EVE-Agent: Evidence-Verifiable Self-Evolving Search Agents
The EVE-Agent, a novel AI framework, introduces the concept of evidence verifiability as essential for reliable self-evolution in search agents. It alters the traditional proposer-solver model: the proposer creates a question, provides an answer, and includes a direct evidence excerpt, while an evidence verifier incentivizes contributions based on credible sources. This design mitigates the risk of self-generated curricula rewarding articulate yet unsubstantiated examples, tackling the challenges posed by unclear and unreliable training signals in agents that evolve independently without data.
Key facts
- EVE-Agent stands for Evidence-Verifiable Self-Evolving Agent.
- The framework is introduced in a paper on arXiv with ID 2605.22905.
- It operates within a proposer-solver framework.
- The proposer generates a question, an answer, and a verbatim evidence span.
- An evidence verifier measures the contribution of the evidence span to the answer.
- The goal is to ensure each generated instance includes source-grounded evidence.
- This prevents the training loop from rewarding unsupported examples.
- The approach is designed for data-free self-evolving search agents.
Entities
Institutions
- arXiv