Overview
Paper: EVE-Agent: Evidence-Verifiable Self-Evolving Agents Authors: Yamato Arai, Yuma Ichikawa arXiv: 2605.22905 Field: AI/NLP
Key points
- The problem: Self-evolving agents based on the Proposer-Solver framework (where one agent generates questions and another answers them, both improving in a closed loop) lack external verification. Without it, the Solver can produce answers that are fluent and internally consistent but factually wrong, and these errors compound over iterations. As the paper states:
- Reward hacking: In self-evolving systems, agents may learn to game internal reward signals — e.g., optimizing for plausibility, length, or formatting rather than factual correctness. Because training data is self-generated, once "cheating" starts, it self-reinforces and pollutes the entire training set.
- Core principle of EVE-Agent: An agent should not train on examples it cannot verify.
- Three-part training example architecture: 1. Question — generated by the Proposer (unchanged from prior work). 2. Answer — generated by the Solver (unchanged). 3. Evidence Span — an exact, verbatim quote from a retrieved document that directly supports the answer (the key innovation).
- Evidence Verifier via marginal accuracy gain: The system measures answer accuracy *without* the evidence span (baseline) versus *with* the evidence span (assisted). The difference — the marginal accuracy gain — quantifies whether the evidence genuinely supports the answer. Gains above a threshold earn positive rewards for Proposer and Solver; low or negative gains are penalized. Policy updates can use standard RL methods (e.g., PPO or GRPO).
- No external supervision needed: EVE-Agent requires no human annotations, ground-truth answers, or external judge models. Verification is automatic, internal, and grounded in measurable contribution — analogous to a student validating hypotheses through experiments rather than a teacher's grading.
- Auditability: Unlike black-box self-generated curricula, every EVE-Agent training example carries an inspectable source span:
- Generality: The method does not modify the underlying model, retriever, search tools, or optimization framework — the evidence-verification layer can be added on top of any existing Proposer-Solver system regardless of LLM or retrieval backend.
- On benchmarks of evidence-grounded correctness, EVE-Agent significantly outperforms previous self-evolving search agents: its answers are not only correct but reliably source-backed, reducing hallucination and fabrication.
- The evidence-verification mechanism (marginal accuracy gain as reward) proves more effective than conventional answer-correctness rewards for evidence quality, answer accuracy, and training stability.
- Dr. Zero (Meta): Also zero-data self-evolving, but evaluation is based mainly on whether answers are "accepted"; EVE-Agent adds an evidence-grounding layer ensuring answers are not merely accepted but supported.
- Multi-Agent Evolve (MAE): Uses a Proposer-Solver-Judge triad, where the judge is a general evaluator susceptible to model bias; EVE-Agent's verifier is objective, based on marginal accuracy gain rather than subjective judgment.
- EvoEnv: Focuses on environment synthesis with verifiability tied to executability (e.g., code that runs); EVE-Agent's evidence-based verifiability applies to broader domains beyond code.
- Epistemology in code: EVE-Agent blends empirical grounding (answers must cite evidence spans) with pragmatic validation (evidence quality defined by its measurable usefulness) — an operational, executable "knowledge verification protocol."
- Anti-confirmation-bias design: Evidence quality is judged by whether it improves accuracy, not whether it matches expectations; bad evidence is penalized regardless of how plausible it seems.
- Applications: More trustworthy RAG systems (forced citation, verifiable retrieval use), Socratic AI tutors that teach verification rather than answers, and automated scientific literature reviewers that produce evidence-based research reports.
- Arai, Y., & Ichikawa, Y. (2026). EVE-Agent: Evidence-Verifiable Self-Evolving Agents. arXiv:2605.22905.
- Meta AI. (2026). Dr. Zero: A Zero-Data Self-Evolving Learning System.
- Chen, Y., et al. (2025). Multi-Agent Evolve: LLM Self-Improve through Co-evolution.
- Singh, A., et al. (2026). Learning to Build the Environment: Self-Evolving Reasoning RL via Verifiable Environment Synthesis. arXiv:2605.14392.
- Ouyang, L., et al. (2022). Training language models to follow instructions with human feedback. NeurIPS.
> "Without verifiable evidence, this loop can reward fluent but unsupported examples, turning the self-generated curriculum into an opaque and potentially unreliable training signal."
> "Each generated instance should include not only an answer but also a source-grounded span whose contribution to that answer can be measured."
> "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."
If errors emerge, one can trace which evidence span, question, or answer was at fault.