The authors designed AlignPred, a compression codec that operates on the KV cache of LLMs.
Local alignment enables predictive coding for error-bounded KV cache compression
AlignPred combines orthogonal rotations of related heads or layers with token-wise and cross-block prediction, then encodes residuals under a configurable reconstruction-error budget.
Chinese Tech
Xu Yang · Jiapeng Zhang · Zhiwei Zuo · Yuxin Chen · Feiqiang Sun · Chengguang Xu · +2 more
Hunan University · Tencent · Central China Normal University
Research Digest··3 min read
The authors present AlignPred, a codec for persistent KV caches in long-running LLM agents.
Why this paper
From Tencent and 2 others
In one line
AlignPred compresses persistent KV caches by aligning related blocks and coding residuals under a configurable error budget.
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.
§