Multimodal large language models used for vision-language grounding generate object coordinates autoregressively, but in dense scenes they frequently fall into a failure mode: the model keeps emitting the same box over and over until its output budget is exhausted. The authors of WinARDet trace this to "historical-answer attention takeover," where repeated coordinate tokens in the growing answer history attract a disproportionate share of attention, crowding out the visual evidence needed to move on to new objects.
To break that loop, they propose Compressed Window Auto-Regression (WinAR). Instead of letting every previous token attend to every new token, WinAR restricts full autoregression to the current generation window and compresses each completed window into a fixed set of latent tokens. Those latent tokens still carry cross-window dependencies, so the model can produce coherent multi-object outputs without being re-exposed to raw historical coordinates.
On the RSDense dataset, which contains 200 to 2,000 objects per image, WinARDet improves the RexOmni F1-score at IoU=0.5 from 32.16 to 39.36 over the baseline, while reducing the redundancy rate from 35.48% to 13.00%. The authors also report that the approach generalizes across decoding strategies and output formats, including parallel box decoding and polygon generation, suggesting it could serve as a practical paradigm for long-horizon grounding tasks across varying scene densities.