Researchers introduced Chain-of-Experience (CoE), a test-time learning approach that allows large language models to improve iteratively using self-feedback and environmental signals instead of relying solely on zero-shot inference. Evaluated across eight models on math, coding, and knowledge tasks, the method delivered substantial gains from self-feedback alone, plus a 5.6% overall accuracy improvement and 19% lower API cost when combining multiple feedback channels. The technique achieved higher accuracy per token than existing test-time strategies while staying robust to weak or misleading feedback signals.
