When to Evict, Not What to Keep: Draft-Guided Eviction for Training-Free KV-Cache Compression

arXiv:2609.33334 · cs.LG, cs.CL · Submitted 2026-09-27 · Read on arXiv

cs.LG, cs.CL

Submitted: 2026-09-27

Updated: 2026-10-08

Terminology

• Decode — 67× in this paper · explained in 2 episodes
• Deferral — 25× in this paper · explained in 2 episodes
• KV cache — 16× in this paper · explained in 3 episodes
• Draft-Guided Eviction — 8× in this paper
• Few-Shot Learning — 8× in this paper · explained in 4 episodes
• Ablation — 7× in this paper · explained in 1 episode
• Steering — 6× in this paper · explained in 2 episodes
• Adaptation — 3× in this paper · explained in 1 episode
• Quantization — 3× in this paper · explained in 12 episodes
• Speculative Decoding — 3× in this paper · explained in 2 episodes
• Accuracy — 2× in this paper · explained in 1 episode
• Alignment — 2× in this paper · explained in 1 episode
• Benchmark — 2× in this paper · explained in 2 episodes
• Benchmarking — 2× in this paper · explained in 3 episodes
• Instruction Tuning — 2× in this paper · explained in 2 episodes
• Transformer — 2× in this paper · explained in 2 episodes
• Adam — 1× in this paper · explained in 2 episodes
• Empirical Methods in Natural Language Processing — 1× in this paper
• Failure Modes — 1× in this paper · explained in 3 episodes
• Fidelity — 1× in this paper · explained in 2 episodes
• Pruning — 1× in this paper · explained in 4 episodes
• Sparsity — 1× in this paper · explained in 4 episodes

Sources

• LookaheadKV: Fast and Accurate KV Cache Eviction by Glimpsing into the Future without Generation
• PyramidKV: Dynamic KV Cache Compression based on Pyramidal Information Funneling
• The Llama 3 Herd of Models
• Ada-KV: Optimizing KV Cache Eviction by Adaptive Budget Allocation for Efficient LLM Inference
• CriticalKV: Optimizing KV Cache Eviction from an Output Perturbation Perspective
• Not All Heads Matter: A Head-Level KV Cache Compression Method with Integrated Retrieval and Reasoning
• Mistral 7B
• KVzip: Query-Agnostic KV Cache Compression with Context Reconstruction
• Qwen2.5 Technical Report
• D2O: Dynamic Discriminative Operations for Efficient Long-Context Inference of Large Language Models
• LOOK-M: Look-Once Optimization in KV Cache for Efficient Multimodal Long-Context Inference
• Lookahead Q-Cache: Achieving More Consistent KV Cache Eviction via Pseudo Query
• Model Tells You Where to Merge: Adaptive KV Cache Merging for LLMs on Long-Context Tasks
• ThinK: Thinner Key Cache by Query-Driven Pruning
• Rethinking KV Cache Eviction via a Unified Information-Theoretic Objective
• Long Context Compression with Activation Beacon
• When Does Value-Aware KV Eviction Help? A Fixed-Contract Diagnostic for Non-Monotone Cache Compression
• Fast KV Compaction via Attention Matching

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