When Language Overwrites Vision: Over-Alignment and Geometric Debiasing in Vision-Language Models

• URL: http://arxiv.org/abs/2605.08245v2
• Date: Wed, 13 May 2026 05:26:30 GMT
• Title: When Language Overwrites Vision: Over-Alignment and Geometric Debiasing in Vision-Language Models
• Abstract summary: Vision-Language Models (VLMs) increasingly power high-stakes applications.<n>They routinely hallucinate, confidently describing content not present in the input.<n>We trace these failure modes to a geometric over-alignment.<n>We propose two complementary remedies: a training-free inference strategy and a bias-aware fine-tuning paradigm.
• Score: 4.309108879640932
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Vision-Language Models (VLMs) increasingly power high-stakes applications, from medical imaging to autonomous systems, yet they routinely hallucinate, confidently describing content not present in the input. We investigate the root causes of these failure modes with a mechanistic analysis focusing on the decoder-based VLMs. We trace these failure modes to a geometric over-alignment: to bridge the modality gap required by attention mechanisms, decoder-based VLMs over-align visual embeddings with the text manifold, injecting a statistical linguistic bias that systematically overshadows fine-grained visual evidence. While prior work either aggressively closes this gap or suppresses hallucinations through expensive black-box decoding strategies, none addresses the underlying geometric cause. We provide the first quantitative characterization of this over-alignment, demonstrating that linguistic bias concentrates in the top principal components of a universal, dataset-agnostic text subspace. Building on this insight, we propose two complementary remedies: a training-free inference strategy and a bias-aware fine-tuning paradigm, both of which explicitly project out this subspace from visual representations. Our methods significantly reduce hallucinations across POPE, CHAIR, and AMBER benchmarks, and improve CLAIR scores on long-form captioning tasks, with the training-free variant adding no computational overhead over the base model.

Related papers

• Disentangling Semantic Attention from Structural Bias in the Attention Manifold [61.68044165974505]
  Multimodal Large Language Models (MLLMs) exhibit disproportionate attention toward certain semantically uninformative visual tokens.<n>We introduce Saliency-guided Purification and Adaptive Redistribution (SPAR), a training-free, plug-and-play intervention.
  arXiv  Detail & Related papers  (2026-07-27T05:30:19Z)
• Staying VIGILant: Mitigating Visual Laziness via Counterfactual Visual Alignment in MLLMs [34.999900039604846]
  Multimodal large language models (MLLMs) extend large language models (LLMs) with visual perception, enabling joint reasoning over images and text.<n>Mechanistic studies indicate that this weakness stems from visual laziness.<n>We propose Visual Information Gain In aLignment (VIGIL), a reinforcement-learning framework that shifts the focus from numerical reward fitting to causal visual grounding.
  arXiv  Detail & Related papers  (2026-06-24T21:12:24Z)
• Diagnosing Visual Ignorance in Vision-Language Models [29.2851901986069]
  Vision-Language Models (VLMs) frequently rely on language priors, producing confident answers that are weakly grounded in visual evidence.<n>In this work, we study language-prior reliance from both mechanistic and behavioral perspectives.<n>We find that language-prior reliance is a systematic routing failure affecting both model internals and benchmark validity.
  arXiv  Detail & Related papers  (2026-06-05T04:16:01Z)
• Mitigating Object Hallucinations in Vision-Language Models through Region-Aware Attention Recalibration [7.647944993041615]
  We introduce a training-free inference strategy to dynamically correct semantic drift without relying on abrupt truncations.<n>By computing an outlier-resistant statistical midpoint across various attention heads, we establish a stable anchor for reliable visual representations.<n>This recalibration process effectively rectifies visual-semantic misalignments while fully preserving generative fluency and language priors.
  arXiv  Detail & Related papers  (2026-05-24T09:14:49Z)
• SEM: Sparse Embedding Modulation for Post-Hoc Debiasing of Vision-Language Models [36.40093869747498]
  We propose Sparse Embedding Modulation (SEM), a post-hoc, zero-shot debiasing framework that operates in a Sparse Autoencoder latent space.<n>SEM identifies and modulates bias-relevant neurons while preserving query-relevant ones.<n>Our results demonstrate that sparse latent representations provide an effective foundation for post-hoc debiasing of vision-language models.
  arXiv  Detail & Related papers  (2026-03-19T15:28:08Z)
• Attention-space Contrastive Guidance for Efficient Hallucination Mitigation in LVLMs [9.043999205886658]
  Hallucinations in large vision-language models often arise when language priors dominate over visual evidence.<n>We propose Contrastive Guidance (ACG), a single-pass mechanism that operates within self-attention layers to construct both vision-language and language-only attention paths.<n>ACG achieves state-of-the-art faithfulness and caption quality while significantly reducing computational cost.
  arXiv  Detail & Related papers  (2026-01-20T08:04:18Z)
• SDCD: Structure-Disrupted Contrastive Decoding for Mitigating Hallucinations in Large Vision-Language Models [4.677212795400693]
  Bag-of-Patches behavior of Visions under weak structural supervision acts as a contributing factor of object hallucinations.<n>We introduce a training-free algorithm called Structure-Disrupted Contrastive Decoding (SDCD)<n>By penalizing tokens that maintain high confidence under this structure-less view, SDCD effectively suppresses the texture-driven bias.
  arXiv  Detail & Related papers  (2026-01-07T01:27:58Z)
• Seeing to Act, Prompting to Specify: A Bayesian Factorization of Vision Language Action Policy [59.44168425139687]
  BayesVLA is a Bayesian factorization that decomposes the policy into a visual-action prior, supporting seeing-to-act, and a language-conditioned likelihood, enabling prompt-to-specify.<n>Experiments show superior generalization to unseen instructions, objects, and environments compared to existing methods.
  arXiv  Detail & Related papers  (2025-12-12T01:59:23Z)
• Attention Guided Alignment in Efficient Vision-Language Models [56.20286899428444]
  Large Vision-Language Models (VLMs) rely on effective multimodal alignment between pre-trained vision encoders and Large Language Models (LLMs)<n>This paper presents a comprehensive analysis of attention patterns in efficient VLMs.<n>We introduce Attention-Guided Efficient Vision-Language Models (AGE-VLM), a novel framework that enhances visual grounding through interleaved cross-attention layers.
  arXiv  Detail & Related papers  (2025-11-21T21:36:48Z)
• Text Speaks Louder than Vision: ASCII Art Reveals Textual Biases in Vision-Language Models [93.46875303598577]
  Vision-language models (VLMs) have advanced rapidly in processing multimodal information, but their ability to reconcile conflicting signals remains underexplored.<n>This work investigates how VLMs process ASCII art, a unique medium where textual elements collectively form visual patterns, potentially creating semantic-visual conflicts.
  arXiv  Detail & Related papers  (2025-04-02T10:47:07Z)
• D-Attn: Decomposed Attention for Large Vision-and-Language Models [29.611769371733672]
  We propose Decomposed Attention (D-Attn), a more flexible attention architecture for large vision-and-language models (LVLMs)<n>D-Attn decomposes the 1-D causal self-attention of LVLMs into visual-to-visual, textual-to-visual, and textual-to-textual attentions.<n>Experiments and analysis validate the effectiveness of D-Attn, demonstrating significant improvements on multiple image benchmarks.
  arXiv  Detail & Related papers  (2025-02-04T00:46:11Z)
• Looking Beyond Text: Reducing Language bias in Large Vision-Language Models via Multimodal Dual-Attention and Soft-Image Guidance [67.26434607115392]
  Large vision-language models (LVLMs) have achieved impressive results in various vision-language tasks. LVLMs suffer from hallucinations caused by language bias, leading to diminished focus on images and ineffective visual comprehension. We propose LACING to address the language bias of LVLMs with muLtimodal duAl-attention meChanIsm (MDA) aNd soft-image Guidance (IFG)
  arXiv  Detail & Related papers  (2024-11-21T16:33:30Z)
• 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)
• Debiasing Multimodal Large Language Models via Penalization of Language Priors [38.97645845493758]
  Multimodal Large Language Models (MLLMs) have become indispensable tools in computer vision and natural language processing.<n>Despite their advancements, our investigation reveals a noteworthy bias: the generated content is often driven more by the inherent priors of the underlying Large Language Models (LLMs) than by the input image.<n>We propose two simple, training-free strategies to rectify these biases and redirect the model's focus toward visual information.
  arXiv  Detail & Related papers  (2024-03-08T12:35:07Z)

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.