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

Hallucination remains a critical reliability bottleneck for LLM-based agents, and fact attribution verifiers serve as the last line of defense. Yet current verifiers produce opaque binary labels, revealing nothing about which evidence was considered or why the judgment was made — leaving agents unable to correct themselves and operators unable to audit decisions. We introduce SEVA (Self-Evolving Verification Agent), which outputs structured evidence alignments, step-by-step reasoning chains, calibrated confidence, and error diagnoses from a six-category taxonomy. To train SEVA beyond SFT, we design a process reward function that scores five verification components independently, allocating 70% of weight to process signals (evidence grounding, reasoning quality, format compliance) and 30% to outcome (label, diagnosis). This design is not optional: we show that standard binary reward suffers from advantage collapse on structured output, where GRPO receives no gradient signal and training stagnates entirely. Process reward resolves this and induces an implicit curriculum — the agent first masters verification behavior (alignment: 0.917→0.997; format: 72%→100%), then improves verification outcomes (F1: 64.9→69.0). Structured output further enables a self-evolution loop where the agent's own errors — now transparent and diagnosable — drive targeted adversarial data generation for iterative improvement. On ClearFacts, SEVA-3B matches GPT-4o-mini (69.0 vs. 69.8 F1) while producing substantially richer, auditable output.