NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·3h ago
AI Research
arXiv CS.AI·May 25·8 min read
EVE-Agent: Evidence-Verifiable Self-Evolving Agents
TL;DR
Self-evolving agents that generate their own training data risk learning from fluent but unsupported answers, creating unreliable feedback loops. EVE-Agent adds an evidence verifier to the proposer-solver framework, requiring each generated answer include a source-grounded text span whose contribution is measurable.
✦ Why It Matters
Engineers can build self-improving systems with auditable, evidence-backed training data without manual annotation overhead.
Key Takeaways
- Adopt evidence verifiability in AI training to improve trustworthiness.
- Utilize the proposer-solver model for generating questions and answers.
- Incorporate evidence spans to support AI-generated responses.
- Evaluate training signals based on the accuracy contribution of evidence.
How It Works
EVE-Agent operates by having a proposer generate a question, an answer, and a verbatim evidence span. An evidence verifier assesses the span's contribution to the answer's accuracy, creating a feedback loop that rewards useful evidence.
This approach ensures that the self-evolving agent learns from reliable sources, enhancing the overall quality of its training data.