Summary
Multimodal language models often miss details in visual questions because the images they see are blurry or too small, so they guess based on language alone. This paper shows that by looking carefully at specific parts of the original image, the models can better find the important area related to the question without needing extra labels or training. The authors introduce a way to measure how much the model’s answer changes when focusing on different image parts, which helps locate the right spot and improve accuracy. However, using this signal to retrain the models did not yet bring improvements, meaning more work is needed to turn this insight into stronger learning.
What this means in practice
- •For multimodal ai developers: Improve fine-grained visual question answering accuracy by selecting image regions that most strongly support model answers without additional labeling or training.
- •For computer vision engineers: Enhance localization of answer-relevant image regions during inference using contrastive signals from language model output distributions.
Authors
Santi Ram Tiwari, Nihal Naik, Devbrat Pandey, Nishant Sinha
Abstract
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.