Verdicts Without Annotated Evidence: Rejection Sampling or Label-Only Post-Training for Evidence Recovery?
Read the original on arXiv AI →The Flow has not summarised this story yet — read it at arXiv AI.
The Flow has not summarised this story yet — read it at arXiv AI.
The paper investigates when a language‑model judge can truly ground its verdicts in code correctness. It shows that current multi‑agent verification methods rely on evidence that is both independent of the answer and distinct between candidates—conditions that fail in code judging. By analyzing two label‑free measurements from the judge’s logs, the authors demonstrate that gating on one measurement allows the system to decline uncertain comparisons, improving accuracy from 20.7% to 36.9% while still answering half of all cases.
MiniVer‑V is a new benchmark for short‑video fact‑checking that contains 195 videos, each annotated with a supported, refuted, or insufficient verdict and 5,510 multimodal evidence units (visual keyframes, speech transcripts, and web‑retrieved sources). The authors propose a two‑layer verification framework that first checks claim‑video consistency using internal evidence, then determines the factual verdict with external corroboration, and employs a sufficiency‑driven greedy search to assemble only the minimal evidence needed for a confident decision. Using Claude Sonnet 4, the method achieves a Macro‑F1 of 0.510 with an average of 4.5 evidence units—about 16 % of the full set—while matching the full‑evidence baseline and improving detection of insufficient cases.
The paper proposes a training‑free detector that uses sentence‑level context sensitivity to identify unsupported content in retrieval‑augmented generation (RAG) answers. By re‑scoring each sentence with full context, no context, and each chunk removed, the method flags the chunk whose removal most reduces a sentence’s likelihood as the likely source. Evaluated on RAGTruth, TofuEval, and RAGBench, the detector outperforms answer‑level faithfulness scores, achieving AUCs up to 0.745 and matching per‑chunk fact‑checkers while using only a fraction of the compute required by large‑language‑model judges.
The paper investigates whether widely used evaluation frameworks for large language models (LLMs) implement a defense called commit‑first judging, which requires a judge to solve a task itself before accepting a candidate answer. Across 24 configurations in eight popular frameworks, none use the full commit‑first method; nine use a weaker variant that is ineffective. In controlled experiments, the weaker variant allowed systems to game the judge, while the full commit‑first approach eliminated this vulnerability but sometimes worsened evaluation when the judge’s own answer was incorrect.
The paper evaluates 21 LLM-as-a-Judge models from nine providers across MT-Bench, JudgeBench, and RewardBench, using agreement, consistency, and bias audit protocols over 118 runs and about 541,000 judgments. It finds that exact-match agreement overstates performance, with Cohen’s kappa lower by 33–41 percentage points, judge rankings shift up to 14 positions, and high test–retest reliability coexists with significant position bias in some production judges. Verbosity bias is minimal across the cohort. The authors propose a Minimum Viable Validation Protocol based on these findings.
arXiv:2609.26550v3 Announce Type: replace Abstract: LLM-as-a-judge scales evaluation, but reasoning judges are slow and costly. We study JEV-as-a-Judge: evaluation with JEV, a decision-only judge tha...