WinARDet: Compressed Window Auto-Regression for Long-Horizon Vision-Language Grounding

WinARDet. Compressed Window Auto-Regression for Long-Horizon Vision-Language Grounding. By Chenxu Wang and eight others, from Nankai University and Beijing Normal University. Published on the fourteenth of October. Its code repository is worked out: about two hundred source files, with a readme.

Detectors built on multimodal large language models treat vision-language grounding as generating coordinates one token at a time. In dense scenes, that goes wrong in a specific way. The model falls into a box repetition loop, emitting the same box again and again until the decoding budget runs out. The authors trace the loop to what they call historical-answer attention takeover. Under full-history autoregression, the coordinate tokens repeated across a growing answer history draw a disproportionate share of attention, and they crowd out the attention sinks the model needs in order to attend to new visual evidence. The more it has written, the less it looks.

Compressed Window Auto-Regression, or WinAR, is built to break that loop. Instead of letting the model attend token by token to its entire output so far, it splits a long answer sequence into consecutive generation windows. Full token-level autoregression is kept only inside the current window. Every earlier window is compressed into a fixed set of latent tokens, and those latents condition the decoding that follows.

The design aims at both halves of the problem. Raw historical coordinates are no longer repeatedly in view, which removes the exposure that triggers the takeover. Yet the compressed windows still carry information forward, so the dependencies across windows that coherent multi-object generation relies on are preserved. The detector built on this mechanism is WinARDet.

It is tested on RSDense, where each image contains at least two hundred objects and as many as two thousand. Against the baseline, WinARDet raises the RexOmni F1-score, at an IoU threshold of one half, from thirty-two point one six to thirty-nine point three six. The redundancy rate falls from thirty-five point four eight percent to thirteen percent. WinARDet also generalizes across decoding strategies and output formats, including parallel box decoding and polygon generation.

A loop that repeats its own past is cured here by showing the model less of that past, not more, while keeping a compressed memory of it. Seen that way, the drop in redundancy is the direct signature of the fix, and the authors present the approach as a practical paradigm for long-horizon grounding across scene densities.