Experience-based learning allows large language model (LLM) agents to improve over time by reusing knowledge from past executions. However, as a new arXiv abstract notes, existing methods typically rely on final task outcomes as feedback for experience. This reliance on outcome alone may miss useful information about how the task was performed.
The paper, titled "Beyond Outcome Feedback: Trajectory-Quality-Guided Experience-based Learning," appears to address this limitation by proposing a trajectory-quality-guided approach. The abstract is truncated, so full details of the method are not available in the source, but the title and opening indicate a shift from outcome-only feedback to considering the quality of the entire trajectory.
Given the single source and its partial text, this digest can only report the stated motivation and the proposed direction. No experimental results or specific mechanisms are described in the available excerpt, so those are not covered here.