The Unseen Bias: How Norm Discrepancy in Pre-Norm MLLMs Leads to Visual Information Loss

• URL: http://arxiv.org/abs/2512.08374v1
• Date: Tue, 09 Dec 2025 08:57:11 GMT
• Title: The Unseen Bias: How Norm Discrepancy in Pre-Norm MLLMs Leads to Visual Information Loss
• Authors: Bozhou Li, Xinda Xue, Sihan Yang, Yang Shi, Xinlong Chen, Yushuo Guan, Yuanxing Zhang, Wentao Zhang,
• Abstract summary: Multimodal Large Language Models (MLLMs) couple pre-trained vision encoders and language models.<n>Their reliance on the ubiquitous Pre-Norm architecture introduces a severe norm disparity between the high-norm visual tokens and the low-norm text tokens.<n>We propose a simple yet effective solution: inserting a single, carefully LayerNorm layer after the visual projector to enforce norm alignment.
• Score: 15.598471176315913
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Multimodal Large Language Models (MLLMs), which couple pre-trained vision encoders and language models, have shown remarkable capabilities. However, their reliance on the ubiquitous Pre-Norm architecture introduces a subtle yet critical flaw: a severe norm disparity between the high-norm visual tokens and the low-norm text tokens. In this work, we present a formal theoretical analysis demonstrating that this imbalance is not a static issue. Instead, it induces an ``asymmetric update dynamic,'' where high-norm visual tokens exhibit a ``representational inertia,'' causing them to transform semantically much slower than their textual counterparts. This fundamentally impairs effective cross-modal feature fusion. Our empirical validation across a range of mainstream MLLMs confirms that this theoretical dynamic -- the persistence of norm disparity and the resulting asymmetric update rates -- is a prevalent phenomenon. Based on this insight, we propose a remarkably simple yet effective solution: inserting a single, carefully initialized LayerNorm layer after the visual projector to enforce norm alignment. Experiments conducted on the LLaVA-1.5 architecture show that this intervention yields significant performance gains not only on a wide suite of multimodal benchmarks but also, notably, on text-only evaluations such as MMLU, suggesting that resolving the architectural imbalance leads to a more holistically capable model.

Related papers

• Same Answer, Different Representations: Hidden instability in VLMs [65.36933543377346]
  We introduce a representation-aware and frequency-aware evaluation framework that measures internal embedding drift, spectral sensitivity, and structural smoothness.<n>We apply this framework to modern Vision Language Models (VLMs) across the SEEDBench, MMMU, and POPE datasets.
  arXiv  Detail & Related papers  (2026-02-06T12:24:26Z)
• AdaTok: Adaptive Token Compression with Object-Aware Representations for Efficient Multimodal LLMs [29.68162972167947]
  We propose an object-level token merging strategy for Adaptive Token compression.<n>Our approach averagely, utilizes only 10% tokens while achieving almost 96% of the vanilla model's performance.
  arXiv  Detail & Related papers  (2025-11-18T06:12:15Z)
• FUDOKI: Discrete Flow-based Unified Understanding and Generation via Kinetic-Optimal Velocities [76.46448367752944]
  multimodal large language models (MLLMs) unify visual understanding and image generation within a single framework.<n>Most existing MLLMs rely on autore (AR) architectures, which impose inherent limitations on future development.<n>We introduce FUDOKI, a unified multimodal model purely based on discrete flow matching.
  arXiv  Detail & Related papers  (2025-05-26T15:46:53Z)
• Multimodal LLM-Guided Semantic Correction in Text-to-Image Diffusion [52.315729095824906]
  MLLM Semantic-Corrected Ping-Pong-Ahead Diffusion (PPAD) is a novel framework that introduces a Multimodal Large Language Model (MLLM) as a semantic observer during inference.<n>It performs real-time analysis on intermediate generations, identifies latent semantic inconsistencies, and translates feedback into controllable signals that actively guide the remaining denoising steps.<n>Extensive experiments demonstrate PPAD's significant improvements.
  arXiv  Detail & Related papers  (2025-05-26T14:42:35Z)
• Demystifying Singular Defects in Large Language Models [61.98878352956125]
  In large language models (LLMs), the underlying causes of high-norm tokens remain largely unexplored.<n>We provide both theoretical insights and empirical validation across a range of recent models.<n>We showcase two practical applications of these findings: the improvement of quantization schemes and the design of LLM signatures.
  arXiv  Detail & Related papers  (2025-02-10T20:09:16Z)
• Analyzing Finetuning Representation Shift for Multimodal LLMs Steering [56.710375516257876]
  We propose to map hidden states to interpretable visual and textual concepts.<n>This enables us to more efficiently compare certain semantic dynamics, such as the shift from an original and fine-tuned model.<n>We also demonstrate the use of shift vectors to capture these concepts changes.
  arXiv  Detail & Related papers  (2025-01-06T13:37:13Z)
• NormXLogit: The Head-on-Top Never Lies [15.215985417763472]
  We propose a novel technique for assessing the significance of individual input tokens.<n>This method operates based on the input and output representations associated with each token.<n>We reveal a significant relationship between a token's importance and the extent to which its representation can resemble the model's final prediction.
  arXiv  Detail & Related papers  (2024-11-25T10:12:27Z)
• Fine-Grained Verifiers: Preference Modeling as Next-token Prediction in Vision-Language Alignment [57.0121616203175]
  We propose FiSAO, a novel self-alignment method that utilizes the model's own visual encoder as a fine-grained verifier to improve vision-language alignment.<n>By leveraging token-level feedback from the vision encoder, FiSAO significantly improves vision-language alignment, even surpassing traditional preference tuning methods that require additional data.
  arXiv  Detail & Related papers  (2024-10-18T03:34:32Z)
• Sycophancy in Vision-Language Models: A Systematic Analysis and an Inference-Time Mitigation Framework [18.54098084470481]
  We analyze sycophancy across vision-language benchmarks and propose an inference-time mitigation framework.<n>Our framework effectively mitigates sycophancy across all evaluated models, while maintaining performance on neutral prompts.
  arXiv  Detail & Related papers  (2024-08-21T01:03:21Z)

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.