跳到正文
RCreddit.com·
暂不在当前实时榜单

An interesting paper on Continuous (Live) Learning: "Infinite-Parameter LLMs: Generating and Adapting Weights from Live Data"

AI 摘要

A new paper, "Infinite-Parameter LLMs: Generating and Adapting Weights from Live Data," explores continuous learning for LLMs. Building on previous TTT research like Titans and recent Dynamic LoRA adapters, this work introduces a similar concept (SHINE). However, it diverges by generating an adapter that is directly applied to existing weights on-the-fly, rather than using LoRA. The authors employ a rigorous Bayesian belief for uncertainty, though the method could potentially be extended to a continuous version by using standard ML techniques.

时间与来源

时间显示为 UTC

显示时区:UTC

本地时区尚不可用,暂时显示 UTC。

收录当时偏移:UTC+02026年9月18日 12:00 UTC

收录
2026年9月18日 12:00
来源类型
开发者社区

本站未收录正文。

前往源站阅读 →
来源·reddit.com