SEVA: Self-Evolving Verification Agent with Process Reward for Fact Attribution

2026-06-28

AI Summary

Large language model agents often suffer from hallucinations due to the absence of auditable and self-correcting fact-checking mechanisms. This work proposes SEVA, a Structured Verification Agent that employs a self-evolving Verify-Reflect-Probe-Refine loop, integrating process-reward-driven reinforcement learning (GRPO) with multi-granularity reward decomposition to jointly optimize reasoning chains, evidence alignment, confidence estimation, and error diagnosis—effectively mitigating dominance collapse. Experimental results show that SEVA-3B achieves an F1 score of 69.0 on ClearFacts, rivaling GPT-4o-mini, while attaining an evidence alignment accuracy of 0.997 and 100% format compliance. Moreover, the self-evolution process drives the model toward specialization rather than generalization.

Abstract

Hallucination is the reliability bottleneck for LLM-based agents, and fact attribution verifiers are the last line of defense -- yet today's verifiers emit only opaque binary labels, leaving agents unable to self-correct and operators unable to audit. We present SEVA, a structured verification agent that emits evidence alignments, step-by-step reasoning chains, calibrated confidence, and a six-category error diagnosis with actionable fixes. Training such an agent with RL is non-trivial: standard binary reward on multi-component output triggers advantage collapse -- within-group reward variance vanishes and the GRPO gradient disappears. We resolve this with a process reward that decomposes verification quality into five independent components weighted 70/30 toward process signals, restoring the gradient and inducing an implicit curriculum -- the agent first masters verification behavior (alignment 0.917 -> 0.997, format 72% -> 100%), then outcomes (F1 64.9 -> 69.0). Structured output further enables a Verify -> Reflect -> Probe -> Refine self-evolution loop, which over four rounds on a 7B model surfaces an unexpected structural finding: each round produces a benchmark-specialist, not a generalist (+15 pp on HaluEval, -10 to -14 pp on TruthfulQA in the same model, persistent at 4x data). On ClearFacts, SEVA-3B matches GPT-4o-mini (69.0 vs. 69.8 F1) while producing substantially richer, auditable output -- confirming a principle that should generalize: for any RL task with multi-component generation, reward granularity must match output granularity.

Problem

• hallucination
• fact attribution
• verification
• auditability
• self-correction

Innovation

• process reward
• structured verification
• self-evolving agent
• fact attribution
• reward granularity