Visual Evidence Sensitivity in VLMs

Physical reasoning requires adapting familiar expectations to the dynamics of the observed world. Existing benchmarks establish broad physical capabilities, yet an overall score leaves unclear whether answers follow visual evidence when familiar rules no longer apply. We introduce InfiniteWorld, a benchmark that pairs physical-reasoning questions with simulator videos whose physical parameters are systematically altered. Its generation pipeline connects simulation, scene screening, paired video translation, question construction, and input checks. Five tasks assess law identification, parameter estimation, temporal reasoning, analytic counterfactual comparison, and law induction. The evaluation defines five matched input settings for each question, comprising text-only evaluation and ordered or shuffled observations from simulator and translated videos. We evaluate a diverse set of video-language models under these matched settings. The results expose a persistent preference for familiar physical rules. Frames improve performance for most models, but gains differ across tasks. Strong accuracy on Earth-normal questions coexists with poor recognition of altered dynamics, and Earth-normal answers remain common when the observed physics changes. Task scores and semantic agreement further distinguish matching physical factors from matching their amount, and correct answers from repeated errors. InfiniteWorld makes these behaviors measurable through paired changes in the evidence supplied for the same question, separating visual evidence from familiar physics.

Multimodal large language models (MLLMs) read images through visual tokens, and under the small per-image budgets that efficient deployment requires, the fine text of documents, charts and infographics becomes illegible. Zooming in recovers the evidence, but the region must be chosen before the model answers. We find that MLLMs know where the evidence is, but not in a form that can be read before they answer: the answer tokens localize it from the middle layers on, whereas the attention available before answering does so only in the top layers, at a depth that varies across models, and diffusely. Supervising the pre-answer attention to point degrades reading, and the benefit of zooming is governed by each image's information gap. Building on these findings, we introduce ours, which lets an MLLM look where its answer would. A dedicated look token, processed in a single prefill and kept out of the answering pass, learns in hindsight to attend as the model's own correct answers do, extracting this knowledge into a concentrated pointer before any answer exists. Its hidden state then selects how to look, from no zoom to crops, square windows, windows that adapt to the pointer and foveated warps, supervised by the counterfactual outcome of every view. ours needs no box annotations, per-model layer search, external localizers or generation before the answer. On Qwen3-VL at three scales and five text-rich benchmarks, it surpasses the strongest baseline by 4.8-9.1 points at 98 visual tokens per view and attention-guided zooming by 11.1-17.0 points, while running faster than attention-guided and grounding-based zooming.

Effective multimodal reasoning requires a model to reliably utilize visual information while sufficiently exploring diverse reasoning paths. To achieve this goal, prior work combines visual dependency and token entropy to reshape sequence-level advantages, but scaling the contribution of token entropy by visual dependency can leave high-entropy tokens with low visual dependency underweighted, despite their important role in reasoning exploration. To address this limitation, we propose PEAR, an adaptive advantage reshaping method for improving perception and exploration in multimodal reasoning. PEAR retains visual dependency as the base signal and adds an entropy-based exploration gain only at high-entropy positions, with lower visual dependency leading to a larger additional gain. The fused signal is then mapped to positive token-level weights that reshape the sequence-level advantage. Experiments on Qwen2.5-VL-3B/7B show that PEAR achieves the highest average performance among evaluated baselines across multimodal reasoning benchmarks, alongside leading results on textual reasoning and general multimodal benchmarks.

This paper studies the problem of test-time robustness scaling for large vision-language models (LVLMs), which aims to mitigate language bias during inference. Existing approaches typically adopt test-time scaling techniques such as contrastive decoding to improve inference performance. Despite the progress, the performance of these approaches is far from satisfactory due to intentional evidence removal during perturbations and redundant inference cost. Towards this end, we propose a novel approach named Margin-gated Contrastive Debiasing with Evidence Restoration (MAGIC). The core of our MAGIC is to evaluate the decision margins of LVLMs to selectively activate textual and visual perturbation branches for efficient debiased inference. In particular, our MAGIC first measures the prediction margins of crucial tokens, and applies a gating function to identify low-confidence responses that are susceptible to language biases. For selected samples, we integrate the token logits from multiple textual and image perturbation branches with the original ones for debiased decision. More importantly, we selectively restore the most informative corrupted views with sharp prediction distributions for ambiguous samples after debiasing, enabling evidence recovery for discriminative decision-making. We further utilize a calibration dataset to actively identify and prune ineffective branches during contrastive debiasing for cost reduction. Extensive experiments on benchmark datasets validate the effectiveness of the proposed MAGIC in comparison with extensive baselines.

Vision-language models (VLMs) struggle with in context learning compared to their textual counterparts. Our preliminary study shows that VLMs rely on textual shortcuts and underutilize the visual demonstrations. We therefore hypothesize that VLMs are insufficiently trained to perform visual in-context learning (vICL). To this end, we construct MIME, a task-agnostic vICL training set without textual shortcuts. We ensure the instruction does not state the rule linking the demonstration images to their answers, so the model must infer it from the visual demonstrations. Fine-tuning on MIME improves vICL accuracy by 8.1 to 18.0 points across three VLMs, preserves single-image performance, and improves multi-image and text only in-context learning. In contrast, stating the rule in the training data sharply reduces this gain. We further propose DTAD (Demonstration Text Attention Dropout), which randomly withholds demonstration labels during training so that the VLM must infer the rule from the demonstration images. While achieving comparable vICL accuracy to fine-tuning on MIME, DTAD yields more consistent attention to the demonstration images. On TrueMICL, Qwen3-VL-8B trained on MIME matches frontier models such as Claude Opus 5, and surpasses them with DTAD. Finally, adding MIME to the instruction-tuning mix of LLaVA 1.5 13B more than doubles its vICL accuracy without hurting single-image abilities, whereas a same-size multi-image set improves it only slightly. These results suggest that vICL data should be a standard component of VLM training.

Long chain-of-thought reasoning improves vision-language models (VLMs), but it also exposes a temporal grounding failure: as generation unfolds, models may drift from image evidence while reinforcement learning still provides only a final-answer reward. We study this gap through visual revisit, a temporally localized reactivation of image evidence during reasoning. Across three VLM backbones, revisit peaks are predictive of correctness and causally linked to performance: masking them degrades accuracy more than masking non-peaks, while injecting revisit-like peaks into failed rollouts partially restores correct answers. Control analyses show that this signal is not explained by raw attention strength, uncertainty, logit margin, or hidden-state norm. We propose RECAP, a GRPO-compatible training method that turns detached revisit traces into a credit-assignment signal. RECAP uses revisit-conditioned advantage estimation, dynamic discounting, and gated value prediction to propagate final rewards through visually important reasoning steps, without extra rewards, additional rollouts, or inference-time changes. Across Qwen2.5-VL-7B, InternVL3-8B, and LLaVA-OV-7B, RECAP consistently improves over GRPO and VPPO, yielding $+6.1$--$+6.5$ HallusionBench gains over the SFT base while preserving reasoning and text-only performance. Results on larger LoRA-tuned models, Chinese zero-shot benchmarks, visual perturbations, and human evaluation further demonstrate robust improvements in image-grounded reasoning.

Visual perception should be iterative and task-directed, with representations refined in response to the question and the need for further processing. While Vision-Language Models (VLMs) increasingly support extended reasoning, their visual tokens are still typically acquired in a single, task-agnostic encoder pass. This creates an evidence-acquisition bottleneck: models may reason longer without refining the visual evidence needed to answer the question. We propose LoopLook, a lightweight encoder-side recurrent perception framework that equips pretrained VLMs with adaptive visual depth. LoopLook recurrently revisits a middle-to-late window of the vision encoder, injecting question representations at each loop to guide visual refinement. A learned stopping mechanism adaptively determines the recurrent depth, while a token-wise Mixture-of-Looks (MoL) combines representations from different loops, allowing each visual token to selectively integrate information acquired across recurrent steps. Across four backbones and eight benchmarks, LoopLook consistently improves performance over the baseline, yielding gains of up to 8.6% on MMStar and 8.1% on HallusionBench. Our analysis shows that LoopLook progressively refines visual attention and adaptively allocates depth according to task difficulty. It also better preserves task-relevant visual information under reduced input resolution, while increasing time to first token (TTFT) by only 5.1-6.5%. By bringing recurrence from reasoning back into perception, LoopLook transforms visual perception from a one-shot glimpse into an adaptive process of recurrent refinement.

Multimodal knowledge editing aims to update visually grounded facts in Multimodal Large Language Models (MLLMs) while preserving unrelated text and visual behaviors. However, standard answer-level metrics can hide evidence misgrounding: an edit may produce the target answer on the original image yet remain sensitive to incidental context rather than the intended visual evidence. To examine this issue in a controlled setting, we construct WaterBird-Edit, an analysis set that decouples bird habitat labels from background context and measures whether edited predictions remain stable when the target evidence is preserved but the background changes. Motivated by this diagnosis, we propose VICE (Visual In-Context Editing), a retrieval-based framework that stores edits in external memory and applies them through visual in-context demonstrations designed to expose task-relevant evidence. Experiments on three standard benchmarks show that VICE achieves strong lifelong editing performance, while results on WaterBird-Edit show that it substantially reduces sensitivity to the controlled background shifts. Code and dataset are available in https://github.com/Wang-Z-W/VICE-main.

In this paper, we investigate why vision-language models (VLMs) produce errors by decomposing them into two sources: visual-representation failures, in which the visual evidence required for the correct answer is not encoded, and visual-utilization failures, in which this evidence is encoded but not successfully used. We propose to estimate the visual-utilization failure rate using lightweight probes and derive theoretical bounds on the true rate that account for probe imperfections. We introduce VELens (Visual Evidence Lens), a rank-based probe that traces visual evidence for both a target object and its ground-truth attribute, such as color, material, or shape, across VLM layers. We evaluate VELens on over 6,000 images using LLaVA-1.6-7B and Qwen2.5-VL-7B. Our method achieves ROC-AUC scores of 0.95 and 0.91, respectively, and outperforms alternative probes in both low-resource and unseen-attribute settings. Using it to diagnose errors, we estimate that visual-utilization failures account for 84.4% of LLaVA errors and 70.0% of Qwen errors, with conservative lower bounds of 77.5% and 61.0%, respectively. Overall, our study identifies visual utilization as the dominant bottleneck in VLMs and suggests that a central opportunity for improvement lies in developing better mechanisms to locate, route, and use the visual information they already encode.

Vision-language trackers (VLTs) accept descriptions intended to assist tracking, but accepting text does not establish that its content changes tracking decisions, particularly as visual evidence deteriorates. We introduce DISTRACT, a matched-intervention benchmark that holds video, target, and distractor fixed while crossing four text priors with four image degradations along a controlled severity axis. We evaluate five VLTs on four datasets using behavioral contrasts, layerwise representation probes, and activation patching. Molmo2 and JointNLT exhibit directed text use. Some effects contract under degradation, whereas others show no significant endpoint change after multiplicity correction. UVLTrack, MMTrack, and DUTrack respond only weakly to changes in text semantics, with no directed behavioral effect surviving FDR correction in the tested settings. Model-internal analyses reveal that at decision-near states, correct-versus-conflicting text is decodable in Molmo2 and JointNLT but not in the other three systems. In the responsive models, decodability peaks at late states where replacement produces little behavioral change, whereas replacing selected earlier states redirects conflict-associated trajectories toward the target. Because matched-correct donors do not significantly outperform unrelated donors, these interventions identify earlier intervention-sensitive states without establishing target-specific semantic rescue. DISTRACT thus exposes when and where text affects tracking under degraded visual evidence.

Vision-language models (VLMs) are exact on small sets and drift into estimation on larger ones, a profile reminiscent of children before they master counting. Children cross the barrier with epistemic actions: pointing at each item while reciting number words, moving counted items aside, partitioning the array. We give VLMs the same actions as dumb image tools with no visual intelligence of their own in a training-free agent loop. Our central claim is about measurement, not intervention: a model that counts by marking leaves a normally hidden serial procedure on the image, and the developmental principles of counting turn those marks into an interpretable audit of serial visual reasoning: every tag is classified against point-level ground truth as on-target, duplicate, phantom, or omitted, the answer is checked against the number of tags, and order irrelevance is tested by forcing opposite traversal orders. Across 9 models, 682 images, and 7 harness conditions we find that (i) the exact-counting range is a model-specific signature spanning nearly two orders of magnitude (R80 from 3 to 150 dots); (ii) as an intervention the finger mostly fails: tags help 4 of 9 models on small-count photographs, hurt imprecise pointers, and rarely beat a verbal routine or a blind point list; what the loop buys is an auditable trace, not accuracy; (iii) as measurement it succeeds: a tag-perturbation probe shows that precise pointers read their answers off the tags, pointing precision tracks which models benefit, dense scenes fail by omission, and models rarely revise after seeing their own tags; and (iv) the capacities dissociate: models follow an instructed traversal order yet often change their answer when it reverses, and children's regular-over-scattered advantage appears only under the routine and survives lattices that defeat a rows-times-columns shortcut.

Vision-language models (VLMs) can infer sensitive information such as geolocation from subtle visual cues, creating privacy risks when responses disclose more detail than the visual context warrants. We study visual-contextual privacy in geolocation disclosure, where a VLM should reveal location information only at the granularity appropriate to the visual context depicted in the image. We first show that this contextually appropriate granularity is recoverable from pre-generation hidden states, even when generated responses are misaligned, with disclosure levels exhibiting an ordered structure along a common hidden-state direction. Motivated by this, we propose an inference-time route-and-steer framework: an ordinal granularity router maps visual-contextual privacy representations to a disclosure route, and a disclosure-level-conditioned steering module applies route-specific hidden-state interventions during decoding. Across four VLM backbones, our method improves visual-context-aware disclosure alignment, better balances over- and under-disclosure, and incurs only modest changes on general capability benchmarks. It further transfers to another dataset with a distinct data distribution and privacy-risk taxonomy.

Vision--language models (VLMs) promise to make computational pathology queryable and able to draft reports, but how much language adds beyond the vision backbone is unclear, because almost every system is evaluated by its developers on test sets of their own choosing. We grade published evidence by how it was produced and test four claims with controlled probes that reuse released checkpoints and public data, varying one input at a time. Report text adds prognostic signal: adding a frozen embedding of the stage-redacted report to a slide embedding raises the pooled C-index by 0.019 for progression-free interval and by 0.016 for overall survival across 1{,}633 TCGA patients, and restoring the stage statement adds nothing. Prompts add no measurable accuracy at inference: with features and pooling fixed, zero-shot classification is statistically indistinguishable from a supervised linear probe for lung and renal cancer subtyping. Slide-level question answering is barely grounded: SlideChat answers 74.82% of SlideBench-VQA(TCGA) items correctly with the slide and 74.41% without it, whereas its language backbone alone reaches 43.52%. Paraphrasing leaves accuracy unchanged but flips 8--23% of answers. Self-reported gains from agentic navigation and few-shot language priors cannot yet be attributed to language. Language thus adds real but narrow value, and we propose minimum controls that would expose these failure modes in future evaluations.

When given an image and a text prompt, vision-language models (VLMs) can produce seemingly sophisticated text responses. However, they are also prone to hallucinate objects in their output, which can have detrimental consequences in safety-critical applications. To detect these hallucinations, certain methods interestingly leverage internal signals from the VLM as it generates the text response; potentially, these methods use the same signals that led to a hallucination to detect it. We conduct an in-depth analysis to understand the robustness of these detection methods. First, we show that a naive baseline of a generated word's token position in the VLM text response achieves competitive performance. Further, we reveal how some detection methods rely on token position and use this heuristic as a shortcut. Second, we explore detection methods' sensitivity to language and vision modalities. Our experiments suggest that strong detection performance is often driven by detecting language-based hallucinations, and when methods rely on token position, they tend to be less visually grounded and exhibit reduced robustness when hallucinations occur earlier in a text response. Lastly, we demonstrate that different detection methods can carry complementary signals; when ensembling methods together, adding weaker methods to an ensemble of higher performing ones still offers performance improvements. Altogether, while there may be useful internal VLM signals to detect hallucinations, our investigation reveals important limitations of the current methods that try to leverage them.

Medical vision-language models often exhibit hidden stratification, where high overall performance masks failures on subtle visual findings and overlapping clinical conditions. We trace this failure mode to modality imbalance during fine-tuning, where strong linguistic priors can suppress visually grounded learning and induce plausible but weakly grounded rationales. Existing data selection methods usually score image-text pairs as holistic units, without separating visually necessary samples from shortcut-solvable ones. To address this problem, we propose ViCoR, a counterfactual-residual-based data selection method for Med-VLM fine-tuning. ViCoR estimates visual necessity through image-dependent optimization residuals, calibrates the resulting scores with medical supervision compatibility, and constructs a diversity-aware fine-tuning subset. Across two backbones, four external generalization benchmarks, and one multi-disease robustness benchmark, ViCoR-selected 20% subsets achieve the strongest average performance compared with full-data fine-tuning and selection baselines. Gradient-level analysis and perturbation audits further show that ViCoR enriches image-dependent training signals and strengthens image-level and evidence-region dependence. The code will be available at https://github.com/htgan-ai/ViCoR.

Can a better robot policy become harder to monitor? We study a fixed vision-language model (VLM) that judges success while a policy improves. In a controlled 20-seed tray-placement experiment, training selects only simulator-verified successes. Success rises from 66.8% to 97.0%, and accepted hover failures fall from 5.0% to 0.9% of executions. Yet the judge accepts 50.0% of remaining hovers, versus 31.1% under untrained sampling. These residual failures more often lie below the tray rim, which can conceal the gap beneath a grasped cube. A separate boundary probe accepts 148/160 below-rim hovers and 0/160 above-rim hovers. We call this change in conditional monitoring sensitivity coverage shift: learning changes which failures reach a fixed judge. A broader diagnosis covers nine judges and 1,392 stimuli. Judge-filtered experiments across eight settings further show heterogeneous effects of state-dependent acceptance. The results distinguish task improvement from failure detectability and motivate reporting absolute misses, conditional misses, and audit coverage together.

Visual token pruning aims to reduce inference cost while preserving the capabilities of the original VLM. Yet existing evaluations rely on aggregate task accuracy and retained token count as proxies for pruning quality and actual efficiency. We reveal two limitations: 1) Aggregate accuracy can obscure whether a pruning method truly identifies critical visual tokens; many examples remain solvable under arbitrary token removal and models can exploit textual shortcuts when visual evidence is degraded. 2) Retained token count is an unreliable proxy for efficiency, as identical token budgets can incur different end-to-end costs depending on where and how pruning is performed. Consequently, comparisons based on aggregate accuracy and retained token counts fail to recover the global performance-efficiency Pareto frontier across methods. We introduce PAA-Bench, a plug-and-play evaluation framework for existing benchmarks, to evaluate whether pruning succeeds through meaningful token selection and whether this success translates into real speedup. To determine when token choice actually matters, we propose Random-Pass@k to measure how often an example remains correct under random pruning, enabling comprehensive evaluation across different levels of selection difficulty. To compare efficiency fairly, we introduce Preservation-Constrained Speedup (PCS) to report the maximum end-to-end speedup each method achieves under a shared preservation target. Experiments across 8 methods, 3 VLM backbones, and 9 benchmarks show that stronger token compression does not necessarily yield greater speedup, method differences widen on examples where token selection matters, and efficiency rankings change with the preservation target.

Best-of-N methods improve text-only LLM reasoning by sampling multiple trajectories before selecting an answer. For Vision-Language Models (VLMs), answer frequency and model confidence alone can be insufficient, as sampled generations also vary in their engagement with the visual input. We find that distributional confidence and visual attention are complementary indicators of rollout quality. Each is limited on its own, while trajectories that show strong signals in both are more likely to be correct. Motivated by this, we introduce VisConf, a rollout-quality score that multiplicatively combines visual engagement with candidate-calibrated Self-Certainty, penalizing rollouts that are weak in either signal. We incorporate Rank-Weighted Consensus (RWC), a tuning-free aggregation rule that converts within-question VisConf ranks into answer-level vote weights. Across three VLMs on MathVista, MMMU-Pro, and MMStar, VisConf with RWC consistently outperforms strong BoN baselines. Our analysis shows particularly large gains when correct trajectories are present but underrepresented in the candidate pool.

Large Vision-Language Models (LVLMs) demonstrate strong performance on multiple-choice question answering (MCQA), yet their reliability under biased conditions remains underexplored. We investigate length bias, the systematic preference for longer answer options, through controlled benchmarks spanning fine-grained visual domains. LVLMs consistently favor longer options even when visually inconsistent, with accuracy dropping 2--38% when correct answers are shorter and increasing 1--70% when longer. This bias persists across model scales. Through layer-wise mechanistic analysis, we reveal that length bias emerges from late-layer suppression of visual evidence in favor of language-side length priors. Blurry-image experiments show that bias intensifies when visual evidence weakens, while abstention tests reveal that LVLMs struggle to refuse incorrect predictions. We propose and evaluate multiple inference-time mitigation strategies; while these provide partial improvements, none achieves deployable bias mitigation without relying on oracle information unavailable at inference time. These results underscore that inference-time interventions are insufficient, pointing to the need for architectural solutions and bias-aware training.

Compact medical vision-language models (VLMs) can answer clinical questions correctly without reliably using the image, relying instead on question wording or learned priors. Standard knowledge distillation copies a teacher's answer preference on the original image, but it leaves unchecked whether that preference is conditional on the visual evidence required by the question. We study this missing target as medical visual evidence dependency: support for the correct answer should decrease when the matched evidence is absent. We introduce Evidence-Conditioned Dependency Distillation (ECDD), which pairs each training image with a control matched to the evidence level and distills the teacher's drop in answer margin across the pair. A lower bound objective transfers positive dependency by penalizing only shortfalls, while original-image supervision preserves task accuracy. The same formulation applies when the evidence is a queried clinical region, a surgical instrument, or whole-image visual content. Across three medical VQA settings, ECDD keeps original-image accuracy competitive with distillation and alignment baselines while making answer support more sensitive to evidence removal. Under the DME headline objective, for example, the positive-case margin response to target rather than random masking rises from 0.786 with KD to 2.023 with ECDD, while original-image accuracy differs by 0.11 points. These results suggest that medical VLM distillation should transfer both the teacher's answer preference and its dependence on visual evidence.

Although reinforcement learning (RL) effectively trains Vision-Language Model (VLM) agents, the mechanisms behind these capability gains are poorly understood. Across four diverse agent environments, we investigate whether the task-success gains from multi-turn RL are accompanied by improved grounded visual question answering about the environment. To assess this, we design environment-specific grounded VQA tasks that reveal improvements in grounded visual understanding. Representational probing shows that vision-encoder probe accuracy remains relatively stable, while LLM-layer probe accuracy improves. Attention analysis further reveals that LLM layers pay more attention to visual tokens during RL training. To test whether increased attention to visual tokens contributes to VQA gains, we introduce a lightweight plug-and-play intervention using a fixed attention bias. Without updating model parameters or computing gradients, this intervention partially restores VQA gains on the untrained model. These results reveal better utilization of visual information as a key mechanism behind the gains from RL.

Vision-language models (VLMs) learn strong statistical regularities during training, which can make them fail to properly respond to visual evidence when input images violate those regularities. It is challenging to disambiguate these prior-induced errors from errors caused by the model missing the underlying visual capabilities. In this work, we use controlled visual in-context learning to encourage the model to focus on the visual evidence, and demonstrate that this alone is frequently sufficient to correct the errors -- suggesting that the models in fact do possess the relevant visual capabilities. Specifically, we show that across five vision-centric tasks, demonstrations yield large gains on prior-conflicting counterfactual examples ($+27.2\%$ for Qwen3-VL 235B and $+15.1\%$ for Qwen3-VL 32B) while leaving accuracy on prior-aligned examples nearly unchanged. These findings generalize across five VLM families with gains from $+9.7\%$ to $+30.3\%$. We analyze visual representations and find that demonstrations indeed redirect attention toward grounding cues while reducing reliance on cues that support the learned prior. Finally, we find that in-context learning is limited when the underlying visual capability is weak or absent. Together, these results show that current VLMs possess visual grounding capabilities that standard evaluations may overlook, underscoring the need for controlled evaluation design in diagnosing and improving model capability.

Vision-language models (VLMs) have achieved strong performance on tasks such as visual question answering, yet small image resizes can turn correct answers into errors. We investigate whether changes in visual configuration, such as image tiling and token arrangement, contribute to this instability. Across seven checkpoints and four benchmarks, equally small resizes cause more correctness flips when they switch configurations. Surprisingly, in over half of these cases, models answer the question incorrectly but can still read the correct answer when told what to read. Furthermore, attention interventions in LLaVA-NeXT suggest that configuration changes can weaken the use of readable information during answering. We therefore guide models using field cues and their own transcriptions. With annotation assistance, these forms of guidance together correct 97.2% of errors with readable information. These findings show that configuration changes can affect how models use information they can still read.

Multimodal large language models (MLLMs) fail at fine-grained visual questions less because they cannot reason than because they never see the evidence: high-resolution images are downsampled before encoding, so the model answers from linguistic priors. The standard remedies are expensive: annotated answers (SFT), hand-engineered verifiers (RLVR), or a large external teacher (on-policy distillation). We ask whether the visual evidence itself can supply the signal for free. We formalize the contrastive evidence gap, the per-token log-likelihood ratio that a model assigns to its own output when conditioned on a question-relevant region versus an irrelevant one, and study it across Qwen2.5-VL-7B, Qwen3-VL-8B, and Qwen3-VL-30B-A3B on V*Bench. Our main positive result is training-free: selecting the candidate crop under which the model's answer distribution is most peaked, using a single-view, label-free criterion, discovers the answer-bearing region with no bounding boxes, training, or labels. It localizes the target 4.4 to 5.1 times better than chance and raises fine-grained accuracy from 70 percent to 85 percent at inference. We further show that the gap is complementary to the model's own confidence. Combining them predicts correctness better than either alone, with AUC up to 0.99, and flags confidently wrong answers, with AUC ranging from 0.97 to 1.00 within the high-confidence subset. All effects concentrate on perception-bottleneck questions and vanish on a global-context control. Finally, we report an honest negative result: converting the same signal into a training method, gated self-distillation (SEG-Distill), does not outperform the base model at pilot scale across three gate designs, while more aggressive gating degrades accuracy. The signal is real, but converting it into training gains remains an open problem.

A radiology report can already answer a clinical question, so it is hard to tell whether a vision-language model also uses the image. ModaLens, a paired image-swap audit, measures how report availability changes image sensitivity: MedGemma-27B on 3,199 paired MIMIC-CXR cases from 293 patients, all 14 questions per case (13 finding-specific and one composite), each image replaced by one from another study, usually of the same patient, with question and report fixed. Under an explicit answer instruction, the model's generated answer changes on 4.26 percent of trials with the report and 20.94 percent without it, a paired increase of 16.7 points (patient-clustered 95 percent CI 15.6 to 17.7), so report availability reduces image-swap sensitivity under this protocol; the original prompt with a lowercase first-token readout gives 4.70 percent against 17.07 percent, and substitutions also move continuous answer scores where the binary prediction does not change. The labels are derived from reports, which limits conclusions about visual correctness; the direction replicates in two further model lineages. Code, the exact prompts and a run record for every number are at https://github.com/criticaldata/MODALENS.

Vision-language models are evaluated by aggregate accuracy on multimodal benchmarks, a practice that implicitly assumes the model uses its visual input. We show this assumption fails on 40%--97% of samples across six VLMs and three perceptual benchmarks: blurring the question-relevant visual region leaves the next-token distribution nearly unchanged. We name this phenomenon the Visual Insensitivity Gap and quantify it with a per-sample Visual Sensitivity Index (VSI). The gap is a property of samples, not of models: VSI ranks correlate across models (grand-mean Spearman rho=+0.40, permutation p<10^-3), so the same samples are flagged insensitive by VLMs sharing no architectural detail beyond a contrastively pretrained vision tower. The mechanism is concrete: on the insensitive samples, a linear probe on each model's own vision tower distinguishes perturbed from clean images at 0.72--0.79 accuracy, yet the model's argmax token changes on only 2%--11% of the same samples, an encoder--LLM gap above 0.65 on every model. Mapping VSI's diagnostic utility cell by cell surfaces a strong regime (multi-choice reasoning on capable VLMs: AUROC=0.85--0.87) and a weak regime (well-calibrated factuality, where softmax confidence already leads). VSI is not a universal best abstention signal; it is a sample-intrinsic indicator of vision-ignoring failure, best used as a conditional ensemble component.

We introduce a Polish-language medical visual question answering (VQA) benchmark, built from Polish Board Certification Examination questions for licensed physicians and dentists pursuing specialist certification. The benchmark comprises image-containing questions spanning diverse medical specialties and visual domains, together with a text-only question answering (QA) control set. We evaluate Polish-oriented, general-purpose open-weight, and commercial vision-language models. The task remains challenging: the best model achieves 79.0% accuracy on the full VQA set, and only GPT-5.6 surpasses the approximate human reference on the subset with available candidate responses; all other evaluated models perform worse than humans. To assess visual grounding, we compare complete inputs with configurations omitting the image, the question, or both, and categorize questions by image importance. Models derive more useful information from the question text than from the image and perform worse on image-dominant questions. Across both QA and VQA, they nevertheless achieve above-chance accuracy from the answer choices alone, showing that non-trivial performance can persist even when key task components are missing.

Multimodal Large Language Models (MLLMs) achieve strong performance by integrating visual inputs with the rich priors of pretrained language models. However, they often fail on vision-centric tasks, especially when visual evidence conflicts with pretrained knowledge. We explore these failures separately using two diagnostic paradigms: (1) probing whether visual information is available, via image reconstruction, and (2) measuring multimodal context sensitivity, the extent to which the model follows visual context versus the language prior. To support the second, we introduce the WhatIfVis, a benchmark spanning five coarse-grained dimensions (spatial-temporal, color, count, size, and weight) whose questions admit answers from either the image or the prior. Our analysis yields three findings: (i) Coarse-grained visual evidence is preserved, as these attributes can be reconstructed from the final-layer image tokens of frozen MLLMs. Failures on questions about these attributes therefore point to post-perceptual utilization, rather than to degraded visual encoding during perception. (ii) Even when explicitly instructed to use or ignore visual evidence, vanilla models (without supervised fine-tuning on the WhatIfVis) show unstable visual context sensitivity. Supervised fine-tuning (SFT) improves this controllability and generalizes across domains, and activation patching further localizes the vision-versus-prior trade-off at architecture-specific depths across all six models. (iii) The vision-versus-prior trade-off is controllable along a learned vector. Applying this steering vector, even without any intent instruction, improves controllability over the vanilla model. Together, these results relocate the bottleneck, indicating that for the coarse attributes we study, MLLMs encode the visual evidence but cannot reliably control their reliance on it.

When do vision-language models need direct access to the image while generating an answer? We study image dependence during answer generation by examining how the visual information needed for the current question becomes available in context. We intervene on direct image access while retaining previously computed states. Across real-image and synthetic tasks, we show that, depending on the generation process, direct access can continue to support accuracy after question processing. Supplying the required attributes as text in the context weakens this dependence. On synthetic tasks, we also examine how dependence changes as the model itself states the required attributes. Before attribute expression, severing access reduces accuracy, and replacing image-side states shifts answers toward the counterfactual content. After sufficient expression, both interventions have smaller effects. Even with an identical generated prefix, dependence differs according to whether image access was available during question processing. Thus, both the visible text and the preceding image access matter. Several of these patterns hold across model families, including Qwen2.5-VL-32B and InternVL3-14B. These findings offer a view of image dependence in terms of the information needed for the current question and the history of image access, beyond generation position alone. This perspective provides a basis for deciding when to reduce visual access during an answer and which visual information to retain for subsequent questions.

When a multimodal large language model answers a visual reasoning question correctly, is the prediction actually supported by the task-critical visual evidence? Correct answers can coexist with flawed reasoning, making accuracy alone an incomplete test of grounding. We introduce VisualFLIP, a paired benchmark with 1,374 images arranged as same-question perturbation pairs across cardinality, attribute, spatial, and logic tasks. Each pair keeps the question fixed but minimally changes the evidence so the gold answer deterministically flips. We evaluate 24 MLLMs with pair accuracy, which requires solving both sides of a pair, and Collapse Rate (CR), which measures how often a model that solves at least one side repeats the same non-empty answer for both images. Together, these metrics show that paired correctness and evidence dependence are related but distinct: capable models can still fail to update after task-critical visual changes, and collapse becomes more severe for some models when the edited image follows an earlier answer in a sequential setting. Further details are available on our project page: https://didizhu-judy.github.io/VisualFLIP/