cs.CVNov 16, 2026

How VLMs Plan What's Next

Authors: Rohit Gandikota, Arnab Sen Sharma, Adam Belfki, David Bau

Organizations: Northeastern University

Abstract

When a vision-language model describes an image, it names one object at a time and it rarely repeats itself, yet how the model plans this order is far from obvious. We characterize the mechanism a VLM uses to decide where to look before it looks. The mechanism has four main parts. First, binding heads match the relevant text tokens against the image and resolve them into pointers to the matched image region. Second, the same binding heads bind each written answer to the region it came from, and the model turns it into a progress signal: a cursor, the position in the image up to which the list has been written. Third, a select step chooses the nearest unwritten region after the cursor, or stops. Finally, gaze heads execute the chosen pointer and return a concept of what is drawn there, which language heads and the late MLPs turn into a word in the language of the question. We validate each part of this mechanism through causal patching in Qwen3-VL models on a synthetic visual retrieval task, and test our mechanisms against alternative hypotheses. We further show that the mechanism holds across input formats, natural photographs, and languages, and that its binding, relay, and gaze heads appear at every model size from 2B to 32B. This work allows us to investigate how a VLM turns a question and an image into a sequence of answers, layer by layer and token by token. More broadly, it opens a way to study how models plan over complex multi-step tasks with long rollouts. Our project is available at https://plan.baulab.info/

Figures & tables

Appendix figures & tables51 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

Jun 12, 2026

Gaze Heads: How VLMs Look at What They Describe

How a vision-language model internally solves the task of describing an image is far from obvious. We find that the model develops a specific mechanism for this: a small set of attention heads in its language-model backbone, which we call gaze heads, whose attention tracks the image region the model is currently describing. We find them with a simple correlation score from a few forward passes, using comic strips as a controlled testbed where narrative order is laid out spatially. These gaze heads do not just track the image tokens being described: redirecting their attention to a chosen region forces the VLM to describe that region instead. A single attention-mask intervention on the top-100 gaze heads, fewer than 9% of all heads, steers the model's answer to any chosen comic panel at 83.1% accuracy, while the same intervention on random heads fails to redirect the answer, and intervening on all heads destroys generation. The same lever also extends to continuous control: switching the gaze target mid-generation makes the model wrap up its current panel description and move to the new one within a few tokens. Beyond comics, the same intervention redirects answers to chosen regions in natural COCO images. The mechanism further recurs across model sizes from 2B to 32B parameters and across other VLM architectures, although some frozen-encoder families show no comparable head set. More broadly, this shows that targeted edits identified through mechanistic analysis can serve as practical inference-time levers for steering multimodal model behavior, without any retraining. Our code, interactive demo, and datasets are available at https://gaze.baulab.info/
Aug 27, 2026

Retrieval Heads Meet Vision: Uncovering How VLMs Locate and Extract Visual Information

Vision-language models (VLMs) can locate an image region referred to by a text prompt and route the corresponding visual evidence to the output, yet the internal mechanism behind this behavior is not understood. Inspired by retrieval heads in large language models, we ask whether VLMs contain an analogous mechanism for visual retrieval. We answer affirmatively by introducing Visual Retrieval Heads (VRHs), a small subset of attention heads (about 1.7-2.6%) that are causally responsible for grounding text descriptions to image regions. To find them, we recast existing head-scoring methods under a unified design space over query tokens, key aggregation, and cross-sample aggregation. We then show that scoring attention from output prediction tokens with a sum over the ground-truth referent region most reliably identifies causal heads. Across eleven VLMs and five referring-expression benchmarks, masking only the top 20 VRHs reduces grounding accuracy by up to 80 percentage points, while masking the same number of random heads has little effect. Beyond replicating the causal-sparse-universal triad established for text retrieval heads, VRHs exhibit several properties not previously reported: they generalize across visual reference tasks, remaining causal on attribute, spatial, counting, and visual-math benchmarks despite being discovered through bounding-box prediction; they are functionally specific, preserving output format while corrupting localization; and they are architecturally shared, transferring causally across VLMs that share an LLM backbone but differ in vision encoder, projector, and instruction tuning.
Nov 16, 2026

Seen but Unsaid: Locating and Repairing Cross-Image Comparison in Vision-Language Models

Vision-language models (VLMs) often fail to match a point or an object across two images, even when the answer can be decoded from their own hidden states. We trace this failure from the visual tokens to the answer letter, in Qwen3-VL (2B-32B), InternVL3.5-8B, and Gemma-3-12B, on synthetic shape matching, same-instance point correspondence (HPatches), and semantic correspondence (SPair-71k), and find two distinct failure points, a small set of attention heads that carries the comparison, and a repair that touches 0.02% of the parameters. Two failure points: exact copies are compared but not read out. An answer-position probe exceeds the calibrated output by 15 to 25 points in every family, for nameable and unnameable shapes alike, and steering along the probe's direction moves no more answers than random directions. Semantic correspondences are present in the visual tokens but never reach the answer. One pathway: redistributing the cross-image attention of 16-32 heads, with its mass preserved, removes 17-22 points of synthetic and 15-32 points of HPatches accuracy, against at most 7 for random or attention-matched heads. A lever that training finds on its own: training only these heads' query and key projections raises SPair-71k accuracy by 28.3 points (Qwen3-VL-8B) and 25.1 points (InternVL3.5-8B), beats both layer-matched and attention-matched heads, and transfers to HPatches and BLINK. An unconstrained all-layer adapter routes 21-51% of its gain through the attention of the same heads, against at most 7% for random heads, and does so by changing what they read (83-86% of the effect) rather than their weights. Every decision rule was frozen before its run, and every reported number is recomputed by independent verification scripts.