Reproduce “Investigating Advanced Reasoning of Large Language Models via Black-Box Environment Interaction” on OpenResearch

About this paper

Congchi Yin, Tianyi Wu, Yankai Shu, Alex Gu, Yunhan Wang, Jun Shao, Xun Jiang, Piji Li

Researchers introduced a "black-box environment interaction" paradigm and the ORACLE benchmark to evaluate Large Language Models' (LLMs) integrated reasoning in genuinely unknown, interactive settings. Experiments across 19 LLMs revealed a critical limitation in their high-level planning for adaptive exploration strategies, with top models achieving significantly lower accuracy than human participants and often falling below 40-50% on harder tasks.

Key insights

• Modern LLMs exhibit a fundamental and universal weakness in "high-level planning capability to develop efficient and adaptive exploration strategies," struggling to effectively utilize real-time feedback for hypothesis refinement.
• LLMs frequently employ sub-optimal, non-adaptive exploration behaviors (e.g., exhaustive, sequential querying) instead of information-theoretically efficient strategies, particularly in tasks with less predictable rules.
• While "extended thinking" capabilities generally improve LLM performance, explicit reflective instructions or increased exploration turns offer negligible gains in tasks demanding deep strategic adaptation, indicating a core limitation.

Summary by alphaXiv.