The authors formulate multi-image reasoning as a visual budget allocation problem: given a set of images and a question, mmVB predicts the minimum visual scale (resolution) needed for each image.
Lightweight module cuts visual tokens 9x in multi-image LLMs
mmVB assigns query-conditioned image budgets, speeding inference 5x while keeping 94.4% of full-resolution accuracy.
Big Tech
Aimon Rahman · Joymallya Chakraborty · Vidit Bansal · Abhishek Dan
Amazon
Research Digest··2 min read
Rahman et al.
Why this paper
From Amazon
In one line
A question-conditioned module assigns per-image visual token budgets before encoding, reducing tokens 9x and prefill FLOPs 7x while keeping 94.4% accuracy.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§