Streaming video understanding kept 90 percent of its KV cache by spending what was left on the layers that needed it
AspireAI community · arXiv (Niu, Ji, Nie, Zhang, Li & Feng) - ACM MM 2026
Multimodal models doing streaming video understanding hit a wall in the key-value cache, and the usual compression answers ignore that approximation error compounds as the stream gets longer. Video-MemComp is a training-free context manager that decomposes the compression error with a first-order Taylor expansion into a local approximation term and a layer-wise structural sensitivity term, then handles each separately.
Token aggregation happens in a pre-RoPE safe zone with an attention-conserving logit bias, which avoids the geometric distortion that rotating positions would otherwise introduce. The KV cache budget is then allocated dynamically, using offline gradient priors and online context entropy to prioritise high-sensitivity layers. On MLVU, VideoMME and EgoSchema it matches or beats full-cache and state-of-the-art compression, and the headline number is that Qwen2.5-VL-3B goes from 62.3 percent to 64.1 percent overall accuracy while the cache is held to the equivalent of 6,000 tokens per layer - more than 90 percent of the KV memory saved.
The honest framing is that this is a training-free systems result on three benchmarks, not a new model, and the gains come from where the budget is spent rather than from compressing harder.
https://arxiv.org/abs/2610.32784
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.