Referring Multiple Regions with Large Multimodal Models via Contextual Latent Steering

2026-05-03

AI Summary

Existing large-scale multimodal models struggle to accurately refer to multiple image regions under visual prompting, primarily due to insufficient exploitation of global contextual information. To address this limitation, this work proposes Contextual Latent Steering (CSteer), a training-free representation editing method applied during inference. CSteer precomputes contextual vectors that implicitly encode region distinctiveness and global attention patterns, thereby endowing general-purpose multimodal models with context-aware multi-region referring capabilities—without any architectural modifications or fine-tuning. The approach outperforms specialized region-referring models across multiple benchmarks, achieving, for the first time, high-precision, training-agnostic, and broadly applicable multi-region reference. This represents a significant advance by reducing reliance on task-specific customized models.

Abstract

Large Multimodal Models (LMMs) have recently demonstrated their proficiency in holistic visual comprehension. However, most of them struggle to tackle region-level perception guided by visual prompts, especially for cases where multiple regions are referred simultaneously, or scenarios where global contexts are necessary for precise visual referring. We introduce Contextual Latent Steering (CSteer), a training-free approach for guiding general LMMs to refer multiple regions contextually, without expensive fine-tuning or architectural modifications. CSteer starts with pre-computing contextual vectors that implicitly represent visual referring behaviors, such as differentiation among regions and attention to global contexts, followed by representation editing during inference time. Experimental results on multiple datasets indicate that general LMMs with CSteer outperform tailored referring LMMs in most cases, suggesting a promising solution in training-free, and setting new state-of-the-art for this field. Code is available at https://github.com/xing0047/csteer.git.

Problem

• Large Multimodal Models
• region-level perception
• visual referring
• multiple regions
• global context

Innovation

• Contextual Latent Steering
• Large Multimodal Models
• Training-free
• Visual Referring
• Region-level Perception