule it acted on.

We present our results in an interactive leaderboard. Use the Task Explorer to walk through real demonstration episodes and try a held-out entity yourself, or the Leaderboard to compare models across all eight tasks.

@misc{zheng2026herosjourneytestingcomplex,
      title={HERO'S JOURNEY: Testing Complex Rule Induction with Text Games},
      author={Anshun Asher Zheng and Kanishka Misra and David I. Beaver and Junyi Jessy Li},
      year={2026},
      eprint={2606.02556},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2606.02556},
}