On V*Bench and HR-Bench, VisLens matches or beats prior visual search baselines, running 8.5–9.9x faster than Thyme and up to 22.2x faster than training-free multi-pass search. Every crop traces back to decoded tokens, so you can see what the model looked at.

VisLens reads the model's own hidden states. A logit lens projects each image region through the LLM head to reveal what it already says; a light tuned-lens makes early layers readable. Match decoded tokens to the query, cluster, merge into crops, answer.

The adapter is where segmentation quality drops: alignment costs spatial detail. LLM layers claw it back. Correct tokens act as anchors, pulling wrong neighbors to the right label. Causal masking starves early tokens of anchors; bidirectional attention fixes that.

Three probes. Linear probing at every layer to track segmentation quality. Attention knockout to test whether tokens repair each other. Bidirectional attention among image tokens to measure what causal masking costs.

MLLMs are being pushed into pixel-level work like segmentation, but nobody knows where in the stack that spatial ability lives, or where it breaks. So we opened the pipeline: vision encoder, adapter, and every single LLM layer.

One recipe, three tasks, zero training. V*Bench: 71.7 → 91.1 (+19.4). LongVideoBench: 52.7 → 60.5 (+7.8). Charades-STA mIoU: 10.4 → 46.8 (+36.4), beating TimeChat and VTimeLLM, both fine-tuned for temporal grounding. Uncertainty can steer perception in real time.

We propose a simple method called UG-Framework. We feed the model a crop, a frame, a time window, and read the output entropy. Low entropy means it found what it needs. Score every candidate, keep the most confident segments, then answer.

MLLMs still struggles with complex visual tasks like visual search, video temporal grounding and long video understanding. Instead of task-specific fine-tuning, we found the model's internal entropy as a useful signal that can boost all tasks in a training-free manner.

We will have two poster presentations and one workshop poster at #ECCV2026 in Malmö🇸🇪! Please come and say hi if you are interested in topics related to multi-modal learning. Details in the thread👇

📢 Fully funded PhD @ Helmholtz Munich + TUM We're hiring PhD students to build trustworthy AI systems for science: explainable AI, VLMs, reliable adaptation, agentic workflows. ⏰ Priority deadline 10 Aug 2026. Application details below 👇