EVE-Agent: Evidence-Verifiable Self-Evolving Agents
The paper introduces EVE-Agent, a self-evolving agent designed to enhance the reliability of AI-generated answers. It emphasizes the importance of evidence verifiability in training self-evolving agents, ensuring that each generated answer is supported by a measurable source. The proposed framework aims to improve the correctness of evidence-grounded responses without relying on human annotations or external labels.
- ▪EVE-Agent operationalizes evidence verifiability in self-evolving search agents.
- ▪The agent generates questions, answers, and supporting evidence spans for each response.
- ▪Experiments demonstrate that EVE-Agent significantly improves the correctness of evidence-grounded answers.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2605.22905 (cs) [Submitted on 21 May 2026] Title:EVE-Agent: Evidence-Verifiable Self-Evolving Agents Authors:Yamato Arai, Yuma Ichikawa View a PDF of the paper titled EVE-Agent: Evidence-Verifiable Self-Evolving Agents, by Yamato Arai and 1 other authors View PDF HTML (experimental) Abstract:Self-evolving agents should not train on examples they cannot justify. Data-free self-evolving search agents offer a scalable route to systems that generate their own questions, answer them, and improve from their own feedback without human annotations. Yet, without verifiable evidence, this loop can reward fluent but unsupported examples, turning the self-generated curriculum into an opaque and potentially unreliable training signal.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
