Arbitration Failure, Not Perceptual Blindness: How Vision-Language Models Resolve Visual-Linguistic Conflicts

• URL: http://arxiv.org/abs/2604.09364v2
• Date: Mon, 13 Apr 2026 13:46:40 GMT
• Title: Arbitration Failure, Not Perceptual Blindness: How Vision-Language Models Resolve Visual-Linguistic Conflicts
• Abstract summary: We show that visual attributes can be linearly decodable from early layers.<n>We show that training-free activation in early layers can improve visual grounding by up to +3.8% with degrading performance in some setups.
• Score: 7.071615105094556
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: When a Vision-Language Model (VLM) sees a blue banana and answers "yellow", is the problem of perception or arbitration? We explore the question in ten VLMs with various sizes and reveal an Encoding-Grounding Dissociation: models that fail to report what they see (and thus provide a wrong answer) still encode the visual evidence as strongly as models that provide the correct answer. Using Multimodal Arbitration Crossover (MAC) analysis with layer-by-layer Logit Lens probing, we track the competition between visual and prior signals across every layer of each model. We show that visual attributes can be linearly decodable from early layers (AUC > 0.86). The accuracy remains nearly identical for both successful and failed samples. However, the gap in the final-layer logit - not the strength of encoding - better predicts grounding outcomes with a correlation of $ρ=$ 0.847. After having studied when VLMs base their answers on image clues rather than prior knowledge, we want to understand the causal relationships. We establish causality through full-sequence activation patching. The standard last-token interventions in LLM interpretability do not affect VLMs. In contrast, replacing the full token sequence at layers identified by MAC alters 60 to 84% of outputs. Partial-token decomposition shows that image tokens carry almost all of the causal impact, while text tokens have none. Scaling addresses the remaining architectural differences to achieve perfect retention. Moving from diagnosis to intervention, we show that training-free activation steering - both linear and sparse autoencoder-guided - in early layers can improve visual grounding by up to +3.8% with degrading performance in some setups. Overall, these findings lead to a clear conclusion: VLMs already see well, but the challenge is acting on what they see. Targeted interventions can help to bridge this gap.

Related papers

• STAR-Pro: Stage-Wise Token Adaptive Reduction with Progressive Refinement for Efficient Large Vision-Language Models [54.273344678561024]
  Large vision-language models (LVLMs) achieve strong multimodal understanding, but the hundreds to thousands of visual tokens they process impose computational overhead.<n>We conduct two complementary analyses of visual token pruning.<n>We propose STAR-Pro, a training-free two-stage framework for visual token pruning.
  arXiv  Detail & Related papers  (2026-09-05T06:18:28Z)
• The Illusion of Visual Tool-Use: A Causal Audit of Thinking with Images [19.417504594148223]
  "thinking-with-images" operations often achieve only marginal or negative gains over direct inference at substantially higher token cost.<n>We ask whether the returned visual evidence causally affects the answer.<n>We use visual tool-use as a causal graph that separates observation-mediated paths from action-induced shortcuts.
  arXiv  Detail & Related papers  (2026-08-06T17:01:08Z)
• VisualFLIP: Do Predictions Depend on Task-Critical Visual Evidence in Multimodal Reasoning? [79.44137632338062]
  When a multimodal large language model answers a visual reasoning question correctly, is the prediction actually supported by the task-critical visual evidence?<n>We introduce VisualFLIP, a paired benchmark with 1,374 images arranged as same-question pairs across cardinality, attribute, spatial, and logic tasks.<n>We evaluate 24 MLLMs with pair accuracy, which requires solving both sides of a pair, and Collapse Rate (CR), which measures how often a model that solves at least one side repeats the same non-empty answer for both images.
  arXiv  Detail & Related papers  (2026-06-05T22:06:07Z)
• From Early Encoding to Late Suppression: Interpreting LLMs on Character Counting Tasks [49.57538588967748]
  Large language models (LLMs) exhibit failures on elementary symbolic tasks such as character counting in a word, despite excelling on complex benchmarks.<n>We uncover a consistent phenomenon across modern architectures, including LLaMA, Qwen, and Gemma.<n>We show that symbolic reasoning failures in LLMs are not due to missing representations or insufficient scale, but arise from structured interference within the model's computation graph.
  arXiv  Detail & Related papers  (2026-04-01T11:40:12Z)
• All You Need Are Random Visual Tokens? Demystifying Token Pruning in VLLMs [43.80391827200227]
  In deep layers, existing training-free pruning methods perform no better than random pruning.<n>Visual tokens progressively lose their salience with increasing network depth.<n>We show that simple random pruning in deep layers efficiently balances performance and efficiency.
  arXiv  Detail & Related papers  (2025-12-08T14:16:01Z)
• Reading Between the Lines: Abstaining from VLM-Generated OCR Errors via Latent Representation Probes [79.36545159724703]
  We propose Latent Representation Probing (LRP) to train lightweight probes on hidden states or attention patterns.<n>LRP improves abstention accuracy by 7.6% over best baselines.<n>This establishes a principled framework for building deployment-ready AI systems.
  arXiv  Detail & Related papers  (2025-11-25T00:24:42Z)
• Beyond Single Models: Mitigating Multimodal Hallucinations via Adaptive Token Ensemble Decoding [41.828387997311474]
  Large Vision-Language Models (LVLMs) have recently achieved impressive results in multimodal tasks such as image captioning and visual question answering.<n>They remain prone to object hallucination -- generating descriptions of nonexistent or misidentified objects.<n>We propose Adaptive Token Ensemble Decoding (ATED), a training-free, token-level ensemble framework that mitigates hallucination by aggregating predictions from multiple LVLMs during inference.
  arXiv  Detail & Related papers  (2025-10-21T06:11:24Z)
• Diffusion Language Models Know the Answer Before Decoding [56.96815863705218]
  Diffusion language models (DLMs) have emerged as an alternative to autoregressive approaches.<n>Our work highlights and leverage an overlooked property of DLMs early answer convergence.<n>We introduce Prophet, a training-free fast decoding paradigm that enables early commit decoding.
  arXiv  Detail & Related papers  (2025-08-27T15:40:25Z)
• Can VLMs Recall Factual Associations From Visual References? [30.821053378797007]
  We identify a systematic deficiency in the multimodal grounding of Vision Language Models (VLMs)<n>Forcing VLMs to rely on image representations of an entity halves their ability to recall factual knowledge.<n>We show that such linking failures are correlated with the expression of distinct patterns in model internal states.
  arXiv  Detail & Related papers  (2025-08-22T16:47:37Z)
• Token Activation Map to Visually Explain Multimodal LLMs [23.774995444587667]
  We propose an estimated causal inference method to mitigate the interference of context to achieve high-quality MLLM explanation.<n>We term this method Token Activation Map (TAM) to highlight the consideration of interactions between tokens.<n>Our TAM method significantly outperforms existing SoTA methods, showcasing high-quality visualization results.
  arXiv  Detail & Related papers  (2025-06-29T14:50:45Z)
• Inference Optimal VLMs Need Fewer Visual Tokens and More Parameters [54.01228554126122]
  Vision Language Models (VLMs) have demonstrated strong capabilities across various visual understanding and reasoning tasks.<n>To reduce inference costs, one can either downsize the Large Language Models (LLMs) or reduce the number of input tokens needed to represent the image.<n>We take the first steps toward designing token compression algorithms tailored for high-compression settings.
  arXiv  Detail & Related papers  (2024-11-05T18:54:21Z)
• Implicit Multimodal Alignment: On the Generalization of Frozen LLMs to Multimodal Inputs [63.29737699997859]
  Large Language Models (LLMs) have demonstrated impressive performance on multimodal tasks, without any multimodal finetuning. In this work, we expose frozen LLMs to image, video, audio and text inputs and analyse their internal representation.
  arXiv  Detail & Related papers  (2024-05-26T21:31:59Z)
• CogCoM: A Visual Language Model with Chain-of-Manipulations Reasoning [61.21923643289266]
  Chain of Manipulations is a mechanism that enables Vision-Language Models to solve problems step-by-step with evidence.<n>After training, models can solve various visual problems by eliciting intrinsic manipulations (e.g., grounding, zoom in) actively without involving external tools.<n>Our trained model, textbfCogCoM, achieves state-of-the-art performance across 9 benchmarks from 4 categories.
  arXiv  Detail & Related papers  (2024-02-06T18:43:48Z)

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.