Learning to Learn-at-Test-Time: Language Agents with Learnable Adaptation Policies

arXiv cs.AI · 18 hours ago

Researchers introduced Meta-TTL, a framework that learns optimal adaptation policies for language agents using test-time learning rather than relying on hand-crafted policies. The approach uses bi-level optimization with evolutionary search across training tasks, and was evaluated on three benchmarks (Jericho, WebArena-Lite, and τ²-bench). Meta-TTL outperformed baseline methods in both in-distribution and out-of-distribution settings, showing that learned adaptation policies generalize to new task distributions.