VERI-DPO: Evidence-Aware Alignment for Clinical Summarization via Claim Verification and Direct Preference Optimization

Weixin Liu, Congning Ni, Qingyuan Song, Susannah Rose, Christopher T. Symons, Murat Kantarcioglu, Bradley Malin, Zhijun Yin

TL;DR

VERI-DPO is a framework that aligns clinical summarizers to be more faithful to medical records without sacrificing detail. By using a specialized "compact verifier" to automatically mine high-quality preference pairs, it significantly reduces hallucinations (unsupported claims) while avoiding the common pitfall where models become overly vague to avoid making mistakes.

What Problem Does This Paper Address?

Generating a Brief Hospital Course (BHC)—the narrative summary of a patient's stay—is a high-stakes task. These summaries are essential for clinical handoffs, but the evidence needed to write them is often fragmented across hundreds of messy, longitudinal notes in an Electronic Health Record (EHR).

While Large Language Models (LLMs) are great at summarizing, they suffer from two major issues in medicine:

  1. Hallucinations: They often make plausible-sounding but unsubstantiated claims (e.g., saying a patient had a surgery they didn't actually have).
  2. "Say-Less" Degeneration: When we try to "align" models to be more truthful using standard methods, they often learn a "shortcut": if they don't say anything specific, they can't be wrong. This results in short, vague summaries that are useless to doctors.

The authors tackle the challenge of creating summaries that are both faithful (supported by evidence) and informative (sufficiently detailed).

Methods

The researchers introduce VERI-DPO (Verifier-Driven Direct Preference Optimization), a three-stage pipeline:

  1. Stage A: Training a Compact Verifier. Instead of relying on expensive human experts for every training example, they fine-tune a lightweight LLM (like Llama-3-8B) to act as a judge. This verifier looks at a claim and a snippet of evidence and predicts one of three labels: Supported, Not Supported, or Not Addressed. Crucially, it uses a "single-token" format for efficiency and can be calibrated to be more or less sensitive to errors.
  2. Stage B: Verifier-Driven Preference Mining. The system generates multiple candidate summaries for a patient. The verifier then "scores" every sentence in those summaries. The authors then "mine" pairs of summaries: a "chosen" one (high coverage, few errors) and a "rejected" one (contains at least one high-confidence contradiction).
  3. Stage C: DPO Alignment. They use Direct Preference Optimization (DPO) to fine-tune the main summarizer. Unlike standard fine-tuning, DPO teaches the model to prefer the "chosen" summary over the "rejected" one, effectively distilling the verifier's "knowledge" of what a good, evidence-backed summary looks like directly into the model.

Datasets

The study uses the MIMIC-III-Ext-VeriFact-BHC dataset, which contains:

  • Data from 100 ICU patients (125 hospital admissions).
  • Over 4,700 clinical notes as reference evidence.
  • 13,070 proposition-level annotations where clinicians labeled claims as Supported, Not Supported, or Not Addressed.
  • The authors used patient-level splits (72 train / 8 dev / 20 test) to ensure the model wasn't just memorizing specific patient cases.

Key Insights and Ideas

  • Contradiction Anchoring: The "aha moment" here is that you don't just want a "better" summary; you want a pair where the rejected summary is specifically anchored by a clear contradiction. This provides a much stronger signal for the model to learn what not to do.
  • The Utility Function: To prevent the "say-less" problem, the authors designed a utility function that penalizes contradictions heavily but also rewards "coverage" (mentioning more supported claims) and penalizes summaries that are too short or repetitive.
  • Single-Token Efficiency: By forcing the verifier to output just one token (A, B, or C), they make the verification process fast enough to scale across thousands of generated candidates during the preference mining stage.

Key Results

The VERI-DPO model was evaluated against a base Llama-3 model and a standard Supervised Fine-Tuning (SFT) baseline.

  • Hallucination Reduction: Under a local verifier judge, the rate of "Not Supported" claims dropped from 10.7% to 1.9%. Even when judged by an external, high-powered model (GPT-4o), the error rate dropped from 11.6% to 6.4%.
  • Beating the "Say-Less" Trap: Unlike many alignment methods that result in shorter text, the VERI-DPO summaries were actually longer on average (2159 characters vs. 1855 for the base model) while having more supported claims (19.1 vs. 15.0).
  • Validity: The percentage of summaries meeting all clinical formatting and length constraints improved from 76.7% to 82.5%.

Limitations and Future Work

  • Small Cohort: The study was limited to 100 patients from a single ICU-focused dataset. Clinical language and documentation styles vary wildly across different hospitals and specialties.
  • Retrieval Dependence: The system is only as good as its retriever. If the "Compact Verifier" can't find the right note in the EHR to check a claim, it might incorrectly label something as "Not Addressed."
  • Automated Judges: While the authors used two different AI judges, they acknowledge that "absolute NS-rates should be treated as proxies." The ultimate test would be a large-scale head-to-head trial with human clinicians.
  • Future Directions: The authors suggest that this "verifier-in-the-loop" approach could be expanded to other complex domains like legal or technical summarization where fact-checking is paramount.

Summary updated

Abstract

Brief Hospital Course (BHC) narratives must be clinically useful yet faithful to fragmented EHR evidence. LLM-based clinical summarizers still introduce unsupported statements, and alignment can encourage omissions ("say-less" degeneration). We introduce VERI-DPO, which uses claim verification to mine preferences and distill them into the summarizer with Direct Preference Optimization (DPO). On MIMIC-III-Ext-VeriFact-BHC (100 ICU patients; patient-level splits), we train a retrieval-augmented verifier to label claim-evidence pairs as Supported, Not Supported, or Not Addressed via a single-token format. The verifier scores sentence-level claims from sampled BHC candidates and aggregates margins into a coverage-aware utility to mine length-controlled, contradiction-anchored preference pairs. On held-out patients, verifier-mined preferences separate candidates by contradiction density, and VERI-DPO reduces Not Supported claim rates from 10.7% to 1.9% (local verifier judge) and from 11.6% to 6.4% (GPT-4o judge), while improving validity from 76.7% to 82.5% and maintaining informative length.
BibTeX
@misc{liu2026veri,
  author = {Weixin Liu and Congning Ni and Qingyuan Song and Susannah Rose and Christopher T. Symons and Murat Kantarcioglu and Bradley Malin and Zhijun Yin},
  title = {{VERI-DPO: Evidence-Aware Alignment for Clinical Summarization via Claim Verification and Direct Preference Optimization}},
  year = {2026},
  eprint = {2603.10494},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2603.10494},
}
Figure