Evidence-First Risk-Controlled Judging Under Noisy Retrieval

LLM judges are increasingly asked to decide whether an answer is factually correct by reading retrieved web text. The judge here is a language model used to grade other models' outputs. The most principled version of this setup routes each case through three options: the judge answers from its own knowledge when confident, retrieves evidence when not, and hands the case to a human otherwise. It picks confidence thresholds so that the fraction of wrong accepted verdicts provably stays below a target. That guarantee rests on a quiet assumption: that retrieved evidence makes the judge more correct as it makes the judge more confident. Work on retrieval-augmented readers shows the opposite can happen. Irrelevant or misleading passages distract the model and can make it confidently wrong.

This note proposes to measure how badly distracting evidence erodes risk-controlled judging, and to fix it with a small decoupled module borrowed from grammatical error correction. A separate, cheap sentence tagger first locates the few sentences that actually bear on the claim. The judge then rules only on those sentences, and the tagger's own confidence is folded into the score used to set the abstention thresholds. In two sentences: distraction makes a judge's confidence a worse signal of its correctness, so a risk-controlled judge must abstain far more often to keep its error guarantee. Separating "find the evidence" from "rule on it," and calibrating on both signals, should recover most of that lost coverage.

Core Claim

Judge, Retrieve, or Abstain calibrates entropy thresholds with Clopper-Pearson bounds under conformal risk control, a statistical recipe for turning a held-out calibration set into a guaranteed error bound. It keeps the false discovery rate (FDR, the share of accepted verdicts that are wrong) below a target $\alpha$ in all 32 configurations it tests. It raises coverage, the share of cases the judge decides rather than abstaining, from 7% to 82% on NQ-Open with Qwen3-8B at $\alpha=0.20$.

That paper explicitly assumes retrieved snippets are clean evidence. Improving Zero-shot Reader by Reducing Distractions from Irrelevant Documents shows readers degrade as more retrieved documents are added, and that they assign very high probability to distracted wrong answers. It motivates the concern.

We need to be precise about what breaks. If calibration and deployment share the same retrieval noise, the conformal guarantee still holds, because it only needs the two sets to be exchangeable. The damage then shows up as lost coverage: confident-wrong verdicts force the thresholds down. The FDR guarantee itself is violated only when the noise level at deployment exceeds the level seen at calibration, which is the normal situation when a retriever, corpus or query mix drifts.

Our claim has three parts. First, both failure modes are large under realistic distractors. Second, a decoupled evidence tagger plus a two-signal calibration score recovers coverage at fixed $\alpha$. Third, the same design shrinks FDR violations under calibration–deployment noise shift. The third part is the one that matters most in practice.

Research Program

The method has three stages, patterned on the detect-then-edit split of Chen et al.. In that grammatical error correction system, an encoder tagger flags erroneous spans and a decoder rewrites only those spans. The split roughly tripled decoding efficiency and slightly improved $F_{0.5}$ on BEA-2019.

In our version, an encoder cross-encoder takes the question, the candidate answer, and each retrieved sentence. It labels the sentence as supporting, refuting, or irrelevant, in the spirit of the atom-level scoring of Stacey et al., which scores isolated facts with a DeBERTa model so surrounding narrative cannot sway the decision. We train it on HotpotQA supporting-fact annotations and answer-bearing NQ sentences, and hold out TriviaQA and PopQA as unseen domains.

The judge then sees only the top-scoring sentences. If none clears a minimal support score, the case is routed to parametric judging or abstention. This mirrors the sentence-level gatekeeper of Asano et al., which blocks inputs from the generative stage and gained over 3 precision points.

Finally, the calibration score combines the two signals. For item $x$, let $H(x)$ be the judge's predictive entropy on the filtered evidence. Let $e(x)\in[0,1]$ be the tagger's maximum support-or-refute probability. We define

$$s(x) = H(x) - \lambda \log e(x),$$

where $\lambda \ge 0$ is chosen on a split disjoint from the calibration set. The verdict is accepted when $s(x) < t$, with $t$ set by the same Clopper-Pearson procedure used in Judge, Retrieve, or Abstain. Setting $\lambda=0$ recovers entropy on filtered text, which is the ablation that separates filtering from scoring.

Adding a term that does not change exchangeability cannot break the guarantee. It can only help or hurt the ranking of correct verdicts above wrong ones, and therefore coverage.

The noise model connects to the judge-bias literature. Justice or Prejudice? (CALM) perturbs the candidate response with fake citations, authoritative tone and padding while keeping facts fixed. It finds GPT-4o especially vulnerable to authority and verbosity bias, but never perturbs the judge's evidence. We inject three kinds of distractors into the retrieved context:

  • random irrelevant passages;
  • topically similar hard negatives that do not contain the answer;
  • CALM-style authority-dressed passages that assert the wrong answer.

The distractor count is swept at 0, 5, 25 and 100. The 25 level matches the chunk count in Ext2Gen, and 100 matches Cho et al.'s hardest setting.

The first measurement is diagnostic. For each noise type and level, we compute how well entropy ranks verdict correctness (area under the ROC curve), the coverage achievable at $\alpha\in\{0.1,0.2\}$, and the empirical FDR when thresholds calibrated at one noise level are applied at another.

Baselines

Parametric-only judging. The naive reference is the judge using its own knowledge with entropy thresholds and no retrieval. It is reported at 7% coverage on NQ-Open with Qwen3-8B at $\alpha=0.20$ in Judge, Retrieve, or Abstain. Any retrieval method must stay well above this floor at every noise level. Cases where retrieval falls below it under heavy distraction would be the clearest evidence of harm.

Judge, Retrieve, or Abstain, unmodified. This is the strongest existing method. It is the only one in the file with provable FDR control for LLM judges, and it reports 82% coverage on NQ-Open (Qwen3-8B, $\alpha=0.20$) and 85% on HotpotQA (Llama-8B), up from 40%. We rerun it with the same judges (Qwen3-8B, Llama-3.1-8B), the same candidate answers, the same calibration and test splits, and the same noisy retrieval pools. First we check that on clean retrieval our reimplementation lands near those numbers. Beating it means higher coverage at equal target $\alpha$ with empirical FDR still at or below $\alpha$ when noise is matched. It also means a strictly smaller FDR overshoot under noise shift, on at least three of the four datasets at 25 or more distractors.

Reranker truncation. A cross-encoder reranker keeps only the top-$k$ passages ($k\in\{1,3\}$), and entropy calibration runs on those. This uses the same encoder backbone and training data as our tagger, so any gain we show comes from sentence-level extraction and the two-signal score rather than from a better retrieval model.

Prompted in-judge extraction. The judge is instructed to quote the relevant sentences first and then rule, which imitates Ext2Gen without its preference training. Ext2Gen reports an average score of 0.557 vs. 0.473 for the default Llama-3.1-8B-Instruct under noisy chunks, beating external compressors. That makes it the most credible argument that extraction belongs inside the model, and it directly answers the file's open tension between decoupled and in-generation extraction.

Distraction-aware score adjustment. This baseline follows DAS from Cho et al. It adds the abstention instruction and reweights evidence by query-generation likelihood. The entropy calibration is kept.

Two ablations isolate the key ingredient. Setting $\lambda=0$ gives filtering alone. Filtering is removed while $\log e(x)$ is kept in the score, which gives the scoring signal alone. All methods share one tuning budget: the same held-out split for choosing $k$, $\lambda$ and prompt variants, and no tuning on the calibration or test sets.

Why not just…?

Why not just calibrate on noisy data and trust the guarantee? This is correct as far as it goes, and we say so. Matched-noise calibration keeps FDR bounded. It does not keep the system useful, though, because confident-wrong verdicts collapse the achievable coverage. It also offers no protection once deployment noise exceeds calibration noise. The shift experiment measures both effects. If matched-noise coverage turns out to barely drop and shifted FDR barely rises, the problem is minor and the proposal should be abandoned rather than rescued.

Why not just retrieve fewer, better passages? This is the most serious threat. A strong reranker might remove most distractors, and Cho et al. show fewer documents help readers. Reranker truncation is therefore a first-class baseline with a matched backbone. Our expectation is that it fails on authority-dressed and hard-negative distractors. Those are topically ideal passages, so a relevance reranker ranks them highly, while a tagger that scores support for the specific candidate answer can label them as refuting or irrelevant. If truncation matches our method on those distractor types, the contribution reduces to "use a reranker," and we will report it that way.

Why not just ask the judge to quote evidence first? Ext2Gen is evidence that in-model extraction beats external compressors for answer generation. But the file flags that in-generation extractors can themselves be distracted when noise resembles evidence, and CALM shows judges are specifically susceptible to authority cues. A judge extracting its own evidence may preferentially quote the authoritative-sounding wrong passage. The prompted-extraction baseline, broken down by distractor type, settles this directly.

Why not just prompt the judge to ignore irrelevant text? RBD found that a modular critic outperformed in-context prompting by 12.8% in evaluation accuracy. CALM found chain-of-thought improved consistency but left GPT-4o exposed to authority and verbosity bias. We include the ignore-irrelevant-text instruction inside the DAS-style baseline rather than as a separate claim.

Proposed Experiments

We use the four datasets on which Judge, Retrieve, or Abstain was evaluated: NQ-Open, TriviaQA, HotpotQA and PopQA. Candidate answers come from a fixed small QA model, and verdict correctness is determined against gold answers. Distractors are drawn from the same retrieval corpus (random and hard-negative), plus authority-dressed passages generated with CALM's perturbation style but placed in the evidence.

The primary metrics are three: empirical FDR against target $\alpha$ with binomial intervals, coverage at $\alpha\in\{0.1, 0.2\}$, and area under the ROC curve of each score for verdict correctness. Secondary metrics are the tagger's sentence-level precision and recall on HotpotQA supporting facts, and wall-clock cost per item. The tagger should add encoder cost only, whereas prompted extraction lengthens decoding.

Shift check (decisive for safety). We calibrate at 5 distractors and test at 25 and 100, for every method. Success means the unmodified entropy method exceeds $\alpha$ by a clear margin while ours stays within the binomial interval of $\alpha$ or overshoots substantially less.

Coverage check (decisive for usefulness). With matched noise at 25 distractors, our method must beat both Judge, Retrieve, or Abstain and reranker truncation in coverage at equal $\alpha$ on the authority and hard-negative conditions.

Ingredient check. The two ablations must show that the two-signal score adds coverage beyond filtering alone. Otherwise the calibration contribution is empty.

Generalization check. The tagger is trained without PopQA and TriviaQA. The gains must persist on those two datasets, which are long-tail and single-hop respectively.

A negative result on the shift check, with matched-noise coverage intact, would still be informative. It would tell practitioners that conformal judging under noisy retrieval needs noise-matched calibration, not a new architecture.