A comprehensive index of artificial intelligence and machine-learning research with AI-generated summaries, citation metrics, and direct links to papers and code.
Unknown
This paper introduces a framework for reliable tool-using agents under communication constraints, focusing on control compression and reliability frontiers.
Nikhil Verma
Focus, an agent-centric architecture inspired by slime mold, enables LLM agents to autonomously compress context, achieving 22.7% token reduction with identical accuracy on SWE-bench Lite.
Xingyu Tan, Xiaoyang Wang, Qing Liu, et al.
SkillZip is a contract-preserving graph compression framework that compresses skill libraries into reversible macros, achieving 3.46x compression with high dependency preservation and verifier reachability.
Unknown
CacheGen reduces KV cache loading delay by 1.67-1.81x versus 8-bit quantization while achieving nearly lossless compression for fast LLM serving.
Unknown
ChunkKV is a novel KV cache compression method that retains the most informative semantic chunks to improve efficiency in long-context LLM inference.
Unknown
Hcattention achieves extreme KV cache compression via heterogeneous attention computing, combining sparse key selection and value offloading.
Unknown
This paper introduces HIGGS, a novel quantization method for large language models that leverages the linearity theorem to push the limits of compression while preserving performance.
Jieke Shi, Junda He, Wenjia Jiang, et al.
Brevis formulates lossless tensor compression as program synthesis, using a typed DSL and A* search to produce compact, bit-exact reconstruction programs for model checkpoints.
Unknown
This paper explores how task type influences KV cache compression and proposes DynamicKV, a task-aware adaptive compression method for long-context LLMs.
Unknown
ThinKV is a thought-adaptive KV cache compression framework that exploits attention sparsity patterns to reduce memory usage in reasoning models.
Unknown
This paper systematically categorizes and explores KV cache compression techniques for transformer-based models, providing a structured analysis of methods to reduce memory overhead.
Gail A. Carpenter, Boriana L. Milenova, Benjamin W. Noeske
Distributed ARTMAP combines MLP and ART advantages for fast, stable supervised learning with improved memory compression.