Visual instruction tuning aligns modalities in intermediate LLM layers, not early ones
A new arXiv paper investigates how visual instruction tuning embeds image features into the layer-wise hierarchy of LLM backbones across diverse vision-language architectures. Using probing analyses and causal interventions, the authors find that instruction tuning routes visual features into intermediate semantic layers, bypassing early unimodal-processing layers. They further show that fine-tuning restricted to these intermediate layers alone preserves full fine-tuning performance on vision-centric benchmarks while reducing training time, suggesting multimodal integration is a localized phenomenon.
Related guides (2)
Related events (8)
MAGIC: Multimodal Alignment & Grounding-aware Instruction Coreset for Vision-Language Models
MAGIC is a training-free coreset selection method for multimodal instruction tuning that uses three intrinsic signals—Multimodal Gain, Bridging Relevance, and Skill-Neuron Signatures—to identify compact, behaviorally faithful training subsets without backpropagation. The method operates in a three-stage pipeline: filtering low-gain examples, ranking by a quality objective, and bucket-wise budget allocation over neuron signatures. On LLaVA-665K and Vision-Flan datasets with 20% data budgets, MAGIC matches or slightly exceeds full fine-tuning performance (100.3% and 101.6% relative) while reducing wall-clock training time by 73.7%. Results transfer to LLaVA-1.5-7B and -13B target models.
LoMo: Local Modality Substitution for Deeper Vision-Language Fusion
This paper identifies a 'carrier sensitivity' problem in Vision-Language Models (VLMs), where replacing textual queries with rendered-image equivalents causes significant performance degradation due to asymmetric roles of text and images in training data. The authors propose Local Modality Substitution (LoMo), a data curation paradigm that reformulates single-modality prompts into interleaved multimodal sequences by dynamically rendering text spans as images, enforcing cross-modal representational invariance. Evaluated across 13 multimodal benchmarks, LoMo improves over standard supervised fine-tuning by 2.67 points on LLaVA-OneVision-1.5-8B and 2.82 points on Qwen3.5-9B. The approach is architecture-agnostic and lightweight, requiring no changes to model architecture.
VLMs May Not Globally Enhance Human Alignment over LLMs During Natural Reading
This paper compares matched LLM and VLM pairs in a text-only setting to isolate the effect of multimodal training history on human-like language processing. Using whole-cortex fMRI and eye-tracking data from natural reading, the authors find that multimodal pretraining does not confer a uniform global advantage in human alignment. However, VLMs show selective advantages when sentences contain stronger visual semantic content, with converging evidence from both neural and behavioral measures. The findings suggest language-internal representations remain the primary driver of human text processing alignment.
Vision-OPD: On-Policy Self-Distillation for Fine-Grained Visual Understanding in MLLMs
Vision-OPD addresses a 'regional-to-global perception gap' in multimodal LLMs, where models answer fine-grained visual questions more accurately when given cropped evidence regions than full images. The method instantiates a crop-conditioned teacher and full-image-conditioned student from the same MLLM, minimizing token-level divergence along on-policy rollouts to transfer regional perception to the full-image policy. This self-distillation requires no external teacher models, ground-truth labels, reward verifiers, or inference-time tools. Benchmarks show competitive or superior performance against larger open-source, closed-source, and agentic 'Thinking-with-Images' models.
Introducing vision to the fine-tuning API
OpenAI has extended its fine-tuning API to support multimodal inputs, allowing developers to fine-tune GPT-4o using both images and text. This enables customization of vision capabilities for domain-specific tasks. The update expands the existing text-only fine-tuning pipeline to handle image-text pairs.
PGT: Procedurally Generated Tasks for Improving Visual Grounding in MLLMs
This paper introduces Procedurally Generated Tasks (PGT), a data-driven framework that overlays geometric primitives on images to create dense supervision signals for fine-grained visual grounding in multimodal large language models. PGT serves both as a training augmentation method and a diagnostic tool to isolate perception failures from semantic priors. Instruction tuning on LLaVA-v1.5-Instruct augmented with PGT data yields gains of up to +20% on the What'sUp benchmark and +13.3% on CV-Bench-2D. The results suggest that spatial reasoning deficits in MLLMs stem primarily from inadequate supervision rather than architectural or resolution constraints.
Preference Optimization for Vision Language Models
This Hugging Face blog post covers the application of Direct Preference Optimization (DPO) to vision-language models (VLMs). It likely discusses how preference learning techniques originally developed for text-only LLMs can be adapted to multimodal settings. The post addresses training methodology for aligning VLMs with human preferences across both visual and textual modalities.
A Dive into Vision-Language Models
This Hugging Face blog post provides a technical overview of vision-language model (VLM) pretraining approaches, covering architectures and training strategies used to align visual and textual representations. It surveys key models and techniques in the multimodal learning space as of early 2023. The post serves as an educational reference for practitioners working with or building VLMs.