A comprehensive index of artificial intelligence and machine-learning research with AI-generated summaries, citation metrics, and direct links to papers and code.
Unknown
Q-dit introduces a post-training quantization method for diffusion transformers that reduces memory and latency while maintaining generation quality.
Unknown
Introduces a KV cache compression method that estimates attention from the distribution of future queries to reduce memory usage in large language model inference.
Unknown
MoMKD introduces a momentum memory mechanism for knowledge distillation to improve model compression in computational pathology.
Unknown
The paper proposes that long-context language models achieve episodic-like order memory through temporal context reinstatement, a mechanism inspired by human memory.
Unknown
Agent Memory Distillation (AMD) transfers hierarchical teacher memory to small LLM agents, enabling efficient knowledge transfer for agent memory.
Unknown
MemGuard proposes persisting verifier signals to govern what LLM-agent memory systems store and when stored information may influence later decisions.
Unknown
ChronoMem introduces version control and semantic rollback for LLM agent memory, enabling commit semantics and global rollback via natural language.
J. Feldmann, N. Youngblood, M. Karpov, et al.
An integrated photonic tensor core using soliton microcombs and phase-change memory achieves over 10^12 MAC operations per second for parallel convolutional processing.
Wei Li, Rui Shao, Jie He, et al.
UniMPA unifies memory, prediction, and action via an action-grounded transition interface to resolve transition realizability gaps in robotic manipulation.
Mohammad Tavakoli, Alireza Salemi, Carrie Ye, et al.
Introduces BEAM, a benchmark for long-term memory in LLMs with up to 10M tokens, and LIGHT, a memory framework inspired by human cognition that improves performance by 3.5%-12.69%.
Rujikorn Charakorn, Edoardo Cetin, Shinnosuke Uesaka, et al.
Doc-to-LoRA (D2L) is a lightweight hypernetwork that meta-learns to perform approximate context distillation in a single forward pass, generating LoRA adapters for a target LLM to reduce inference latency and KV-cache memory.
Lukas Radl, Michael Steiner, Mathias Parger, et al.
StopThePop introduces a hierarchical rasterization approach for 3D Gaussian Splatting that eliminates popping artifacts and view inconsistencies with minimal overhead, enabling faster rendering and reduced memory.