Research map / Vision-language and multimodal / Multimodal LLMs

Are VLMs Seeing or Just Saying? Uncovering the Illusion of Visual Re-examination

ICML 2026 · Vision-language and multimodal › Multimodal LLMs · approach: VLM capability analysis · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces VisualSwap, a probe that swaps the image after a VLM's self-reflective statement, and VS-Bench (800 image pairs) to test whether VLMs genuinely re-examine visual input; finds large accuracy drops, worse for thinking models, and attributes it to weak visual attention during self-reflection.

Problem

Self-reflective statements such as "let me check the figure again" in VLM reasoning may not trigger genuine visual re-examination and may instead reflect learned textual patterns. The paper tests whether VLMs actually re-ground in the image after such statements.

Approach

After a model reasons over image I_a and emits a reflection prompt, the image is swapped for a visually similar but answer-different I_b within the same generation; the drop from base accuracy on I_b measures failure of visual re-examination. Combined with multi-turn user-instruction comparisons and attention-score analysis to explain the gap.

Key results

• Probe accuracy drops by up to 60% on average; ERNIE-4.5-VL-28B-A3B-Thinking falls from 79.9% to 19.6%.
• Thinking variants degrade ~3x more than Instruct counterparts (Qwen3-VL-32B: 48.3% vs 17.9% average drop).
• Explicit multi-turn user instruction restores Qwen3-VL-235B-A22B-Thinking from 34.1% to 85.4% (base 88.8%).

Surprising findings

• Thinking models are nearly 3x more vulnerable than instruct counterparts, and larger models do not fare better (Qwen3-VL-235B drop 54.6% vs 39.4% for 8B Thinking).
• Swapping at naturally generated reflection points yields equal or lower accuracy than standardized prompt injection.

Limitations

• Evaluation limited to open-weight models; closed-source models were excluded because they do not support the probing mechanism.
• Benchmark is small (800 pairs) and built with human-in-the-loop image generation using Nano Banana Pro.
• Attention amplification is presented only as a proof-of-concept on two Qwen3-VL-8B variants.

Experimental setup

Datasets: MathVista MathVerse MathVision MMMU-Pro

Benchmarks: VS-Bench

Metrics: accuracy performance degradation (Delta) CLIP similarity SSIM LPIPS visual attention score

Baselines: Qwen3-VL-8B Instruct Qwen3-VL-8B Thinking Qwen3-VL-32B Instruct Qwen3-VL-32B Thinking Qwen3-VL-30B-A3B Instruct Qwen3-VL-30B-A3B Thinking Qwen3-VL-235B-A22B Instruct Qwen3-VL-235B-A22B Thinking ERNIE-4.5-VL-28B-A3B Instruct ERNIE-4.5-VL-28B-A3B Thinking Kimi-VL-A3B Instruct Kimi-VL-A3B Thinking

Builds on: MathVista MathVerse MathVision MMMU-Pro VLMEvalKit vLLM

Closest papers by meaning

• Through the Lens of Contrast: Self-Improving Visual Reasoning in VLMs (ICLR 2026 · similarity 0.89)
• Look-Back: Implicit Visual Re-focusing in MLLM Reasoning (AAAI 2026 · similarity 0.89)
• Do VLMs Perceive or Recall? Probing Visual Perception vs. Memory with Classic Visual Illusions (CVPR 2026 · similarity 0.89)
• When to Think and When to Look: Uncertainty-Guided Lookback (CVPR 2026 · similarity 0.89)
• VLMs have Tunnel Vision: Evaluating Nonlocal Visual Reasoning in Leading VLMs (NeurIPS 2025 · similarity 0.89)
• BLINK-Twice: You see, but do you observe? A Reasoning Benchmark on Visual Perception (NeurIPS 2025 · similarity 0.89)