Technical Summary: Seeing without Looking: Do Vision-Language Benchmarks Really Test Vision?

Problem Statement

Vision-Language Models (VLMs) have achieved high scores on numerous benchmarks designed to evaluate grounded visual understanding and object hallucination. However, it remains unclear whether these scores genuinely reflect a model's reliance on fine-grained visual evidence or if they stem from the exploitation of coarse visual cues, representational redundancy, or strong language priors. The authors observe a critical disconnect: on widely used benchmarks like POPE, randomly removing a substantial fraction of image tokens (up to 75%) results in only a marginal degradation in accuracy. This suggests that standard benchmark metrics may not be sensitive enough to detect the loss of detailed visual grounding, potentially leading to overconfident conclusions about model capabilities in visual reasoning and hallucination mitigation.

Methodology

The authors conduct a comprehensive, multi-level diagnostic analysis across multiple open-source VLMs (including LLaVA-1.5, Qwen3-VL, Gemma-3, InternVL3, and Molmo) and benchmarks (POPE, A-OKVQA, MME, and AMBER). The study employs three primary categories of interventions:

1. Visual-Level Interventions:

   • Token Dropping: Randomly removing image tokens at the embedding level before they enter the language decoder.
   • Global Degradation: Systematically weakening visual evidence through "No-Image" (dummy black image), "Black" occlusion (masking specific image regions), and "Blur" (mixing with noise).
   • Entity-Level Manipulation: Focusing on Ground Truth positive samples, the authors use segmentation models (Grounding DINO, SAM2) to perform precise interventions:
     • BlackMask: Masking only the pixels of the queried entity.
     • BlackBox: Occluding the entire bounding box of the entity (removing local context).
     • EntitySwap: Replacing the queried entity with an unrelated object using image editing models.

2. Task-Formulation Interventions:

   • Expanded Answer Space: Introducing an explicit "Unknown" option to test if models can express uncertainty when visual evidence is degraded.
   • Open-Ended Generation: Reformulating binary questions into open-ended generation tasks to evaluate the ranking and probability of the target entity in the output distribution (using Mean Reciprocal Rank and Mean Probability).

3. Representational Analysis:

   • Layer-wise Geometry: Analyzing the vision encoder's internal states to measure spatial discriminability. The authors compute intra-block vs. inter-block cosine similarity, k-means spatial compactness, and the effective rank of token representations across network layers to determine if spatial distinctions degrade as information propagates deeper into the model.

Key Results

• Robustness to Visual Loss: On POPE, accuracy remains nearly unchanged even when 75% of image tokens are dropped. Similarly, global degradations like heavy blurring or partial black masking cause disproportionately small drops in accuracy compared to the severity of the visual corruption.
• Entity-Level Insensitivity: Even when the specific queried entity is completely masked (BlackMask) or replaced (EntitySwap), models often maintain high "Yes" rates. While the decision margin (logit difference between "Yes" and "No") decreases, the distribution remains shifted toward affirmative answers, suggesting that surrounding scene context or language priors sustain the prediction.
• Failure to Express Uncertainty: When an "Unknown" option is provided, most models rarely select it, even under severe visual degradation (e.g., No Image or high noise). They continue to produce confident binary judgments.
• Weak Representation of Entities: In open-ended generation tasks, the target entity often receives a low rank in the model's generation distribution, even when visible. This indicates the model does not strongly encode entity-specific information in its output probabilities.
• Degradation of Spatial Discriminability: The representational analysis reveals that as visual tokens pass through deeper layers of the vision encoder, inter-block similarity increases while intra-block similarity decreases. By the final layers, tokens from different spatial regions become directionally indistinguishable, and the effective rank of the representation collapses. This suggests a loss of fine-grained spatial structure within the model's internal processing.

Key Contributions

1. Identification of a Measurement Gap: The paper identifies that standard benchmark metrics (top-1 accuracy) often fail to degrade significantly even when corresponding visual evidence is substantially weakened or removed, indicating a lack of sensitivity to fine-grained grounding.
2. Multi-Level Diagnostic Framework: The authors develop a systematic framework involving token-level reduction, global degradation, entity-level occlusion, semantic manipulation, and decision-level analysis to probe the true dependencies of VLM performance.
3. Evidence of Shallow Visual Reliance: The study demonstrates that current VLMs can maintain correct predictions despite the loss of detailed visual evidence, relying instead on coarse scene cues, redundancy, and language priors. This reveals a systematic gap between benchmark success and genuine reliance on question-relevant visual evidence.

Significance and Claims

The paper argues that current benchmarks are insufficient for reliably evaluating fine-grained visual grounding in VLMs. The authors claim that high benchmark scores may overestimate a model's visual capabilities and its robustness to visual perturbations. Consequently, methods designed to improve visual grounding, mitigate hallucinations, or enhance inference efficiency (e.g., by dropping image tokens) may appear effective simply because they exploit dataset priors or coarse cues rather than true visual understanding.

The authors conclude that evaluation must move beyond top-1 accuracy to explicitly test whether model predictions respond appropriately to changes in visual evidence. They suggest that the observed stability in performance is not a sign of robust visual understanding but rather a symptom of shallow reliance on visual inputs, necessitating more diagnostic evaluation protocols to accurately assess model capabilities.