Abstract
Many visual requests -- ``the object to open this bottle'', ``the person not wearing a helmet'' -- require reasoning, not just category matching. Pure open-vocabulary detectors need an explicit phrase; vision-language models (VLMs) can reason yet ``see but mis-speak'', attending to the right region but returning the wrong box or label. We argue this is a \emph{binding} failure: in coordinate-as-text VLMs localization passes through the autoregressive head, coupling it to language generation; in two-stage pipelines the model's intent is squeezed through a single class string. We present HKVLM, which removes localization from the language path. A frozen, language-aligned detector emits class-agnostic region proposals; a frozen language model encodes reasoning instructions as referential query embeddings; a lightweight \emph{alignment hook} binds queries to regions by contrastive retrieval and bipartite assignment in a shared embedding space. A perception-grounded faithfulness veto forbids naming an object that no region supports. Only the hook is trained, targeting small-data cold-start settings where monolithic VLM tuning struggles. We formalize a \emph{say-vs-see} decomposition separating localization error (SeeErr) from binding error (SayErr), and evaluate on RefCOCO/RefCOCO+/RefCOCOg and POPE. With frozen Grounding DINO and Qwen2.5-VL, training only the hook lifts grounding accuracy by $50$--$90\times$ over untrained cross-space matching; the faithfulness veto raises POPE accuracy from near-chance ($0.50$) to $0.66$--$0.76$ and reduces hallucination from ${\sim}0.99$ to $0.23$--$0.43$, with gains from $200$ expressions. Increasing proposals from $M{=}50$ to $M{=}300$ improves grounding by $19$--$24\%$ without retraining, confirming that residual error is perceptual (SeeErr) rather than binding (SayErr).
Abstract (translated)
URL
https://arxiv.org/abs/2606.28862
