Upcoming publication
Manuscript / August 2026
Multi-Step Reasoning for Implicit Data Retrieval
A Benchmark and an SFT+GRPO Framework for Document Field Extraction with Vision-Language Models
Abstract
Many document fields are not printed as a single value. They must be derived by locating visual evidence, selecting the relevant items and performing arithmetic or categorical reasoning. Our paper defines this challenge as implicit data retrieval and introduces THINK-RETRIEVE-THINK, a structured training framework for grounded document reasoning.
Key results
95.9%Field accuracy after SFT+GRPO
4,151Held-out document-field pairs
<0.1%Incomplete-generation rate
77.1OmniDocBench score, up from 61.8
