ValueDiff: Value-Geometric KV Cache Eviction for Sink-Suppressed LLMs

arXiv:2609.23314 · cs.LG, cs.AI · Submitted 2026-09-20 · Read on arXiv

cs.LG, cs.AI

Submitted: 2026-09-20

Updated: 2026-09-20

Comments: 10 pages, 3 Figues

Code: https://github.com/gkamradt/LLMTest_NeedleInAHaystack

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Terminology

• KV cache — 17× in this paper · explained in 3 episodes
• Cosine Similarity — 10× in this paper · explained in 3 episodes
• Decode — 10× in this paper · explained in 2 episodes
• Alignment — 8× in this paper · explained in 1 episode
• Benchmark — 7× in this paper · explained in 2 episodes
• Calibration — 5× in this paper · explained in 7 episodes
• Transformer — 5× in this paper · explained in 2 episodes
• Accuracy — 4× in this paper · explained in 1 episode
• Fidelity — 3× in this paper · explained in 2 episodes
• Ablation — 2× in this paper · explained in 1 episode
• Empirical Methods in Natural Language Processing — 2× in this paper
• In International Conference on Learning Representations — 2× in this paper
• In International Conference on Machine Learning — 2× in this paper
• Reconstruction Error — 2× in this paper · explained in 2 episodes
• Trace — 2× in this paper · explained in 2 episodes
• Aggregation — 1× in this paper · explained in 1 episode
• Anisotropy — 1× in this paper · explained in 3 episodes
• Edge Devices — 1× in this paper · explained in 2 episodes
• Failure Modes — 1× in this paper · explained in 3 episodes
• Quantization — 1× in this paper · explained in 12 episodes
• Sparsity — 1× in this paper · explained in 4 episodes
• Speculative Decoding — 1× in this paper · explained in 2 episodes
• Transformer Models — 1× in this paper · explained in 2 episodes

Sources

• Apple Intelligence Foundation Language Models
• Fast On-device LLM Inference with NPUs
• Accelerating Mobile Language Model via Speculative Decoding and NPU-Coordinated Execution
• Attention Sink in Transformers: A Survey on Utilization, Interpretation, and Mitigation
• Why do LLMs attend to the first token?
• ManifoldKV: Training-Free KV Cache Compression via Euclidean Outlier Detection
• A Simple and Effective $L_2$ Norm-Based Strategy for KV Cache Compression
• Gated Attention for Large Language Models: Non-linearity, Sparsity, and Attention-Sink-Free
• gpt-oss-120b & gpt-oss-20b Model Card
• Gemma 2: Improving Open Language Models at a Practical Size
• KVzip: Query-Agnostic KV Cache Compression with Context Reconstruction
• CAOTE: KV Cache Selection for LLMs via Attention Output Error-Based Token Eviction
• What are you sinking? A geometric approach on attention sink
• The Llama 3 Herd of Models
• Qwen2.5 Technical Report
• RULER: What's the Real Context Size of Your Long-Context Language Models?
• Measuring Mathematical Problem Solving With the MATH Dataset
• Let's Verify Step by Step
• Anisotropy Is Inherent to Self-Attention in Transformers

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