Reducing Detail Hallucinations in Long-Context Regulatory Understanding via Targeted Preference Optimization
Fuente:
arXiv
Saved in:
| Main Authors: | , , , , , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2026
|
| Subjects: | |
| Online Access: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866910165713813504 |
|---|---|
| author | Liu, Yang Chong, Bin Lin, Yuhan Zhang, Chongyang Zheng, Hao Zhang, Ziyi Liang, Jiayu Ran, Ran Li, Qian Xu, Kefu |
| author_facet | Liu, Yang Chong, Bin Lin, Yuhan Zhang, Chongyang Zheng, Hao Zhang, Ziyi Liang, Jiayu Ran, Ran Li, Qian Xu, Kefu |
| contents | 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. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2604_23113 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Reducing Detail Hallucinations in Long-Context Regulatory Understanding via Targeted Preference Optimization Liu, Yang Chong, Bin Lin, Yuhan Zhang, Chongyang Zheng, Hao Zhang, Ziyi Liang, Jiayu Ran, Ran Li, Qian Xu, Kefu Social and Information Networks 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. |
| title | Reducing Detail Hallucinations in Long-Context Regulatory Understanding via Targeted Preference Optimization |
| topic | Social and Information Networks |
| url | https://arxiv.org/abs/2604.23113 |