AdaExploreBench: Benchmarking LLM Agents for Proactive Exploration and Test-time Adaptation in Evolving Environments

AdaExploreBench. Benchmarking LLM Agents for Proactive Exploration and Test-time Adaptation in Evolving Environments. By Rui Wang and seventeen others, from the Hong Kong University of Science and Technology, the Hong Kong University of Science and Technology, Guangzhou, and four other institutions. Published on the nineteenth of October. Its code repository is worked out: about one hundred and thirty source files, with a readme.

Large language model agents are being deployed more and more into environments that are unfamiliar to them and that keep changing. Working there requires an agent to explore on its own initiative, to understand how the environment behaves, and to adapt as that behaviour shifts. Existing benchmarks rarely test this. They seldom ask whether an agent can discover for itself how an environment works, or whether it notices and adjusts when the environment evolves.

AdaExploreBench targets exactly those two abilities. The first is proactive exploration: whether an agent, dropped into an environment it does not understand, goes looking for the evidence it needs rather than waiting for it. The second is adaptation at test time: whether, once the environment evolves, the agent revises what it believes instead of carrying on with knowledge that no longer applies. Both are about the agent's own initiative, and that is what existing benchmarks rarely test.

The dynamics of each environment are instantiated as hidden rules. The agent is not told them. It has to go out and gather evidence, and then infer the rules from what it gathered.

On top of that sits implicit rule drift. At some point the latent rule changes, and nothing announces it. The only sign is that performance gets worse. The agent has to read that degradation as a signal that what it learned earlier is now stale, and go back to exploring.

Eight leading language models were tested, and proactive, adaptive exploration remains hard. GPT-6 Sol, as one example, reaches only sixty-one point nine percent calibrated accuracy.

Two control conditions locate the difficulty. Giving the agent sufficient evidence up front, or removing the rule drift, each raises average performance significantly, by close to ten percent.

Those controls draw a sharp line through agent ability. Handed informative evidence, current agents can often reason from it well. What they struggle with is going out to acquire the right evidence in the first place, and recognising when evidence they already hold has stopped being true. The bottleneck, on this benchmark, is not inference but inquiry.