WinARDet: compressed window autoregression stops box repetition loops in dense grounding
NankaiUniversity · hf · 2026-10-21
Nankai University and BNU propose WinAR, addressing MLLM-based detectors looping on identical boxes in dense scenes. They attribute the failure to historical-answer attention takeover: under full-history autoregression, repeated coordinate tokens attract disproportionate attention, crowding out attention sinks needed for new visual evidence. WinAR partitions long answer sequences into generation windows, restricts full token-level autoregression to the current window, and compresses prior windows into a fixed set of latent tokens, cutting exposure to raw historical coordinates while preserving cross-window dependencies. Validated on RS Dense with 200–2,000 objects.
More from Multimodal
- Topaz joins Adobe; Starlight Fast upscales 720p AI video to 4K/60fps — Mr_AllenT · 2026-10-21
- FLUX 3 Klein on RTX 3090: int8 Doubles Speed, but FLUX.2 klein 4B Wins on Quality — strata2signal · 2026-10-21
- EvaOmni: an evidence-grounded benchmark for audio-visual reasoning in omnimodal LLMs — VisionXLab · 2026-10-21
- Qwen-Image 2.1 edit keeps producing plastic skin, negative prompts don't help — NorthAppropriate4458 · 2026-10-21
- AI-generated illustration visualizes Paz Soldán's sci-fi story 'Artificial' and its Ship of Theseus theme — dasein47 · 2026-10-21
- AI-generated Halloween video of a woman walking through skeleton crowds goes viral — Stephane_AI · 2026-10-21