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.
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.
The paper investigates whether the tokens highlighted by large language models (LLMs) as key to their predictions are behaviorally consistent when those tokens are manipulated. Using sufficiency, comprehensiveness, and counterfactuality metrics, the authors evaluate closed‑source GPT‑4o‑mini and open‑source models Gemma3, Granite8B, and DeepSeek on IMDB and Steam review datasets. Results show that GPT‑4o‑mini consistently follows expected patterns, Gemma3 and Granite8B maintain sufficiency but falter under stricter interventions, and DeepSeek variants exhibit structural inconsistencies, indicating that explanation reliability varies across model families and sizes.
The paper investigates how multimodal large language models decide to refuse answering visually unanswerable questions. It examines three control mechanisms—instruction phrasing and placement, post‑training methods, and internal activation interventions—across three models on the MoHoBench benchmark. The study finds that refusal behavior is highly sensitive to instruction details, can be amplified by post‑training (sometimes leading to mixed refusal and answer attempts), and can be modestly altered by activation interventions without causing indiscriminate refusals.
CounterCount is a diagnostic dataset that pairs factual and counterfactual images to test whether Vision‑Language Models (VLMs) rely on visual evidence or on world knowledge when counting attributes. Humans achieve near‑perfect accuracy on both types of images, whereas VLMs drop from 91.6% to 49.5% on counterfactuals, often defaulting to canonical counts. The study shows that VLMs exhibit a System‑1‑like bias, underutilize visual evidence, and that targeted input manipulations can mitigate this bias.
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.
arXiv:2603.04419v3 Announce Type: replace-cross Abstract: Vision-language models produce different object and use descriptions under different persona prompts, but low overlap alone does not identify...
The paper introduces Synergistic Faithfulness (𝔽_syn), a new metric based on the Shapley Interaction Index that isolates the joint contribution of visual and textual modalities in Vision‑Language Models (VLMs). It demonstrates that existing explainers overemphasize visual salience and fail to capture true cross‑modal synergy, and provides a scalable evaluation framework for auditing VLM reasoning. The authors evaluate nine XAI methods across three VLM architectures and datasets, revealing significant shortcomings in current approaches.
The paper introduces CAIT, a benchmark of 400 synthetic scenes featuring counter‑intuitive actions that challenge multimodal large language models (MLLMs). Human participants and proprietary models like Claude and Gemini perform well, but standard open‑source instruction‑tuned MLLMs fail, largely due to a strong language prior that overrides contradictory visual evidence. The study shows that Chain‑of‑Thought reasoning can help but introduces new issues, while targeted fine‑tuning and structured prompting can reduce reliance on language priors and improve visual grounding.
arXiv:2608. 00076v2 Announce Type: replace-cross Abstract: Multimodal large language models (MLLMs) increasingly support high-stakes decision making by combining complementary information from images and text.
The paper introduces IMAVB, a 500‑clip benchmark that tests whether omnimodal large language models can detect when a textual premise contradicts their visual or audio input. Experiments on eight open‑source models and Gemini 3.1 Pro reveal a Representation‑Action Gap: internal states encode mismatches, yet the models rarely reject false premises, exhibiting under‑rejection or over‑rejection. A probe‑guided logit adjustment improves rejection behavior, suggesting the main bottleneck is in translating perception to action rather than in perception itself.
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.
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.