KV-Kaizen: Learning Context-Adaptive Cache Compression Choices
cs.AI
Submitted: 2026-09-29
Updated: 2026-09-30
Code: https://github.com/ml-explore/mlx
Sources
- Qwen2.5 Technical Report
- PyramidKV: Dynamic KV Cache Compression based on Pyramidal Information Funneling
- Nectar: Neural Estimation of Cached-Token Attention via Regression
- DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model
- Reducing Transformer Key-Value Cache Size with Cross-Layer Attention
- Stochastic KV Routing: Enabling Adaptive Depth-Wise Cache Sharing
- Gemma 4 Technical Report
- Categorical Reparameterization with Gumbel-Softmax
- The Concrete Distribution: A Continuous Relaxation of Discrete Random Variables
- Estimating or Propagating Gradients Through Stochastic Neurons for Conditional Computation
- Instruction-Following Evaluation for Large Language Models
- Training Verifiers to Solve Math Word Problems
- RULER: What's the Real Context Size of Your Long-Context Language Models?
- Mistral 7B
- Model Tells You What to Discard: Adaptive KV Cache Compression for LLMs
- Fast Transformer Decoding: One Write-Head is All You Need
- KVSharer: Efficient Inference via Layer-Wise Dissimilar KV Cache Sharing
- CommonKV: Compressing KV Cache with Cross-layer Parameter Sharing
- Cross-layer Attention Sharing for Pre-trained Large Language Models
- DeepSeek-V3 Technical Report
Related papers
- MAVEN-T: Reinforced Heterogeneous Distillation for Real-Time Multi-Agent Trajectory Prediction
- Model Discovery Agent: LLM-assisted Bayesian experiment design for data-efficient discovery of mechanistic world models
- Agents' Last Exam
- The Clinician's Veto: Navigating Trust, Liability, and Uncertainty in Autonomous AI Prescribing
- MindHelper: Closed-Loop Embodied Mental-State Reasoning for Precision Intervention
- Incumbent Advantage: Brand Bias and Cognitive Manipulation Dynamics in LLM Recommendation Systems