CoRe: A Comprehensive Framework for Cross-Image Comparative Reasoning in Vision-Language Models

2026-07-14

AI Summary

This work addresses the challenge that existing vision-language models struggle to simultaneously achieve fine-grained attribute grounding and global consistency in cross-image comparative reasoning. To this end, the authors propose CoRe, a unified framework featuring three key contributions: the creation of CoRe-Bench, the first fine-grained benchmark for cross-image comparative reasoning; the automatic construction of CoRe-20K, a large-scale triplet dataset leveraging multi-expert collaboration and structured visual metadata; and the design of TriSR, a structured reward mechanism optimized jointly with GRPO. Evaluated on CoRe-Bench, the proposed method outperforms the strongest baseline by a substantial margin of 28.2 percentage points while maintaining competitive performance on standard multimodal benchmarks.

Abstract

Cross-image comparative reasoning remains challenging for vision-language models (VLMs), especially when correct prediction requires fine-grained attribute grounding and globally consistent reasoning. We present CoRe, a unified framework for this problem. CoRe includes: (i) CoRe-20K, a large-scale triplet-based training set automatically constructed from structured visual metadata through a multi-expert collaborative pipeline, covering counting, depth, distance, and spatial relations; (ii) TriSR, a structured reward framework that jointly supervises attribute grounding, judgment alignment, and triplet consistency under GRPO optimization; and (iii) CoRe-Bench, the first benchmark dedicated to fine-grained cross-image comparative reasoning. Experiments show that CoRe substantially outperforms existing VLMs on CoRe-Bench while remaining competitive on standard multimodal benchmarks, achieving a 28.2-point gain in partial accuracy over the strongest baseline.

Problem

• cross-image comparative reasoning
• vision-language models
• fine-grained attribute grounding
• globally consistent reasoning

Innovation

• cross-image comparative reasoning
• vision-language models
• structured reward
• triplet-based training
• fine-grained attribute grounding