Better with Experience: Self-Evolving LLM Agents for Evidence-Grounded Health Community Notes
Title: Leveraging Experience: Self-Improving LLM Agents for Evidence-Based Health Community Notes
Original: arXiv:2606.02215v1 Announce Type: new
Abstract: Large Language Model (LLM)-enhanced Community Notes present a scalable solution for the prompt, evidence-based correction of health-related misinformation on social media platforms. Nevertheless, current systems typically reset their context with each new post, failing to leverage valuable correction insights gained from previous instances. To address this limitation, we present EvoNote, an agentic framework designed to facilitate the self-evolution of health Community Notes generation. This is achieved through an evolving experience memory that retains knowledge from prior misinformation correction episodes. The framework’s foundation lies in fine-grained credit assignment, where EvoNote links trajectory-level feedback to specific health-note qualities, subsequently distilling these insights into action-level memory. This memory aids in claim analysis, evidence gathering, and note composition.
We assessed EvoNote using MM-HealthCN, a multimodal benchmark comprising 1,200 instances of user-flagged health posts, accompanied by human-written Community Notes and crowd-sourced helpfulness ratings. Utilizing a human-validated hierarchical utility judge, we found that notes generated by EvoNote were favored over their human-written counterparts in 89.6% of cases. Furthermore, on a distinct dataset of "Needs More Ratings" posts lacking crowd verdicts, EvoNote successfully produced helpful notes for 82.0% of cases. The system also dramatically accelerated the correction process, reducing the median time to generate a candidate correction from over 13 hours in the human-note pipeline to less than 2 minutes. Our analyses attribute these improvements to more robust evidence utilization and the reuse of effective correction strategies, suggesting that self-evolving note generation represents a promising approach for managing health misinformation.
Source: arXiv Generated at: 2026-06-02 00:00:00 UTC
