WIRE contracts the decoder input before the first layer: a question-conditioned bidirectional selective scan compresses each frame's full-resolution embeddings into carrier tokens, and hypernetworks at selected layers read the compressed forward pass plus a readout of removed features to generate LoRA updates for attention and feedforward projections.
Weight updates carry video context, cutting prefill cost in frozen video LLMs
WIRE compresses visual embeddings into carrier tokens and generates temporary LoRA updates, reducing decoder prefill time by over 3x at 256 to 1024 frames while keeping accuracy competitive.
Top University
Ameen Ali · Tamim Zoabi · Lior Wolf · Chaim Baskin
Tel Aviv University · Ben-Gurion University of the Negev
Research Digest··2 min read
The authors present WIRE, which compresses video embeddings into carrier tokens and generates temporary LoRA weight updates for a frozen video LLM.
Why this paper
From Tel Aviv University and Ben-Gurion University of the Negev
In one line
Temporary low-rank weight updates, generated per video and question, let a frozen video language model run on far fewer visual tokens without losing task-relevant video information.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors (2 noted)
- ✓Reports numbers on named benchmarks
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§