Transcoders Trace Visual Grounding and Hallucinations in Vision-Language Models

• URL: http://arxiv.org/abs/2605.22902v1
• Date: Thu, 21 May 2026 17:34:39 GMT
• Title: Transcoders Trace Visual Grounding and Hallucinations in Vision-Language Models
• Abstract summary: Generative Vision-Language Models (VLMs) perform well on multimodal reasoning, but how visual inputs are transformed to text remains poorly understood.<n>We adopt a function-centric framework based on Transcoders, sparse approximations of sublayers that act as a causal proxy for layer-wise.<n>We show that function-centric circuit decomposition yields interpretable and predictive accounts of multimodal computation inVLMs.
• Score: 7.425292797106273
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Generative Vision-Language Models (VLMs) perform well on multimodal reasoning, but how visual inputs are transformed to text remains poorly understood. Existing interpretability work on VLMs uses Sparse Autoencoders (SAEs), which decompose static residual representations and miss the functional updates that drive cross-modal interaction. We adopt a function-centric framework based on Transcoders, sparse approximations of MLP sublayers that act as a causal proxy for layer-wise computation. Applied to Gemma 3-4B-IT, the framework decomposes the model into interpretable computational pathways linking image patches to directions in token generation. Transcoder attributions produce stronger and more stable effects on visually grounded tokens under patch ablation than SAE attributions, and align better with semantically relevant image regions. A False Visual Grounding counterfactual analysis confirms that the recovered pathways are specific to vision-language interaction.Finally, we perform a structural analysis of hallucinated generations, by extracting graph-based indicators from circuit traces produced by the transcoders. A logistic classifier over these mechanistic graph features predicts hallucinations at AUC $0.68$. These results show that function-centric circuit decomposition yields interpretable and predictive accounts of multimodal computation in VLMs.

Related papers

• Compressing AI Traffic: Standardized Neural Network Coding of Visual-Token Representations in Split Vision-Language Inference [15.357950338341285]
  Video-MME accuracy remains close to the uncompressed reference up to a 98% reduction of the transmitted BF16 tensor and only then collapses.<n>Open-ended MLVU generation shows the same plateau-and- robustness profile under an LLM judge.
  arXiv  Detail & Related papers  (2026-09-01T13:07:32Z)
• 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)
• DifFRACT: Diffusion Feature Reconstruction and Attribution for Circuit Tracing [8.306039957986394]
  We extend transcoder-based circuit tracing to multimodal diffusion transformers.<n>We find mechanisms underlying attribute binding and cross-stream semantic propagation.<n>Our results provide causal explanations for systematic generation errors.
  arXiv  Detail & Related papers  (2026-06-14T13:02:44Z)
• OneDrive: Unified Multi-Paradigm Driving with Vision-Language-Action Models [69.2503510410147]
  We present a unified autonomous driving framework built upon a pretrained VLM.<n>We demonstrate that pretrained VLM attention exhibits strong transferability beyond pure language modeling.<n>Experiments on end-to-end autonomous driving benchmarks demonstrate state-of-the-art performance.
  arXiv  Detail & Related papers  (2026-04-20T07:50:00Z)
• From Drop-off to Recovery: A Mechanistic Analysis of Segmentation in MLLMs [50.185593677108436]
  Multimodal Large Language Models (MLLMs) are increasingly applied to pixel-level vision tasks, yet their intrinsic capacity for spatial understanding remains poorly understood.<n>We investigate segmentation capacity through a layerwise linear probing evaluation across the entire MLLM pipeline: vision encoder, adapter, and LLM.
  arXiv  Detail & Related papers  (2026-03-18T00:22:15Z)
• ForgeryVCR: Visual-Centric Reasoning via Efficient Forensic Tools in MLLMs for Image Forgery Detection and Localization [62.03035862528452]
  ForgeryVCR is a framework that materializes imperceptible traces into explicit visual intermediates via Visual-Centric Reasoning.<n>ForgeryVCR achieves state-of-the-art (SOTA) performance in both detection and localization tasks.
  arXiv  Detail & Related papers  (2026-02-15T11:14:47Z)
• Med-CRAFT: Automated Construction of Interpretable and Multi-Hop Video Workloads via Knowledge Graph Traversal [13.216513001286812]
  textbfPipelineName is a novel neuro-symbolic data engineering framework.<n> Med-CRAFT extracts structured visual primitives from raw video streams and instantiates them into a dynamic Spatiotemporal Knowledge Graph.<n>We instantiate this pipeline to produce M3-Med-Auto, a large-scale medical video reasoning benchmark.
  arXiv  Detail & Related papers  (2025-11-30T19:24:10Z)
• Hybrid CNN-ViT Framework for Motion-Blurred Scene Text Restoration [2.0855516369698845]
  We introduce a hybrid deep learning framework that combines convolutional neural networks (CNNs) with vision transformers (ViTs)<n>The architecture employs a CNN-based encoder-decoder to preserve structural details, while a transformer module enhances global awareness through self-attention.<n>We show that the proposed method attains 32.20 dB in PSNR and 0.934 in SSIM, while remaining lightweight with 2.83 million parameters and an average inference time of 61 ms.
  arXiv  Detail & Related papers  (2025-11-08T17:48:58Z)
• Patch-as-Decodable-Token: Towards Unified Multi-Modal Vision Tasks in MLLMs [88.68484904214142]
  We introduce Patch-as-Decodable Token (PaDT) to generate both textual and diverse visual outputs.<n>Central to PaDT are Visual Reference Tokens (VRTs), derived from visual patch embeddings of query images.<n>We show PaDT consistently achieves state-of-the-art performance, even compared with significantly larger MLLM models.
  arXiv  Detail & Related papers  (2025-10-02T12:23:57Z)
• Hyperdimensional Probe: Decoding LLM Representations via Vector Symbolic Architectures [12.466522376751811]
  Hyperdimensional Probe is a novel paradigm for decoding information from the Large Language Models vector space.<n>It combines ideas from symbolic representations and neural probing to project the model's residual stream into interpretable concepts.<n>Our work advances information decoding in LLM vector space, enabling extracting more informative, interpretable, and structured features from neural representations.
  arXiv  Detail & Related papers  (2025-09-29T16:59:07Z)
• Interpreting vision transformers via residual replacement model [8.97847158738423]
  How do vision transformers (ViTs) represent and process the world?<n>This paper addresses this long-standing question through the first systematic analysis of 6.6K features across all layers.<n>We introduce the residual replacement model, which replaces ViT computations with interpretable features in the residual stream.
  arXiv  Detail & Related papers  (2025-09-22T07:00:57Z)
• Discrete Diffusion VLA: Bringing Discrete Diffusion to Action Decoding in Vision-Language-Action Policies [62.653984010274485]
  Vision-Language-Action (VLA) models adapt large vision-language backbones to map images and instructions into robot actions.<n> prevailingAs either generate actions auto-regressively in a fixed left-to-right order or attach separate or diffusion heads outside the backbone.<n>We present Discrete Diffusion VLA, a unified-transformer policy that models discretized action chunks with discrete diffusion.
  arXiv  Detail & Related papers  (2025-08-27T17:39:11Z)
• VISIT: Visualizing and Interpreting the Semantic Information Flow of Transformers [45.42482446288144]
  Recent advances in interpretability suggest we can project weights and hidden states of transformer-based language models to their vocabulary. We investigate LM attention heads and memory values, the vectors the models dynamically create and recall while processing a given input. We create a tool to visualize a forward pass of Generative Pre-trained Transformers (GPTs) as an interactive flow graph.
  arXiv  Detail & Related papers  (2023-05-22T19:04:56Z)
• Hybrid Routing Transformer for Zero-Shot Learning [83.64532548391]
  This paper presents a novel transformer encoder-decoder model, called hybrid routing transformer (HRT) We embed an active attention, which is constructed by both the bottom-up and the top-down dynamic routing pathways to generate the attribute-aligned visual feature. While in HRT decoder, we use static routing to calculate the correlation among the attribute-aligned visual features, the corresponding attribute semantics, and the class attribute vectors to generate the final class label predictions.
  arXiv  Detail & Related papers  (2022-03-29T07:55:08Z)

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.