Checking the Verdict, Not the Facts: Compact Evidence-Strength Verifiers for Medical Evidence Synthesis

When a language model summarizes a set of clinical trials, the most dangerous error is often not a made-up number. It is a wrong verdict: the summary says a treatment "reduces mortality" when the studies disagree, or it states as settled what the evidence only tentatively suggests. Clinician evaluations of ChatGPT on Cochrane reviews named this the "certainty illusion," turning tentative findings into absolutes (Evaluating large language models on medical evidence summarization). On MS^2, a large multi-document benchmark of medical systematic reviews, models agree with the review on the direction of the effect only about half the time (MS^2). Today's faithfulness checkers miss this kind of error by design. They split a summary into small claims and ask whether each one is supported by some passage of the source. A verdict claim can be supported by one trial and contradicted by the pooled evidence, so it passes every pairwise check while being wrong.

We propose to train a small, domain-specialized verifier, a 1.5B–8B model that scores outputs, that judges the conclusion instead of the individual facts. The summarizer first writes a compact evidence table: for each study, the population, intervention, comparator and outcome (PICO, the standard schema of evidence-based medicine), plus the effect direction and a hedging cue. The verifier reads this table and predicts the direction and certainty that the pooled evidence supports. It then checks whether the summary's stated verdict and its hedging match. We use this verifier in two places. During training it mines preference pairs for alignment. At inference it reranks a few candidate summaries. The bet is that a compact model can do expert-level evidence synthesis if the reasoning is split into an extraction step and a learned aggregation step, rather than being forced into a single pass over thousands of tokens.

Core Claim

Verdict-level errors in multi-document evidence summaries come from failing to aggregate evidence across documents, not from failing to ground individual sentences. They can be cut substantially by a compact verifier that predicts the supported verdict from a structured evidence table, and they cannot be cut as much by proposition-level checkers of any size. A second, falsifiable claim concerns model size. Self-critique for faithfulness reportedly fails at 1.5B parameters (Learning from Self Critique and Refinement for Faithful LLM Summarization). We expect that with an external specialized verifier and a structured intermediate, the size below which alignment stops helping drops from 7B to 1.5B or 3B.

Relation to Prior Work

Two papers motivate the gap. The Cochrane evaluation (Evaluating large language models on medical evidence summarization) catalogs certainty illusions and PICO scrambling. It also finds that feeding full abstracts instead of focused results made outputs worse, which suggests that long raw context hurts verdict formation. MS^2 (MS^2) shows that individual study conclusions match the review's conclusion only 53.8% of the time. That figure constrains our design. A hand-coded rule such as "majority of studies point the same way" would disagree with the expert review almost half the time. The aggregator must therefore be learned from the roughly 20,000 reviews, with the simple vote count kept only as a baseline.

Three papers supply methods we reuse but extend in a specific way. VERI-DPO (VERI-DPO) fine-tunes an 8B verifier to label claim–evidence pairs and uses it offline to build preference data for direct preference optimization (DPO, a method that trains a model to prefer one output over another). It also pairs this with a coverage reward to avoid vague, short summaries. Our verifier is still compact and still used to mine preferences, but its unit of judgment is the pooled verdict rather than a proposition–snippet pair. VERI-DPO also notes that it defaults to "Not Addressed" when its retriever misses context, a failure that cross-document aggregation targets directly.

SCRPO (Learning from Self Critique and Refinement) distills atomic-fact self-critique into DPO. It argues against test-time refinement on latency grounds and itself reports missing "nuanced meaning shifts" such as clinical certainty. We accept the latency argument, so training-time alignment is our main deployment. We address the nuance failure with a dedicated certainty target.

Guided continued pre-training (Improving Factuality in Clinical Abstractive Multi-Document Summarization by Guided Continued Pre-training) injects PICO structure into BART pre-training. We instead make PICO an explicit intermediate output that a verifier can inspect, and we test whether the two approaches stack.

On the reasoning side, two-stage prompting (TSGP) and quote-then-score (Unleashing LLMs' Proficiency in Zero-shot Essay Scoring) show that making small models extract before they judge lets them beat larger single-pass systems. HaluAgent (Small Agent Can Also Rock!) shows that a 7B model can become a strong hallucination detector from 2,000 trajectories. However, it only detects errors, works sentence by sentence, and is not specialized to a domain.

One cited paper pulls the other way. Domain-conditional PMI decoding (Mitigating Hallucination in Abstractive Summarization with Domain-Conditional Mutual Information) treats domain priors as a cause of hallucination. This raises the possibility that a domain-specialized summarizer drifts toward a "default verdict," for example reflexively favorable treatment language. We test this directly instead of assuming specialization helps.

Research Program

The pipeline has three trained components built on one open backbone family at 1.5B, 3B and 8B.

The first component is an extractor. It turns each study abstract into a table row: PICO fields, effect direction (increase, decrease, or no difference), and a hedging cue taken from the study's own language. Training rows are drafted by a teacher model and filtered by agreement with the study conclusion sentences.

The second component is the verdict verifier. Given the table and a candidate conclusion sentence, it outputs a distribution over the supported direction and a three-level certainty label, then scores agreement with what the candidate states. It is supervised by the reviews' own conclusions in MS^2 training data. Certainty labels come from hedging lexicons applied to review conclusions and are validated on a clinician-annotated subset. As a reference point, the vote-count aggregate is

$$\hat d = \operatorname{sign}\Big(\sum_i w_i d_i\Big),$$

where $d_i \in \{-1, 0, +1\}$ is study $i$'s extracted effect direction and $w_i$ its sample size. The learned verifier must beat this rule on review agreement to justify itself.

The third component is the summarizer, aligned with DPO on preference pairs. The rejected summary differs from the chosen one in verdict direction or certainty level, according to the verifier. To avoid rewarding vague hedging, we add a coverage term that counts table rows referenced in the summary, following VERI-DPO's coverage utility.

At inference the cost is one extraction pass per study, one synthesis pass, and optionally best-of-4 reranking by the verifier. That is far from the $14.45-per-document multi-agent loop of SummQ (Learning to Summarize by Learning to Quiz).

Baselines

The naive reference point is the same backbone fine-tuned with ordinary supervised learning on MS^2 inputs and reference summaries, decoded greedily. The strongest published numbers on this data are modest. MS^2 reports ROUGE 27.56 for BART and direction agreement of about 50% (MS^2). Guided continued pre-training reports a 15% reduction in ΔEI distance, a measure of mismatch in evidence-inference direction, relative to a Longformer baseline (Improving Factuality...). We will retrain the BART-based guided-pretraining model on our splits.

The strongest competing alignment method is VERI-DPO-style training with a proposition-level verifier. We build it at the same 8B size, from the same teacher budget, with the same DPO steps and coverage term; only the verifier's unit of judgment differs. Its reported result is a drop in unsupported claims from 10.7% to 1.9% on MIMIC-III BHC (VERI-DPO). No reported result found on Lacuna for VERI-DPO or SCRPO on MS^2 or Cochrane. We also run SCRPO-style self-critique DPO at each model size, domain-conditional PMI decoding on the supervised baseline, and prompted GPT-4o given the full input or the extracted table. For GPT-4o, the clinician-preference reference is that ChatGPT summaries of main results were preferred 26% of the time, against 34% for human references (Evaluating large language models on medical evidence summarization).

Three ablations isolate the key ingredient. The first keeps the table but drops the verifier, testing whether extraction alone does the work. The second keeps the verifier but has it read raw abstracts instead of the table. The third replaces the learned verifier with the vote-count rule above.

Beating the field means the following. Direction agreement and certainty agreement must be significantly higher than the proposition-verifier baseline at equal compute. Proposition-level support must be no worse. Summary length and table coverage must be no lower, so that the gain cannot be explained by the model saying less.

Why not just…?

Why not just rerank samples with an off-the-shelf faithfulness checker such as SummaC? SummaC aggregates sentence-pair entailment scores (SummaC). It reaches 74.4% balanced accuracy on news inconsistency benchmarks. Its pairwise design is exactly what lets a verdict supported by one study pass. We include best-of-4 with SummaC as a baseline, and on our verdict-error test subset it should perform near chance. If it does not, the core claim is weakened.

Why not just extract-then-summarize with prompting, as two-stage scaffolds do? This is the real threat. The "More is Less" finding suggests that focused input alone may recover much of the gain. The table-without-verifier ablation settles it. If that ablation closes most of the gap to the full system, we will report that the scaffold is the contribution and that the verifier adds little.

Why not just use VERI-DPO's proposition verifier, which already curbs unsupported claims? It answers a different question: whether some passage supports a claim. VERI-DPO also reports a residual 6.4% unsupported rate under a GPT-4o judge and a dependence on retrieval. The matched comparison above tests whether verdict-level supervision adds value beyond proposition-level supervision.

Why not just prompt a frontier model? Prompted ChatGPT is exactly where the certainty illusions were observed. On clinical-trial inference, models above 70B showed diminishing returns (SemEval-2024 Task 2). We still run GPT-4o, both with and without the table, as a ceiling reference.

Proposed Experiments

We train on MS^2 and test on its held-out reviews and on Cochrane/MSLR. We report ROUGE for continuity, direction agreement, ΔEI, certainty agreement against lexicon labels, and clinician-rated certainty agreement on 200 reviews. We also report proposition support rate, summary length, and table coverage to detect the say-less trap.

To check that the verifier transfers to clinical-trial inference, we evaluate it zero-shot on NLI4CT. There, the best reported F1 is 0.8, from Mixtral-8x7B (SemEval-2024 Task 2).

Three checks decide the proposal. First, a constructed subset where the reviews and the majority of studies disagree: here the learned verifier must beat the vote-count rule, or the learned aggregator is unjustified. Second, the size sweep: aligned 1.5B and 3B summarizers must improve over their supervised baselines, where SCRPO-style self-critique regresses. Third, the domain-prior test: we measure whether verdict errors after specialization skew toward a default direction, and whether domain-conditional PMI decoding removes any residual skew on top of our method.