VestigeKV: The NoPE-MLA KV Cache Carries Its Own Sparse-Attention Signal in a Vestigial Branch
cs.LG, cs.CL
Submitted: 2026-09-03
Updated: 2026-09-09
Code: https://github.com/fan-wenjie/vestigekv
Terminology
Sources
- DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model
- Training Transformers for KV Cache Compressibility
- The Impact of Positional Encoding on Length Generalization in Transformers
- Kimi Linear: An Expressive, Efficient Attention Architecture
- SnapKV: LLM Knows What You are Looking for Before Generation
- MatryoshkaKV: Adaptive KV Compression via Trainable Orthogonal Projection
- Kamera: Unified Position-Invariant Multimodal KV Cache for Training-Free Reuse
- Irminsul: MLA-Native Position-Independent Caching for Agentic LLM Serving
- Transformers are Multi-State RNNs
- Quest: Query-Aware Sparsity for Efficient Long-Context LLM Inference
- KeepKV: Achieving Periodic Lossless KV Cache Compression for Efficient LLM Inference
- Length Generalization of Causal Transformers without Position Encoding
- Efficient Streaming Language Models with Attention Sinks
- H$_2$O: Heavy-Hitter Oracle for Efficient Generative Inference of Large Language Models
Related papers
- Polynomial-Augmented Neural Networks (PANNs) with Weak Orthogonality Constraints for Enhanced Function and PDE Approximation
- AIRL-S: Unifying Reinforcement Learning and Search-Based Test-Time Scaling via Adversarial Inverse Reinforcement Learning
- Transformers as Bayesian In-Context Experimenters: Smoothness-Adaptive Efficient ATE Estimation
- Convergence issues in Relational Concept Analysis based on AOC-posets
- Beliefs Beyond Posteriors: Local-Consistency Optimisation for Bayesian Neural Networks
- Understanding Diffusion Models via Ratio-Based Function Approximation with SignReLU Networks