Composition, Not Conversation: VLMs Lose the Scene, Not the Thread

Benchmark accuracy is often implicitly assumed to reflect grounded visual understanding in vision-language models (VLMs), yet it remains unclear to what extent such scores truly reflect reliance on visual evidence. Motivated by a surprising observation that removing a substantial fraction of image tokens only degrades model performance very slightly on a widely used hallucination benchmark, we systematically investigate this mismatch in a set of open-source VLMs. Our analysis spans multiple levels of granularity, spanning global visual degradation, localized occlusion, question reformulation, answer-space expansion, and decision-level analyses beyond standard accuracy. We further complement these behavioral results with a layer-wise analysis of vision-token geometry. Throughout the experiments, we find that although VLMs do incorporate visual input, their predictions are less sensitive to the loss of fine-grained visual evidence that standard accuracy should have suggested. Even when the final prediction remains unchanged, the model's internal support for the correct answer may already be weakened. We further complement a representation-level analysis, which shows increasing similarity among visual tokens in deeper layers, providing a possible explanation for our findings. Together, these results suggest that current benchmarks are not sufficient to reliably evaluate fine-grained visual grounding in VLMs.

Vision-language models (VLMs) answer visual questions by combining visual information extraction with downstream problem solving. We investigate a fundamental question: Does an incorrect answer necessarily reflect a failure in visual extraction or problem solving? A model may succeed at both abilities when tested separately yet still fail on the original multimodal question, a distinction that overall answer accuracy cannot reveal. To study this, we perform a question-level empirical analysis across multiple VLMs and visual domains. We define an exactly scorable task state (i.e., the visual information sufficient to solve a question) and use it to test whether the same model can extract the required state, solve the question from the ground-truth state, and answer the original multimodal question. We find that composition failures, where extraction and solving both succeed but direct answering fails, account for 17.7% to 75.6% of direct-answering errors across multiple VLMs and datasets. To address this failure mode, we introduce a simple yet effective method, termed State Realization Tuning (SRT). SRT fine-tunes LoRA adapters attached to the language-model layers while keeping the pretrained VLM weights frozen. It trains the model to output the ground-truth task state before the final answer in a single autoregressive response. SRT improves over standard supervised fine-tuning by 1.7 to 14.1 percentage points and repairs 92.5% to 98.1% of diagnosed composition failures. A single LoRA adapter trained with SRT also improves performance across substantially different task-state structures. Our work shows that having both visual extraction and problem-solving capabilities does not guarantee correct multimodal answering. Requiring the model to first output the visual information needed to solve the question can help bridge this gap.

Counterfactual analysis is widely used to study evidence use in vision-language models, but its diagnostic value is limited on well-posed tasks: when several cues independently support the same answer, removing one may not change the prediction. We propose monocular metric object-size estimation as an ill-posed diagnostic setting for evidence selection: because physical size cannot be determined from a single uncalibrated image, models must rely on imperfect cues category priors, target appearance, local context, apparent image size, and scene geometry. We assemble Metric VQA ($10{,}813$ dimension queries from Objectron and $331$ tape-measured in-the-wild scenes) and evaluate $12$ open-weight VLMs ($3$--$397$,B parameters) with counterfactual analysis decomposing six visual and language evidence channels. Even the largest VLMs tested (Qwen3-VL-235B, Qwen3.5-397B, InternVL3.5-241B) trail a text-only frontier LLM on the in-the-wild split. The diagnostic analysis shows: target identity is the most load-bearing cue, target pixels and local context help only some models, apparent size shifts predictions without a directional readout, and global scene geometry is largely unused. We analyze LoRA fine-tuning as an actionable intervention specific to metric estimation: while the task is learnable, the models do not learn to leverage scene geometry.