Verdict-Sufficient Evidence Pruning for Faithful Fact-Checking Explanations

When a language model explains why a claim is true or false, it usually writes from a pile of retrieved documents. Most of that pile is irrelevant to the verdict, and the model often cites the wrong passage for a sentence or cites none at all. This proposal changes what the pruner keeps. Current tools that trim retrieved text before generation keep whatever looks relevant to the question. We instead train a fast pruner to keep only the smallest set of evidence sentences that is still enough to reach the verdict. That set can be evidence that supports the claim, or evidence that refutes it. The explanation is then written only from those sentences, one cited sentence per kept piece of evidence. The bet is that removing everything not needed for the verdict removes most of the chances to mis-attribute, and that a model trained to find sufficiency, not relevance, can do this at the speed of an ordinary context pruner.

The need is documented. When GPT-4 wrote fact-checking explanations from machine-retrieved evidence, only 9% of them were fully and accurately cited, compared with 31% when it was given gold evidence (Evaluating Evidence Attribution in Generated Fact Checking Explanations). Users still rated these explanations as helpful, which makes the failure worse. The gap between 9% and 31% is the cost of noisy evidence, and a pruner that reliably recovers something close to gold-like minimal evidence should close much of it.

Core Claim

The claim is that the training label matters more than the pruning architecture. We take a pruner of the same form as Provence, which is a small encoder that scores each retrieved sentence keep or drop in one forward pass (Provence: efficient and robust context pruning for retrieval-augmented generation). If it is trained on verdict-sufficiency labels instead of Provence's "sentences an LLM cited when answering" labels, we expect it to produce explanations with much higher attribution accuracy at the same token budget. We also expect it to handle refuted and not-enough-information claims correctly, which existing minimal-evidence methods cannot.

We define sufficiency with local minimality, so that the definition can actually be checked. Let $c$ be a claim, $S$ a set of evidence sentences, and $V(c,S)\in\{\text{Supported},\text{Refuted},\text{NEI}\}$ the verdict of a fixed verifier model. The set $S$ is a locally minimal sufficient group for verdict $y$ if

$$V(c,S)=y \quad\text{and}\quad V(c,S\setminus\{s\})\neq y \ \text{ for every } s\in S.$$

Checking this takes $|S|+1$ verifier calls. This sidesteps the NP-hardness that Minimal Evidence Group Identification for Claim Verification proves for globally minimal groups. For $y=\text{NEI}$ the minimal group is empty, so the generator is forced to say the evidence is insufficient. That directly targets the failure in e-FEVER, where few-shot GPT-3 invented justifications for not-enough-information claims.

Relation to Prior Work

Three lines of work supply parts of the method but leave the combination open.

The minimal-evidence-group work supplies the search procedure. It predicts full or partial support for each sentence and merges partial supporters, pruning the search with set-theoretic rules. It is offline and slow, and its authors explicitly leave contradictory evidence to future work. We reuse its search to generate training labels and extend it to refutation.

The refutation extension is shaped by LOREN (LOREN: Logic-Regularized Reasoning for Interpretable Fact Verification). LOREN uses an asymmetric rule: a claim is refuted if any one part is refuted, but supported only if every part is. Under this rule, a minimal refuting group only needs to defeat one culprit part plus its bridging facts. Refuting groups should therefore usually be smaller and easier to find than supporting groups, which is a prediction we test.

Provence and its relatives supply the fast inference form (AttentionRAG, XProvence). Their objective is token efficiency and answer accuracy; none of them measures attribution in the explanation.

Two further papers constrain the design. The attribution study motivates it by showing that the evidence gap exists. Decomposition Dilemmas shows that aggressive claim decomposition adds context loss and retrieval noise for capable verifiers. For that reason we do not decompose the claim before retrieval; the culprit structure is found only inside the label search.

Research Program

The work has three parts.

First, we build labels. On the training splits of FEVER (FEVER), HoVer (HoVer), and WiCE, we retrieve evidence and run the extended minimal-group search with an open verifier. The search looks for supporting groups on supported claims and refuting groups on refuted claims, and every group must pass the local-minimality check above. One complication is that refutation is not always monotone: a newly added sentence can reconcile a conflict, as in Retrieval-Augmented Generation with Conflicting Evidence. We therefore record how often a superset of a refuting group stops refuting, and we exclude those claims from training while reporting them separately.

Second, we train the pruner. A DeBERTa-sized encoder with Provence's layout reads the claim together with the retrieved sentences. It outputs keep/drop decisions for each sentence and a three-way verdict head, trained jointly. The joint training follows the finding in Generating Fact Checking Explanations that training evidence extraction alongside veracity helps both tasks.

Third, we constrain generation. The generator LLM receives only the kept sentences, each with its own identifier. It must write one sentence per kept sentence, citing it, and it may not add uncited sentences.

Baselines

The naive reference is the setup Xing et al. measured. GPT-4, or an open model of comparable size, receives the full retrieved evidence and is asked for a verdict and a cited explanation. The reported number to beat is 9% fully and accurately cited explanations with machine-retrieved evidence on PolitiHop, with 31% under gold evidence as a practical ceiling (Evaluating Evidence Attribution in Generated Fact Checking Explanations).

The strongest existing input pruner is Provence, run off the shelf and then with the same generation step. On attribution the file gives no scalar accuracy; it reports only 50–80% token reduction and a 2x speedup. So no reported result was found on Lacuna for Provence on attribution. We run it ourselves at matched kept-token budgets.

The strongest fact-checking pipeline with an explicit redundancy filter is BiDeV (BiDeV). Its reported numbers are relative: an average 3.88% macro-F1 gain across HoVer and Feverous-s, and 17.72% on 4-hop claims over fine-tuned baselines. We rerun it with the same retriever and generator and compare against its absolute scores.

The offline minimal-group search itself, run at inference, is the quality upper bound for our distilled pruner. Its paper reports only relative precision gains, 18.4% on WiCE-MEG and 34.8% on SciFact-MEG, not absolute values.

The key ablation retrains our exact encoder on Provence-style relevance labels, using the same data, steps, and hyperparameter search budget. This isolates the label definition.

All systems share one retriever, one generator, and one tuning budget, and are compared at matched average kept tokens. Beating a baseline means a statistically significant gain in fully accurate citation rate, with verdict macro-F1 no more than one point lower.

Why not just…?

Why not just prune with Provence and prompt for citations? Provence keeps whatever an answering LLM cited, which favors topical relevance. For a fact-check, that includes background sentences that are true and relevant but not needed for the verdict, and these are exactly the sentences a generator mis-cites. The relevance-label ablation tests this directly. If it matches our pruner at equal token budget, the minimality idea adds nothing, and we will report that.

Why not just generate from full evidence and drop explanation sentences that fail an entailment check? This is the most serious threat. Post-hoc filtering can fix false citations, but it cannot restore a verdict that was driven by a distractor. It also leaves in sentences that are correctly cited but unnecessary. We run this filter with the same verifier we use for labels. We then compare citation accuracy, verdict F1, and explanation length. If the filter matches us on all three, our contribution shrinks to speed and to NEI handling.

Why not just ask the LLM to select evidence identifiers first, then explain? Select-then-explain prompting costs nothing to try, and frontier models may do it well. We include it as a baseline. Our counter-argument rests on its behavior under noisy retrieval, since the 9% figure is itself a frontier model working from noisy evidence. We also compare cost: our pruner is a single encoder pass, while prompted selection is a long-context LLM call.

Proposed Experiments and Decisive Checks

We evaluate on four test sets. PolitiHop tests attribution out of domain. HoVer 2–4 hop tests multi-hop claims; its 81.2% oracle accuracy against 14.9% for its baseline shows how much evidence quality matters. WiCE-MEG and SciFact-MEG test minimality precision against human annotations. FEVER refuted and NEI subsets test counter-evidence and abstention.

Attribution is measured with automated citation masking and recovery, and humans annotate a 300-explanation subset to validate the automated measure. We also report verdict macro-F1, kept tokens, and latency.

Three results would decide the question. The first is whether the sufficiency-trained pruner beats the relevance-trained ablation on fully accurate citation rate at matched tokens. The second is whether it narrows the gap between 9% and 31% on PolitiHop by at least half. The third is whether its NEI outputs abstain rather than invent support, where e-FEVER shows the naive approach fails. A secondary check tests the LOREN-derived prediction that refuting groups are smaller than supporting groups. If they are not, the asymmetric extension needs rethinking.