WinARDet fixes looping in MLLM detectors: F1 rises from 32.16 to 39.36

Multimodal LLMs that locate objects from a text description handle dense scenes poorly. When an image contains hundreds of targets, the model often starts emitting the same box over and over until it hits the decoding limit. The authors of WinARDet propose a mechanism called Compressed Window Auto-Regression (WinAR) that breaks this loop. On the RSDense benchmark, where a single image holds 200 to 2000 objects, F1 at IoU=0.5 rose from 32.16 to 39.36 relative to the RexOmni baseline, and the share of redundant predictions fell from 35.48% to 13.00%.

Where the repetition loop comes from

MLLM detectors usually frame grounding as autoregressive coordinate generation: the model writes out boxes one token at a time, and each new token depends on the entire preceding answer history. This works on scenes with a few objects. In dense scenes the history keeps growing, and coordinate tokens pile up in it.

The authors call the cause of the failure historical-answer attention takeover. Repeated coordinates in the growing history attract a disproportionate share of attention. They crowd out the attention sinks the model needs to look at new visual evidence. The model stops "looking" at the image and keeps quoting itself. The loop is self-reinforcing: the more repeats in the history, the more attention they get, and the more likely the next duplicate becomes.

How the compressed window works

WinAR cuts a long answer sequence into consecutive generation windows. Full token-level autoregression is kept only inside the current window. Each previous window is compressed into a fixed set of latent tokens, and these tokens serve as the condition for further decoding.

The logic is as follows. Raw coordinates of past boxes no longer sit permanently in the context, which weakens the trigger for attention takeover. Dependencies across windows are not lost, though: the latent tokens tell the model what has already been found, so it does not start over. The size of the past representation is fixed for each window, so its contribution to the context does not grow with the number of objects.

Key figures from the paper:

• F1 at IoU=0.5 on RSDense: 32.16 → 39.36 (+7.2 points, roughly a 22% relative gain);
• share of redundant predictions: 35.48% → 13.00%;
• scene density in the benchmark: 200 to 2000 objects per image;
• tested modes: parallel box decoding and polygon generation.

Where it could help and what to check

The strength of the work is that the method is not tied to a single output scheme. The authors show that WinARDet transfers across decoding strategies and formats: it works with both parallel box decoding and polygon generation. That matters because segmentation contours are also long coordinate sequences and suffer from the same repetition.

For teams building detection on VLMs, the takeaway is practical. If a model counts or localizes many small objects (aerial imagery, warehouses, dense urban scenes, crowds), measure the redundancy rate separately instead of looking only at F1. Duplicates show up in this metric immediately, while on medium-density scenes they may not appear at all. The idea of compressing history into latent tokens could also apply to any task with long structured output.

There are caveats. An absolute F1 of 39.36 shows that dense grounding remains a hard task: most objects are still localized imprecisely or missed. The description does not show what the compression costs in latency and compute, or whether it requires fine-tuning the model from scratch. The results come from a single dense benchmark, so claims of broad applicability need independent measurements on other data. Still, the attention-takeover diagnosis gives a concrete point of intervention, and this approach is easier to test than yet another increase in model size or context length.