Where Activation Sparsity and KV-Cache Sparsity Cross in LLM Decoding

arXiv:2609.33889 · cs.LG, cs.CL, cs.PF · Submitted 2026-09-27 · Read on arXiv

cs.LG, cs.CL, cs.PF

Submitted: 2026-09-27

Updated: 2026-09-27

Code: https://github.com/js-lee-AI/ByteCross

Terminology

• Decode — 50× in this paper · explained in 2 episodes
• Sparsity — 46× in this paper · explained in 4 episodes
• Perplexity — 34× in this paper · explained in 3 episodes
• Activation Sparsity — 15× in this paper · explained in 1 episode
• Transformer — 9× in this paper · explained in 2 episodes
• KV cache — 8× in this paper · explained in 3 episodes
• Speculative Decoding — 7× in this paper · explained in 2 episodes
• Quantization — 5× in this paper · explained in 12 episodes
• Accuracy — 4× in this paper · explained in 1 episode
• Sparse Attention — 4× in this paper · explained in 5 episodes
• Admissibility — 2× in this paper · explained in 2 episodes
• Benchmark — 1× in this paper · explained in 2 episodes
• Elias — 1× in this paper · explained in 2 episodes
• Generative Model — 1× in this paper · explained in 2 episodes
• Generative Models — 1× in this paper · explained in 2 episodes
• Large Language Model (LLM) — 1× in this paper · explained in 3 episodes
• Pruning — 1× in this paper · explained in 4 episodes
• RAG — 1× in this paper · explained in 3 episodes
• Transformer Models — 1× in this paper · explained in 2 episodes
• Verification — 1× in this paper · explained in 2 episodes

Sources

• Taming Throughput-Latency Tradeoff in LLM Inference with Sarathi-Serve
• GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints
• Medusa: Simple LLM Inference Acceleration Framework with Multiple Decoding Heads
• PyramidKV: Dynamic KV Cache Compression based on Pyramidal Information Funneling
• KnapSpec: Self-Speculative Decoding via Adaptive Layer Selection as a Knapsack Problem
• FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness
• LayerSkip: Enabling Early Exit Inference and Self-Speculative Decoding
• The Llama 3 Herd of Models
• FlashDecoding++: Faster Large Language Model Inference on GPUs
• Efficient Memory Management for Large Language Model Serving with PagedAttention
• CATS: Contextually-Aware Thresholding for Sparsity in Large Language Models
• DART: Draft-Agreement Routing for Training-Free Adaptive Thinking Budgets in Hybrid Reasoning Models
• Fast Inference from Transformers via Speculative Decoding
• SnapKV: LLM Knows What You are Looking for Before Generation
• EAGLE: Speculative Sampling Requires Rethinking Feature Uncertainty
• AWQ: Activation-aware Weight Quantization for LLM Compression and Acceleration
• Deja Vu: Contextual Sparsity for Efficient LLMs at Inference Time
• KIVI: A Tuning-Free Asymmetric 2bit Quantization for KV Cache
• Pointer Sentinel Mixture Models
• ReLU Strikes Back: Exploiting Activation Sparsity in 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