EVE-Agent: Evidence-Verifiable Self-Evolving Agents

Yamato Arai, Yuma Ichikawa

arXiv:2605.22905 · 2026-05-25 공개 · arXiv · PDF

self-evolving-agents search-agents auditability model-backbone retriever curriculum-generation training-signal source-grounded

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. We argue that evidence verifiability is a prerequisite for trustworthy self-evolution in search agents: each generated instance should include not only an answer but also a source-grounded span whose contribution to that answer can be measured. We introduce EVE-Agent, an Evidence-Verifiable Self-Evolving Agent that operationalizes this principle through a modification to the proposer--solver framework. The proposer generates a question, an answer, and a verbatim evidence span. An evidence verifier then rewards the span according to the marginal accuracy gain when the evidence is provided. This produces a training signal that favors evidence that genuinely helps answer the question, without requiring oracle answers, human labels, or external annotations. EVE-Agent leaves the backbone model, retriever, search tool, and optimization framework unchanged. Experiments show that EVE-Agent substantially improves evidence-grounded correctness over prior self-evolving search agents. The resulting curriculum is not merely self-generated but auditable by construction: each training example carries an inspectable source span that explains why it should be trusted.

한국어 요약

📋 한 줄 요약

**[Self-Evolving Search Agent / Evidence Verification]** EVE-Agent가 proposer가 question·answer·verbatim evidence span을 함께 생성하면, verifier가 evidence 제공 시 marginal accuracy gain으로 보상 — 오라클·인간 라벨 없이도 evidence-grounded correctness 대폭 향상.

🎯 핵심 기여도

💡 핵심 아이디어

Self-evolving 에이전트의 신뢰성은 evidence verifiability에 의해 결정되며, span의 marginal contribution을 측정해 보상하면 외부 라벨 없이도 auditable·trustworthy curriculum이 자연스럽게 형성된다.

🔬 기술적 접근법

📊 주요 결과

💭 의의 및 한계

**의의**: Self-evolving 학습의 truthfulness·auditability 문제 해결, evidence를 first-class 학습 신호로 격상, 검색 도구 의존 LLM 에이전트의 신뢰성 향상에 일반 적용 가능. **한계**: Verbatim span 의존 — paraphrastic·multi-hop evidence 표현 한계, marginal gain 측정의 계산 비용, search 도메인 외 적용 일반화 필요.

🚀 실용적 활용