The paper introduces an interventional protocol to assess how vision‑language models (VLMs) explain the impact of missing modalities on their predictions. By comparing the models’ self‑explanations with actual changes observed after restoring missing inputs, the study finds that VLMs routinely overstate the sufficiency of available evidence and underestimate the effect of adding back missing modalities. Across eight open‑weight VLMs and four tasks, the discrepancy between predicted and realized changes is substantial, revealing systematic mischaracterization of modality dependence.
By Aydin Javadov, Daniel Schoess, Florian von Wangenheim
arXiv:2604. 14888v3 Announce Type: replace-cross Abstract: Recent advances in vision language models (VLMs) offer reasoning capabilities, yet how these unfold and integrate visual and textual information remains unclear.
By Danae S\'anchez Villegas, Samuel Lewis-Lim, Nikolaos Aletras, Desmond Elliott
The paper shows that controlled visual in‑context learning can help vision‑language models (VLMs) overcome errors caused by training priors. By providing demonstrations, the models focus more on visual evidence, yielding significant accuracy gains on counterfactual examples while leaving prior‑aligned performance largely unchanged. This effect is observed across five VLM families and five vision‑centric tasks, indicating that many VLMs possess latent visual grounding abilities that standard tests may miss.
By Kun Wang, Xindi Wu, Sanghyuk Chun, Olga Russakovsky, Esin Tureci
arXiv:2609. 31450v1 Announce Type: cross Abstract: Medical vision-language model (VLM) post-training is commonly evaluated through answer accuracy.
By Wang Jingxin
Aligned vision‑language models (VLMs) are designed to combine grounded visual reasoning with safe generation. The study finds that when safety constraints are applied, these models often abstain from answering questions that they could answer under default instruction, yet visual evidence continues to influence the decoding process. The authors show that safety‑induced abstention alters late‑stage hidden‑state dynamics, and that targeted interventions can restore grounded answering without retraining or changing visual inputs.
The study evaluates whether signals from a frozen vision‑language model, InstructBLIP, can reliably indicate answerability in visual question answering. Using a 10K subset of UNK‑VQA, the authors test predictive uncertainty, cross‑modal alignment, visual grounding, visual dependence, internal representations, and learned probes, finding that all signals yield weak discrimination (AUROC 0.50‑0.54). Even interventions that remove question‑relevant visual evidence or increase probe capacity do not substantially improve answerability detection, underscoring that visual dependence and related metrics are not robust proxies for answerability.
By Minjun Kim, Hyunmin Song, Changwon Lim
The paper investigates how safety‑aligned large language models (LLMs) can still be tricked by jailbreaks or over‑refuse benign requests. By analyzing residual‑stream updates in decoder‑only Transformers, the authors find that harmfulness is more readily decoded from attention updates—especially in early layers—while the MLP component is primarily responsible for writing refusal responses. Post‑training enhances MLP‑driven refusal writing without improving harmfulness decodability, and causal interventions confirm that downstream MLP updates are crucial for the refusal shift caused by upstream attention changes. Building on these insights, the authors propose Sentinel, a system that combines an attention‑derived perception score with an MLP‑derived decision score to better detect jailbreaks and over‑refusals across multiple models.
By Zixuan Weng, Haoqi He, Weifei Jin, Yue Huang, Yang Bai
The paper investigates how visual interventions—training‑free edits applied to multimodal large language models—affect hallucinations and prediction quality. By applying a visual‑residual edit across successive decoder layers and matching its RMS strength, the authors identify an Evidence Anchoring Boundary (EAB) where the intervention’s influence drops sharply; effects before the EAB are 13.78–26.10× larger than after it. They further compare counterfactual subtraction across different question types and introduce calibrated selective routing, which only alters predictions when calibration supports an intervention, improving attribute accuracy by 1.56–1.81 percentage points while avoiding the harms of universal subtraction.
By Qiji Zhou, Junyi Hu, Peiyu Wang, Yue Zhang
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
arXiv:2603.22161v3 Announce Type: replace
Abstract: Metacognition -- assessing the quality of one's own cognitive performance -- guides adaptive behavior across species. Substantial research demonstr...
By Dharshan Kumaran, Nathaniel Daw, Simon Osindero, Petar Veli\v{c}kovi\'c, Viorica Patraucean
The paper investigates how language‑model agents use external procedural skills and discovers a bias called Procedural Exposure Bias (PEB), where repeated exposure to a skill’s pattern makes the model favor its associated action even when the context is unchanged. PEB is much weaker when the skill is presented visually rather than textually, mainly due to visual encoding rather than layout. Experiments across multiple environments show that visual skills lead to higher task success than textual ones, highlighting the importance of skill representation for agent decision‑making.
By Qing Yang, Yuqi Xiong, Pengcheng Huang, Zhipeng Xu, Zhenghao Liu, Zulong Chen, Yu Gu, Ge Yu
arXiv:2608.23313v1 Announce Type: new
Abstract: Vision-language model safety benchmarks typically evaluate only final responses: whether a model refuses, warns, or complies. This outcome-level view c...
By Xuetong Li, Gaofeng Liu