When Sinks Help or Hurt: Unified Framework for Attention Sink in Large Vision-Language Models
Paper • 2604.03316 • Published
LSG checkpoints — When Sinks Help or Hurt (ECCV 2026 Oral, Spotlight)
Layer-wise Sink Gating (LSG) gate checkpoints for
"When Sinks Help or Hurt: Unified Framework for Attention Sink in Large Vision-Language Models"
(arXiv:2604.03316). All checkpoints are
gate-only (non_lora_trainables.bin, well under 3 MB each); backbones are
loaded separately and remain unchanged.
| folder | backbone | configuration |
|---|---|---|
llava15_L10_PEFT/ |
liuhaotian/llava-v1.5-7b | single layer L10 (paper main) |
llava15_stack_L3_6_7_10_19_PEFT/ |
liuhaotian/llava-v1.5-7b | 5-layer stack {3,6,7,10,19} (paper main) |
crossarch_phi3_L20/ |
xtuner/llava-phi-3-mini-hf | Appendix F Track A, L20 |
crossarch_qwen25_L4/ |
Zhang199/TinyLLaVA-Qwen2.5-3B-SigLIP | Appendix F Track B, L4 (prior 1.25) |
Each folder ships the matching sink_cfg.json. The two LLaVA-1.5 folders keep the _PEFT suffix because the loader activates gate-only loading when the model_base directory name contains PEFT. Usage instructions:
github.com/JH-GEECS/lsg_public.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support