Coverage-Controlled Preference Mining from Noisy Claim Verification for Evidence-Grounded Generation
arXiv:2603.10494v2 Announce Type: replace-cross Abstract: Evidence-grounded generation produces summaries whose claims should be supported by supplied evidence, but claim-level verifiers provide noisy feedback and can reward models that simply say less. We study this problem in clinical Brief Hospital Course summarization, where outputs must remain grounded in patient-specific EHR evidence. We introduce VERI-DPO, a preference-mining framework that converts noisy claim verification into coverage-controlled summary-level preferences. For each evidence-window prompt, VERI-DPO samples multiple can
The proliferation of evidence-grounded AI generation models highlights the urgent need for robust verification mechanisms that do not compromise output quality, especially in sensitive domains like clinical summarization.
This research addresses a core challenge in generative AI: ensuring factual accuracy and grounding, which is critical for trust and adoption in high-stakes applications.
The introduction of VERI-DPO provides a novel methodology for training generative models using preference mining from noisy feedback, potentially leading to more reliable and controllable AI outputs.
- · AI developers
- · Healthcare sector (AI-assisted summarization)
- · Content generation platforms
- · AI safety researchers
- · Generative AI models with poor grounding
- · Manual claim verification processes
Improved factual correctness and reduced hallucinations in evidence-grounded generative AI systems.
Increased adoption of AI in domains requiring high precision and trustworthiness, such as medical records or legal documents.
New benchmarks and evaluation metrics emerge that specifically measure coverage-controlled and preference-mined model performance, shifting development focus.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.LG