Paper Note | JJ's Blog

📝 KVFlow: Efficient Prefix Caching for Accelerating LLM-Based Multi-Agent Workflows

📝 ThunderAgent: A Simple, Fast and Program-Aware Agentic Inference System - ICML'26 Spotlight

📝 DroidSpeak: KV Cache Sharing Across Fine-tuned Model Variants - NSDI'26

📝 EPIC: Efficient Position-Independent Caching for Serving Large Language Models - ICML'25

📝 TwinQuant: Learnable Subspace Decomposition for 4-Bit LLM Quantization - ICML'26

📝 LEARNPRUNER: RETHINKING ATTENTION-BASED TOKEN PRUNING IN VISION LANGUAGE MODELS - ICLR'26

📝 DFlash: Block Diffusion for Flash Speculative Decoding - ICML'26

📝 DOCPRUNE: Efficient Document Question Answering via Background, Question, and Comprehension-aware Token Pruning - CVPR'26

📝 Empirical Recipes for Efficient and Compact Vision-Language Models

📝 Variation-aware Vision Token Dropping for Faster Large Vision-Language Models - CVPR'26

📝 PRUNE REDUNDANCY, PRESERVE ESSENCE: VISION TOKEN COMPRESSION IN VLMS VIA SYNERGISTIC IMPORTANCE–DIVERSITY - ICLR'26

📝 MMTok: Multimodal Coverage Maximization for Efficient Inference of VLMs - ICLR'26

📝 ONE PATCH DOESN’T FIT ALL: ADAPTIVE PATCHING FOR NATIVE-RESOLUTION MULTIMODAL LARGE LANGUAGE MODELS - ICLR'26

📝 Video Compression Commander: Plug-and-Play Inference Acceleration for Video Large Language Models - EMNLP'25

📝 TALON: Confidence-Aware Speculative Decoding with Adaptive Token Trees

📝 VLCACHE: Computing 2% Vision Tokens and Reusing 98% for Vision–Language Inference

📝 Accelerating Streaming Video Large Language Models via Hierarchical Token Compression - CVPR'26

📝 HyperVL: An Efficient and Dynamic Multimodal Large Language Model for Edge Devices

📝 VisionThink: Smart and Efficient Vision Language Model via Reinforcement Learning - NeurIPS'25

📝 SpecBranch: Speculative Decoding via Hybrid Drafting and Rollback-Aware Branch Parallelism - ICLR'26

📝 Taming the Long-Tail: Efficient Reasoning RL Training with Adaptive Drafter - ASPLOS'26

📝 Empower Vision Applications with LoRA LMM - Eurosys'25

📝 Venus: An Efficient Edge Memory-and-Retrieval System for VLM-based Online Video Understanding - INFOCOM'26

📝 Elastic On-Device LLM Service - MobiCom '25

📝 RServe: Overlapping Encoding and Prefill for Efficient LMM Inference

📝 ModServe - Modality- and Stage-Aware Resource Disaggregation for Scalable Multimodal Model Serving - SoCC'25

📝 Speculate Deep and Accurate - Lossless and Training-Free Acceleration for Offloaded LLMs via Substitute Speculative Decoding - NeurIPS'25

📝 Jupiter: Fast and Resource-Efficient Collaborative Inference of Generative LLMs on Edge Devices - INFOCOM'25

📝 KVzip: Query-Agnostic KV Cache Compression with Context Reconstruction - NeuIPS'25 Oral

📝 SuffixDecoding: Extreme Speculative Decoding for Emerging AI Applications - NeurIPS'25 Spotlight

📝 Breaking the Wall: Unifying Edge GPUs and NPUs into Pipeline Parallelism for Efficient LLM Fine-Tuning

📝 TINY BUT MIGHTY: A SOFTWARE-HARDWARE CO DESIGN APPROACH FOR EFFICIENT MULTIMODAL IN FERENCE ON BATTERY-POWERED SMALL DEVICES - ICLR'26

📝 Efficiently Serving Large Multimodal Models Using EPD Disaggregation - ICML'25

📝 ElasticMM: Efficient Multimodal LLMs Serving with Elastic Multimodal Parallelism - NeurIPS'25 Oral

📝 ChunkKV: Semantic-Preserving KV Cache Compression for Efficient Long-Context LLM Inference - NeurIPS'25

📝 Stop Looking for “Important Tokens” in Multimodal Language Models: Duplication Matters More - EMNLP'25

📝 MoVE-KD: Knowledge Distillation for VLMs with Mixture of Visual Encoders -CVPR 2025

📝 DartQuant: Efficient Rotational Distribution Calibration for LLM Quantization - NeurIPS'25

📝 HEADINFER: Memory-Efficient LLM Inference by Head-wise Offloading

📝 Efficient Multi-modal Large Language Models via Progressive Consistency Distillation - NeurIPS'25

📝 SPINQUANT: LLM QUANTIZATION WITH LEARNED ROTATIONS - ICLR'25

📝 ViSpec: Accelerating Vision-Language Models with Vision-Aware Speculative Decoding - NeurIPS'25

📝 SpecEdge: Scalable Edge-Assisted Serving Framework for Interactive LLMs - NeurIPS'25 Spotlight

📝 Ada-KV: Optimizing KV Cache Eviction by Adaptive Budget Allocation for Efficient LLM Inference - NeurIPS'25

📝 NOT ALL HEADS MATTER: A HEAD-LEVEL KV CACHE COMPRESSION METHOD WITH INTEGRATED RETRIEVAL AND REASONING - ICLR'25

📝 QSVD: Efficient Low-rank Approximation for Unified Query-Key-Value Weight Compression in Low-Precision Vision-Language Models - NeurIPS'25 Spotlight

📝 SPECVLM: Enhancing Speculative Decoding of Video LLMs via Verifier-Guided Token Pruning - EMNLP'25

📝 AdaSPEC: Selective Knowledge Distillation for Efficient Speculative Decoders - NeurIPS'25 Spotlight

📝 R-KV: Redundancy-aware KV Cache Compression for Reasoning Models - NeurIPS'25

📝 Beyond Attention or Similarity: Maximizing Conditional Diversity for Token Pruning in MLLMs - *NeurIPS'25

📝 Cache-to-Cache: Direct Semantic Communication Between Large Language Models - ICLR'26