AdaExploreBench benchmarks LLM agents' proactive exploration
An arXiv paper posted on 19 October introduces AdaExploreBench, a benchmark from Rui Wang and 17 co-authors for proactive exploration and test-time adaptation in evolving environments. The benchmark encodes environment dynamics as hidden rules agents must actively gather evidence for and infer, and under implicit rule drift the latent rule changes without announcement, forcing agents to notice stale knowledge from degraded performance and re-explore. Across experiments on eight leading LLMs, the authors report that proactive, adaptive exploration remains difficult: GPT-6 Sol reaches only 61.9% calibrated accuracy, while supplying sufficient evidence or removing rule drift raises average performance by close to 10%.
The authors frame this as a critical bottleneck: agents can often reason well once informative evidence is available, but struggle to acquire the right evidence and to recognize when earlier evidence is no longer valid.
Computer Science > Computation and Language
arXiv:2610.25950 (cs)
[Submitted on 19 Oct 2026]
Title:AdaExploreBench: Benchmarking LLM Agents for Proactive Exploration and Test-time Adaptation in Evolving Environments
Authors:Rui Wang, Zishen Lai, Jiayu Liu, Cheng Qian, Jeonghwan Kim, Heng Wang, Xuejun Zhang, Jiateng Liu, Bingxuan Li, Shujin Wu, Qing Zong, Tianshi Zheng, Chunyang Li, Qihan Lin, Chen Li, Dadi Guo, Xiusi Chen, Yangqiu Song
View a PDF of the paper titled AdaExploreBench: Benchmarking LLM Agents for Proactive Exploration and Test-time Adaptation in Evolving Environments, by Rui Wang and 17 other authors