Skip to content

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

🕒 Published (v1): 2026-05-21 17:35 UTC · Source: Arxiv · Venue: CVPR 2026 · link

Why this paper was selected

CVPR 2026; reveals VLM benchmark scores don't require grounded visual evidence

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

Standard VLM benchmarks significantly overestimate models' reliance on fine-grained visual evidence: accuracy on POPE, A-OKVQA, and MME degrades only marginally even when up to 75% of image tokens are removed or the queried entity is fully occluded. The paper provides a multi-level diagnostic framework—spanning token dropping, global degradation, entity-level occlusion, entity swapping, question reformulation, and representational analysis—to expose this measurement gap. The findings suggest that high benchmark scores are substantially sustained by coarse visual cues, representational redundancy, and language priors rather than faithful grounded visual reasoning.

Problem

Benchmark accuracy on widely used VLM hallucination and grounding benchmarks (e.g., POPE) is implicitly treated as a proxy for fine-grained visual grounding. However, it is unclear whether high scores reflect genuine reliance on question-relevant visual evidence or are instead supported by coarse cues, redundant visual representations, or strong language-side priors. No prior work has systematically characterized how benchmark performance responds to progressive, multi-level weakening of visual evidence.

Method

The authors apply a hierarchy of input interventions to seven open-source VLMs (LLaVA-1.5-7B, Qwen3-VL-4/8/32B, Gemma3-12B, InternVL3-8B, Molmo-7B-D-0924) across four benchmarks (POPE, A-OKVQA, MME, AMBER):

  1. Token-level dropping: randomly remove a fraction \(\sigma \in [0,1]\) of image tokens after the vision encoder, evaluating POPE accuracy vs. drop ratio.
  2. Global image degradation: (a) full black image, (b) black occlusion covering proportion \(r\) of area, (c) additive noise blending \(\tilde{I} = (1-\alpha)I + \alpha\epsilon\).
  3. Entity-level occlusion: restrict evaluation to GT=Yes POPE samples; localize the queried entity with Grounding DINO + SAM2; apply BlackMask (entity pixels only), BlackBox (full bounding box), or EntitySwap (replace entity with an unrelated object via Gemini-3).
  4. Decision-margin analysis: examine \(\Delta = \log p(\text{yes}) - \log p(\text{no})\) distributions beyond top-1 accuracy.
  5. Task-formulation interventions: add an explicit "unknown" option to probe uncertainty awareness; reformulate as open-ended generation and evaluate target-entity rank and MRR.
  6. Layer-wise representational analysis: compute intra- vs. inter-block cosine similarity, k-means spatial compactness, and effective rank \(\text{erank}(U^{(\ell)}) = \exp\!\bigl(-\sum_j \bar{\sigma}_j \log \bar{\sigma}_j\bigr)\) across vision encoder layers to assess spatial discriminability collapse.

Key Contributions

  • Identifies that VLM benchmark accuracy does not degrade proportionally to progressive visual evidence removal—even 75% token dropping causes only ~3% accuracy decline on POPE.
  • Develops a multi-level diagnostic framework spanning global degradation, entity-level occlusion/swap, question reformulation, and probability-level analysis.
  • Demonstrates via decision-margin distributions that internal model confidence is weakened by entity removal even when top-1 accuracy is unchanged.
  • Shows through layer-wise geometric analysis that vision-encoder tokens converge toward a low-dimensional, spatially indiscriminable subspace in deeper layers, providing a mechanistic hypothesis for robustness to visual perturbation.
  • Establishes that observed robustness is not an artifact of yes/no format: it persists in open-ended generation (AMBER) and when an "unknown" option is available.

Results

  • Token dropping on POPE: at \(\sigma = 0.75\), accuracy drops by only ~3% for Qwen3-4B and LLaVA-1.5-7B; Qwen3-32B and Gemma3-12B show non-monotonic behavior, slightly outperforming baseline at \(\sigma = 0.25\).
  • Global degradation (Table 5): under 75% black occlusion, POPE accuracy for most models remains 0.76–0.81 (vs. baseline 0.94–0.98); no-image collapses all models to ~0.50 (chance). Similar pattern on A-OKVQA and MME.
  • Entity-level occlusion (Table 1): BlackMask reduces POPE Yes Rate from ~0.93–0.99 to 0.67–0.94; BlackBox further reduces to 0.44–0.84. Models still affirm the presence of a removed entity at high rates.
  • EntitySwap (Table 2): Yes Rate should ideally be 0; observed values range 0.27–0.63 across models, confirming failure to update on entity substitution.
  • Open-ended MRR (Table 4): EntitySwap reduces MRR by 68.6%, mean rank worsens from 391.9 to 943.0 (+140.6%), median rank from 85 to 458 (+438.8%), mean prediction probability drops 79.4%.
  • Unknown-option selection (Table 3): most models choose "unknown" at near-zero rates even under complete image removal; Gemma3-12B and Qwen3-32B are exceptions under 75% noise (0.78 and 0.81).

Limitations

  • Analysis is confined to open-source models; closed-source frontier VLMs are not evaluated.
  • Entity localization pipeline (Grounding DINO + SAM2) may introduce noise or fail on small/occluded entities, affecting entity-level intervention quality.
  • Token-dropping experiments operate post-vision-encoder, where bidirectional attention may have already mixed spatial information across patches, potentially underestimating the effect of true spatial removal.
  • The representational analysis of vision-token geometry is reported qualitatively/descriptively; the paper is truncated before full quantitative results are presented.
  • No new benchmark or training-side remedy is proposed; the work is purely diagnostic.
  • AMBER results are limited to coarse-grained degradations, leaving fine-grained open-generation interventions unexplored.

Relevance to Vision-Language Models

This paper directly challenges the validity of the dominant evaluation paradigm for VLMs by showing that hallucination and grounding benchmarks are systematically insensitive to the loss of fine-grained visual evidence. For researchers tracking VLMs, this is a foundational calibration concern: methods reporting POPE/MME improvements may be advancing language-prior exploitation rather than genuine visual grounding. The multi-level diagnostic framework and the decision-margin + MRR metrics proposed here offer concrete tools for constructing more discriminative evaluations. The layer-wise convergence of vision-encoder token representations also connects to ongoing work on token compression and efficient inference, suggesting those methods should be benchmarked with grounding-sensitive metrics rather than standard accuracy.