Retrieval Heads Meet Vision: Uncovering How VLMs Locate and Extract Visual Information

• URL: http://arxiv.org/abs/2608.27417v1
• Date: Thu, 27 Aug 2026 17:43:58 GMT
• Title: Retrieval Heads Meet Vision: Uncovering How VLMs Locate and Extract Visual Information
• Abstract summary: Vision-language models (VLMs) can locate an image region referred to by a text prompt and route the corresponding visual evidence to the output.<n>We introduce Visual Retrieval Heads (VRHs), a small subset of attention heads that are causally responsible for grounding text descriptions to image regions.<n>We show that scoring attention from output prediction tokens with a sum over the ground-truth referent region most reliably identifies causal heads.
• Score: 41.72898733384826
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Vision-language models (VLMs) can locate an image region referred to by a text prompt and route the corresponding visual evidence to the output, yet the internal mechanism behind this behavior is not understood. Inspired by retrieval heads in large language models, we ask whether VLMs contain an analogous mechanism for visual retrieval. We answer affirmatively by introducing Visual Retrieval Heads (VRHs), a small subset of attention heads (about 1.7-2.6%) that are causally responsible for grounding text descriptions to image regions. To find them, we recast existing head-scoring methods under a unified design space over query tokens, key aggregation, and cross-sample aggregation. We then show that scoring attention from output prediction tokens with a sum over the ground-truth referent region most reliably identifies causal heads. Across eleven VLMs and five referring-expression benchmarks, masking only the top 20 VRHs reduces grounding accuracy by up to 80 percentage points, while masking the same number of random heads has little effect. Beyond replicating the causal-sparse-universal triad established for text retrieval heads, VRHs exhibit several properties not previously reported: they generalize across visual reference tasks, remaining causal on attribute, spatial, counting, and visual-math benchmarks despite being discovered through bounding-box prediction; they are functionally specific, preserving output format while corrupting localization; and they are architecturally shared, transferring causally across VLMs that share an LLM backbone but differ in vision encoder, projector, and instruction tuning.

Related papers

• UniProbe: A Learnable Token-Level Hallucination Detector for Large VLMs using Multi-Structural Internal Representations [61.85916280854257]
  Large Vision-Language Models (LVLMs) achieve impressive visual reasoning and dialogue capabilities, yet frequently hallucinate content unsupported by the visual input.<n>We introduce textbfUniProbe, a lightweight, unified, learnable detector that models a frozen LVLM's heterogeneous computational trace from a single forward pass.
  arXiv  Detail & Related papers  (2026-08-11T12:01:59Z)
• UniRef-UAV: A Multimodal Benchmark for Universal Referring in UAV Imagery [69.87589293875112]
  Unmanned aerial vehicles (UAVs) increasingly rely on visual grounding capabilities to localize task-relevant targets.<n>Existing referring expression comprehension benchmarks and methods are largely built around text-only queries and single-object outputs.<n>We introduce emphUniversal Referring, a generalized UAV referring task that jointly expands the query modality and the output cardinality.
  arXiv  Detail & Related papers  (2026-07-09T09:09:31Z)
• Interpretable Zero-shot Referring Expression Comprehension with Query-driven Scene Graphs [18.414159451507153]
  Zero-shot referring expression comprehension (REC) aims to locate target objects in images given natural language queries.<n>Existing Vision-Language Models (VLMs) commonly address zero-shot REC by measuring feature similarities between textual queries and image regions.<n>We propose textbfSGREC, an interpretable zero-shot REC method leveraging query-driven scene graphs as structured intermediaries.
  arXiv  Detail & Related papers  (2026-03-26T04:05:30Z)
• Visual Self-Refine: A Pixel-Guided Paradigm for Accurate Chart Parsing [76.2602505940467]
  Existing models often struggle with visually dense charts, leading to errors like data omission, misalignment, and hallucination.<n>Inspired by the human strategy of using a finger as a visual anchor'' to ensure accuracy when reading complex charts, we propose a new paradigm named Visual Self-Refine (VSR)<n>The core idea of VSR is to enable a model to generate pixel-level localization outputs, visualize them, and then feed these visualizations back to itself, allowing it to intuitively inspect and correct its own potential visual perception errors.
  arXiv  Detail & Related papers  (2026-02-18T13:40:53Z)
• IVC-Prune: Revealing the Implicit Visual Coordinates in LVLMs for Vision Token Pruning [27.75049214892312]
  Large Vision-Language Models (LVLMs) achieve impressive performance across multiple tasks.<n>A significant challenge, however, is their prohibitive inference cost when processing high-resolution visual inputs.<n>We propose textbfIVC-Prune, a training-free, prompt-aware pruning strategy that retains both IVC tokens and semantically relevant foreground tokens.
  arXiv  Detail & Related papers  (2026-02-03T03:39:31Z)
• SparseMM: Head Sparsity Emerges from Visual Concept Responses in MLLMs [74.2538340966038]
  We investigate how Multimodal Large Language Models (MLLMs) process visual inputs by analyzing their attention mechanisms.<n>We reveal a surprising sparsity phenomenon: only a small subset of attention heads in LLMs actively contribute to visual understanding.<n>We introduce SparseMM, a KV-Cache optimization strategy that allocates asymmetric computation budgets to heads in LLMs based on their visual scores.
  arXiv  Detail & Related papers  (2025-06-05T17:59:55Z)
• Expand VSR Benchmark for VLLM to Expertize in Spatial Rules [11.320245739677826]
  Visual spatial reasoning is a basic part of human cognition which requires fine-grained perception on cross-instance.<n>There is still a lack of sufficient quantity and quality evaluation and optimization datasets for Vision Large Language Models(VLLMs) specifically targeting visual positional reasoning.<n>We found current VLLMs to exhibit a contradiction of over-sensitivity to language instructions and under-sensitivity to visual positional information.<n>To our knowledge, we expanded spatially positioned image data controllably using diffusion models for the first time and integrated original visual encoding.
  arXiv  Detail & Related papers  (2024-12-24T07:13:17Z)
• Rephrase, Augment, Reason: Visual Grounding of Questions for Vision-Language Models [59.05769810380928]
  Rephrase, Augment and Reason (RepARe) is a gradient-free framework that extracts salient details about the image using the underlying vision-language model. We show that RepARe can result in a 3.85% (absolute) increase in zero-shot accuracy on VQAv2, 6.41%, and 7.94% points increase on A-OKVQA, and VizWiz respectively.
  arXiv  Detail & Related papers  (2023-10-09T16:57:57Z)
• Zero-shot Visual Relation Detection via Composite Visual Cues from Large Language Models [44.60439935450292]
  We propose a novel method for zero-shot visual recognition: RECODE. It decomposes each predicate category into subject, object, and spatial components. Different visual cues enhance the discriminability of similar relation categories from different perspectives.
  arXiv  Detail & Related papers  (2023-05-21T14:40:48Z)
• Visual Spatial Reasoning [35.5155400193075]
  We present a dataset containing more than 10k natural text-image pairs with 66 types of spatial relations in English. We show how the dataset includes challenging linguistic phenomena, such as varying reference frames. We demonstrate a large gap between human and model performance: the human ceiling is above 95%, while state-of-the-art models only achieve around 70%.
  arXiv  Detail & Related papers  (2022-04-30T23:03:49Z)

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.