Composing continual learning mechanisms lifts long-horizon retention from 1.2% to 34.9%
mdredze · x · 2026-09-09
The paper "Continual Learning Mechanisms Compose for Long-Horizon Memorization" tackles long-horizon memorization: learning many tasks sequentially, catastrophic forgetting in sequential fine-tuning loses about 98.8% of knowledge after 100 tasks.
The authors combine four existing mechanisms — generative replay, self-distillation, weight regularization, and merged LoRA — each preserving complementary forms of prior information.
Together they reach 34.9% final retention, with replay + merged LoRA showing super-additive benefits. The finding: composing existing mechanisms is a very strong yet simple recipe.
More from Research
- EdgeRock open-sourced: 72 MiB KV cache at 256K context, 99.1% smaller than Qwen3.5-4B — alifcoder · 2026-10-27
- EdgeRock-4B holds KV cache at 72 MiB across 256K context, 99.1% smaller than Qwen3.5-4B — alifcoder · 2026-10-27
- TRL adds Score Centering, a simple fix for RLHF training instability — lvwerra · 2026-10-27
- Latent-Foresight: world models that learn to predict their own future — abursuc · 2026-10-27
- UK AISI papers: no good safeguards yet against misuse of biological AI tools — davidmanheim · 2026-10-27
- CMU's XYBench: LLMs answer queries literally and reinforce user misconceptions — valentina__py · 2026-10-27