Decomposed On-Policy Distillation for Vision-Language Reasoning: Steering Gradients for Visual Grounding

• URL: http://arxiv.org/abs/2606.00564v1
• Date: Sat, 30 May 2026 06:34:37 GMT
• Title: Decomposed On-Policy Distillation for Vision-Language Reasoning: Steering Gradients for Visual Grounding
• Abstract summary: We challenge the standard monolithic view of Vision-Language Model (VLM) distillation by mathematically decomposing the loss into two distinct components: the language prior and visual grounding.<n>We introduce Visual Gradient Steering (VGS), a method that dynamically reorients the update vector to prioritize the visual subspace.<n> Experimental results on multiple distillation settings and complex multimodal benchmarks demonstrate that VGS significantly outperforms the standard monolithic formulation of on-policy distillation.
• Score: 76.37424140943456
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: While on-policy distillation offers dense supervision for training small reasoning models, its optimization dynamics in the multimodal domain remain under-explored. In this work, we challenge the standard monolithic view of Vision-Language Model (VLM) distillation by mathematically decomposing the loss into two distinct components: the language prior and visual grounding. Our analysis uncovers that gradient vectors for these components are nearly orthogonal, indicating that the objective of aligning with the teacher's language distribution is geometrically independent from the objective of matching its visual perception. Consequently, standard optimization passively follows a suboptimal compromise trajectory that implicitly balances the two objectives. Hypothesizing that visual grounding constitutes the primary bottleneck for vision-language reasoning, we introduce Visual Gradient Steering (VGS), a method that dynamically reorients the update vector to prioritize the visual subspace. Experimental results on multiple distillation settings and complex multimodal benchmarks demonstrate that VGS significantly outperforms the standard monolithic formulation of on-policy distillation, achieving superior grounding with minimal training overhead.

Related papers

• ExACT: Exemplar-Driven Calibrated Refinement for Training-Free Visual Grounding in Remote Sensing Images [51.53960096934913]
  Remote sensing visual grounding (RSVG) aims to locate specific objects in high-resolution RS imagery using free-form natural language descriptions.<n>Recent advances in multimodal large language models (MLLMs) show great potential for such open-vocabulary RSVG.<n>We propose Exemplar-driven Calibrated Refinement (ExACT), a training-free framework driven by a one-shot visual prompting mechanism.
  arXiv  Detail & Related papers  (2026-06-27T13:50:39Z)
• PointVG-R: Internalizing Geometric Reasoning in MLLMs for Precise Pointing Localization via Visual Chain of Thought [19.680475189691965]
  Pointing-based visual grounding requires models to precisely locate target objects by deciphering complex spatial relationships between the visual scene and pointing gestures.<n>We propose PointVG-R, a reasoning-guided Multi-modal Large Language Model (MLLM)<n>We design a novel geometric reasoning pipeline that simulates the iterative cognitive process humans employ when interpreting pointing gestures.
  arXiv  Detail & Related papers  (2026-06-23T13:06:51Z)
• GEM: Generative Supervision Helps Embodied Intelligence [60.28190143133428]
  Embodied Vision-Language Models (VLMs) have demonstrated impressive performance and generalization in robotics.<n>A significant gap remains between the high-level semantic focus of standard text-guided pre-training paradigms and the low-level spatial and physical knowledge critical for execution in embodied environments.<n>We introduce GEM, a Generative-supervised Embodied vision-language Model designed to bridge this divide.
  arXiv  Detail & Related papers  (2026-05-27T14:39:42Z)
• Visually-Guided Policy Optimization for Multimodal Reasoning [60.035908460318126]
  We propose Visually-Guided Policy Optimization (VGPO) to reinforce visual focus during policy optimization.<n> VGPO initially introduces a Visual Attention Compensation mechanism that leverages visual similarity to localize and amplify visual cues.<n>We show that VGPO achieves better visual activation and superior performance in mathematical multimodal reasoning and visual-dependent tasks.
  arXiv  Detail & Related papers  (2026-04-10T14:22:38Z)
• ProVG: Progressive Visual Grounding via Language Decoupling for Remote Sensing Imagery [20.725910735285762]
  Remote sensing visual grounding aims to localize objects in remote sensing imagery according to natural language expressions.<n>We propose textbfProVG, a novel RSVG framework that improves localization accuracy by decoupling language expressions into global context, spatial relations, and object attributes.<n>ProVG consistently outperforms existing methods, achieving new state-of-the-art performance.
  arXiv  Detail & Related papers  (2026-04-02T10:59:19Z)
• ThinkJEPA: Empowering Latent World Models with Large Vision-Language Reasoning Model [53.15040805435013]
  Vision-language models (VLMs) provide strong semantic grounding and general knowledge by reasoning over uniformly sampled frames.<n>We propose a VLM-guided JEPA-style latent world modeling framework that combines dense-frame dynamics modeling with long-horizon semantic guidance.
  arXiv  Detail & Related papers  (2026-03-23T17:59:42Z)
• Perception-Aware Multimodal Spatial Reasoning from Monocular Images [57.42071289037214]
  spatial reasoning from monocular images is essential for autonomous driving.<n>Current Vision-Language Models (VLMs) still struggle with fine-grained geometric perception.<n>We propose a perception-aware multimodal reasoning framework that equipsVLMs with explicit object-centric grounding ability.
  arXiv  Detail & Related papers  (2026-03-07T02:05:12Z)
• Dynamic Multimodal Activation Steering for Hallucination Mitigation in Large Vision-Language Models [22.535916867005955]
  Large Vision-Language Models (LVLMs) exhibit outstanding performance on vision-language tasks but struggle with hallucination problems.<n>We propose Dynamic Multimodal Activation Steering, a training-free approach for hallucination mitigation.
  arXiv  Detail & Related papers  (2026-02-25T09:10:00Z)
• FRISM: Fine-Grained Reasoning Injection via Subspace-Level Model Merging for Vision-Language Models [20.47311573790516]
  We propose FRISM (Fine-grained Reasoning Injection via Subspace-level model Merging), a fine-grained reasoning injection framework based on subspace-level model merging.<n>Experiments demonstrate that FRISM effectively improves reasoning capabilities without compromising the model's original visual capabilities.
  arXiv  Detail & Related papers  (2026-01-29T02:36:19Z)
• GoViG: Goal-Conditioned Visual Navigation Instruction Generation [69.79110149746506]
  We introduce Goal-Conditioned Visual Navigation Instruction Generation (GoViG), a new task that aims to autonomously generate precise and contextually coherent navigation instructions.<n>GoViG exclusively leverages raw egocentric visual data, substantially improving its adaptability to unseen and unstructured environments.
  arXiv  Detail & Related papers  (2025-08-13T07:05:17Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.