The authors compute a spatial-density score from cosine distances among visual tokens within local image-grid neighborhoods at an early decoder layer.
Spatial density guides faster, safer token pruning in vision-language models
TwinPrune uses one spatial-density score to reduce prefill computation and decode-time memory, while exposing a benchmark flaw that can conceal damage from cache eviction.
Chinese Tech
Kai Liu · Anqi Li · Junxian Li · Zhixin Wang · Zhikai Chen · Renjing Pei · +1 more
Shanghai Jiao Tong University · Huawei Technologies Ltd
Research Digest··2 min read
Liu et al.
Why this paper
From Huawei Technologies Ltd and Shanghai Jiao Tong University
In one line
First-token benchmarks cannot test decode-time eviction, while a shared spatial-density score efficiently prunes visual tokens during both VLM prefill and decode.
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.
§