#Backward Transfer - AI Papers Daily

#Backward Transfer

1 篇

TFGN: Task-Free, Replay-Free Continual Pre-Training Without Catastrophic Forgetting at LLM Scale

TFGN：无需任务标签与回放的LLM级持续预训练架构

🧭 Training ★ 深度解读

Independent Researcher

提出一种内部架构叠加层TFGN，在无回放、无任务ID的条件下，实现LLaMA 3.1 8B等多尺度的近零灾难性遗忘与>99.59%的梯度正交性

Training | 2026-05-14 | cs.LG | score 10/10

Continual Learning Catastrophic Forgetting LLM Pre-training Gradient Orthogonality Backward Transfer