The paper investigates why multimodal large language models (MLLMs) struggle with vision‑centric tasks when visual evidence conflicts with pretrained language knowledge. Using image reconstruction and a new WhatIfVis benchmark, the authors show that MLLMs preserve coarse‑grained visual attributes but fail to consistently use them, and that supervised fine‑tuning and activation patching can improve controllability of visual context sensitivity. The study demonstrates that the main bottleneck lies in the models’ inability to reliably regulate their reliance on visual evidence rather than in visual perception itself.
By Jiaang Li, Chengzu Li, Zhaochong An, Yifei Yuan, Xi Liu, Serge Belongie, V\'esteinn Sn{\ae}bjarnarson
The paper investigates a failure mode in multimodal large language models called multimodal contextual sycophancy, where external text can override conflicting image evidence. A diagnostic set of 998 cases independently varies visual evidence, commonsense priors, and external text to probe when this failure occurs. Experiments across six models show that a System‑2 Visual Arbitration (S2VA) approach, which withholds text from the visual witness, significantly improves performance over direct witness reports, with the best information boundary varying by model and context source.
By Yi-Cheng Lai, Hen-Hsen Huang
The paper introduces Visual Retrieval Heads (VRHs), a small fraction of attention heads in vision‑language models that are causally responsible for grounding text descriptions to image regions. By recasting head‑scoring methods and evaluating across eleven VLMs and five benchmarks, the authors show that masking the top 20 VRHs can drop grounding accuracy by up to 80 percentage points, while random masking has little effect. VRHs generalize across various visual reference tasks, preserve output format while corrupting localization, and transfer causally across models sharing an LLM backbone.
By Chanho Park, Daehyeon Choi, Jihyun Lee, Minhyuk Sung
arXiv:2606.18974v3 Announce Type: replace
Abstract: Unified multimodal models (UMMs) interleave generated ''visual thoughts'' (VTs) with text reasoning to improve spatial tasks. This incurs roughly a...
By Pengyu Li, Zhitao Gao, Lingling Zhang, Muye Huang, Yuanming Li, Fangzhi Xu, Jun Liu
arXiv:2608.21819v1 Announce Type: cross
Abstract: Reliable image captioning in Vision-Language Models (VLMs) requires captions to be both precise and complete, avoiding unsupported object mentions wh...
By Jihyung Ko, Eunji Jung, Hyeongsub Kim, Ziseok Lee, Jae Won Cho, Sanghyun Jo, Kyungsu Kim
The paper introduces a free, label‑free visual evidence signal that improves fine‑grained vision‑language reasoning. By selecting image crops that maximize the model’s answer distribution peak, the method locates answer‑bearing regions without training or annotations, boosting accuracy from 70 % to 85 %. The evidence gap also complements model confidence, enabling better correctness prediction and error flagging.
By Santi Ram Tiwari, Nihal Naik, Devbrat Pandey, Nishant Sinha
The paper investigates how general vision‑language models (VLMs) develop specialized optical character recognition (OCR) capabilities. By applying a causal intervention protocol, the authors identify sparse, stable OCR‑head sets in several VLMs and show that these heads largely overlap with textual retrieval/copy heads found in general VLMs. The study concludes that full‑sequence OCR functions as a dense multimodal copy‑and‑paste mechanism, and that when a VLM is fine‑tuned for OCR, it largely preserves the same head identities while redistributing their functional and causal strengths.
By Yuanxiang Huangfu, Hanmeng Zhong, Linqing Chen, Jeffrey Tiong Jee Hui
arXiv:2509. 22415v3 Announce Type: replace-cross Abstract: Multimodal large language models (MLLMs) have achieved strong vision-language performance, yet their token-level visual evidence remains difficult to inspect.
By Jiawei Liang, Jianjie Huang, Ruoyu Chen, Xianghao Jiao, Siyuan Liang, Shiming Liu, Xiaochun Cao
The paper introduces MIME, a task‑agnostic visual in‑context learning (vICL) dataset that removes textual shortcuts by omitting explicit rules linking demonstration images to answers, forcing models to infer the rule visually. Fine‑tuning vision‑language models on MIME boosts vICL accuracy by 8.1–18.0 points across three VLMs while preserving single‑image performance and improving multi‑image and text‑only in‑context learning. The authors also propose DTAD, a technique that randomly drops demonstration labels during training to encourage visual inference, achieving comparable accuracy with more consistent attention to demonstration images and outperforming frontier models on the TrueMICL benchmark.
By Nimrod Shabtay, Asaf Yehudai, Avihu Dekel, Eli Schwartz, Raja Giryes
Multimodal LLMs can see a document, but they often can't read it reliably. Small text, tables, visual cues, and topological elements still trip them up under direct visual inference, even when the pag...
The paper demonstrates that vision‑language models (VLMs) possess a small set of attention heads, called Visual Retrieval Heads (VRHs), that are causally responsible for linking text prompts to specific image regions. By adapting head‑scoring techniques from language models, the authors identify VRHs as the heads whose attention from output prediction tokens, summed over the ground‑truth referent region, most reliably indicates causal grounding. Experiments across eleven VLMs and five referring‑expression benchmarks show that masking the top 20 VRHs can drop grounding accuracy by up to 80 percentage points, while random masking has little effect, and that VRHs generalize across diverse visual tasks and transfer across models sharing an LLM backbone.
The paper introduces EviSpec, a training‑free compiler that generates complementary evidence specifications to improve high‑resolution multimodal large language models (MLLMs). By explicitly guiding visual search with structured evidence specifications, EviSpec achieves significant relative gains—up to 14.8% over random evidence—across five MLLMs and three benchmarks, and also sets new state‑of‑the‑art results on VQA and hallucination‑focused tasks.
By Zhongkuan Mao, Wenzhuo Zhao, Xianjie Liu, Yidong Wang, Zhao Gao, Ronghao Xian, Yao Jiang, Yi Zhang, Liangjian Wen, Keren Fu