Reducing Detail Hallucinations in Long-Context Regulatory Understanding via Targeted Preference Optimization

2026-04-24

AI Summary

This study addresses the critical issue of detail hallucination in large language models when processing lengthy regulatory documents, where errors in key parameters—such as thresholds, units, and ranges—compromise safety. The work formalizes this problem for the first time and introduces a fine-grained taxonomy encompassing five error types. To mitigate these inaccuracies, the authors propose DetailDPO, a novel framework that constructs minimally perturbed contrastive samples differing only along a single detail dimension and leverages fine-grained human preference annotations to guide targeted Direct Preference Optimization (DPO). Applied to long-context models including Qwen2.5 and Llama-3.1, DetailDPO focuses gradient updates on detail-relevant tokens. Evaluated across context lengths from 8K to 64K, the method reduces detail error rates by 42%–61% compared to baselines and significantly improves accuracy in cross-domain applications such as finance and healthcare.

Technology Category

• Natural Language Processing: Safety and Robustness
• Machine Learning: Large Multimodal Models (LMMs)
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Economics, Online Markets and Human Computation: LLM based quality controls for crowd work
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions

Abstract

Large language models (LLMs) frequently produce \emph{detail hallucinations} when processing long regulatory documents, including subtle errors in threshold values, units, scopes, obligation levels, and conditions that preserve surface plausibility while corrupting safety-critical parameters. We formalize this phenomenon through a fine-grained \emph{Detail Error Taxonomy} of five error types and introduce \textbf{DetailBench}, a benchmark built from 172 real regulatory documents and 150 synthetic documents spanning three jurisdictions, with human-annotated detail-level ground truth comprising 13,000 preference pairs. We propose \textbf{DetailDPO}, a targeted preference optimization framework that constructs contrastive pairs differing in exactly one detail dimension, concentrating DPO gradient signal on detail-bearing~tokens. We provide theoretical analysis showing why \emph{minimal detail perturbation} pairs yield gradient concentration under mild assumptions. Experiments on the Qwen2.5 family (7B, 14B, 72B) and Llama-3.1-8B across three context-length tiers (8K--64K tokens) show that DetailDPO reduces the Detail Error Rate by 42--61\% relative to baselines, with consistent gains across all five error types and cross-domain transfer to financial and medical documents.

Problem

• detail hallucinations
• regulatory understanding
• long-context
• safety-critical errors
• large language models

Innovation

• detail hallucination
• preference optimization
• DetailDPO
• long-context understanding
• regulatory compliance