Decocted Experience Improves Test-Time Inference in LLM Agents

Researchers present a new approach to improve Large Language Model performance without updating model parameters by using 'decocted experience' - extracting and organizing key insights from previous interactions to guide better reasoning. The method shows effectiveness across reasoning tasks including math, web browsing, and software engineering by constructing better contextual inputs rather than simply scaling computational resources.

• →New method improves LLM performance without parameter updates by leveraging contextual experience rather than just computational scaling.
• →Decocted experience involves extracting essence from past interactions and organizing it coherently for future reasoning tasks.
• →The approach addresses the cost inefficiency of naive test-time compute scaling by providing better guidance for exploration.
• →Validation across multiple domains including math reasoning, web browsing, and software engineering shows broad applicability.
• →Context construction quality is more critical than simply increasing inference-time computation for complex agentic tasks.