A research team has built a verification agent that does more than flag bad facts — it explains why they're wrong and suggests fixes.
Current fact-attribution verifiers attached to large language models tend to spit out a binary pass/fail with no explanation, making it nearly impossible for downstream agents to self-correct or for operators to audit decisions. SEVA changes that by emitting structured output: evidence alignments, step-by-step reasoning, calibrated confidence scores, and a six-category error diagnosis. Training it with reinforcement learning required a custom fix — standard binary reward caused what the authors call "advantage collapse," where gradient signals vanish entirely. Their solution was a process reward that breaks verification quality into five weighted components, restoring the training signal and letting the model learn behavior before outcomes.
The benchmark numbers are the interesting part. A SEVA-3B model matches GPT-4o-mini on the ClearFacts dataset (69.0 vs. 69.8 F1) while producing far more auditable output — a meaningful result if it holds outside controlled conditions. More surprising is what happened during a four-round self-evolution experiment on a 7B model: each round produced a benchmark specialist, not a generalist. The model gained 15 percentage points on HaluEval while losing 10 to 14 points on TruthfulQA in the same run, and the effect persisted even at four times the data volume.
That specialization finding is the paper's sharpest edge. The field has long debated whether small, task-specific models can reliably replace frontier-scale ones for verification work — SEVA's results suggest the answer is "sometimes, for specific benchmarks," which is a more honest answer than most product launches will give you.