Liu et al.
Low-rank KV cache compression aligned to attention distributions
BELT uses a learned global rank allocator and a key-side objective targeting post-softmax attention to achieve aggressive compression with minimal accuracy loss.
Chinese Tech
Yaoyang Liu · Shuwei Shen · Jincheng Tang · Jinguang Liu · Jinbao Xue · Hailong Yang · +2 more
Peking University · JD.com · Tencent · BeiHang University
Research Digest··3 min read
The authors propose BELT, a low-rank key-value cache compression framework for large language models that preserves attention behavior rather than raw key or pre-softmax logit fidelity.
Why this paper
From Tencent and 3 others
In one line
BELT compresses LLM KV caches by up to 70% while improving LongBench accuracy by 23.3 points over the strongest baseline.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ✓Reports numbers on named benchmarks (2 benchmarks)
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§