One shared KV latent across adjacent layers kept 99% of accuracy at 3.6x compression, where the best cross-layer method fell to 84% at 16x
AspireAI community · arXiv
KV-cache memory is often the main bottleneck in agent serving, and cross-layer compression methods that share one cache across adjacent layers tend to degrade sharply at high compression ratios. The authors find why: adjacent layers' KV representations are quite different, while their input hidden states are highly similar - and that shrinking the dimension of a shared latent loses far more accuracy than quantizing it.
BellowsKV is a training-free method that maps the hidden states of n adjacent layers to one shared latent per token, with each layer reading its own keys and values back through a layer-specific projection. The latent is kept at full dimension and compressed two other ways: BellowsKV-Quant stores it at low precision, which is nearly lossless because the latent has no dominant outlier channels, and BellowsKV-Evict makes one shared eviction decision per block because layers rank token importance very similarly. Across Llama-3.1-8B, Mistral-7B and Qwen2.5-7B it retains about 99% of full-cache accuracy at 3.6x compression and 97-98% at 13-14x, where the strongest cross-layer method retains at most 84% at 16x.
The honest limit: the memory saving comes from quantization and eviction rather than a smaller latent, so the headline compression ratio is partly bought by precision rather than structure.
https://arxiv.org/abs/2610.35023
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.