Composing continual learning mechanisms lifts retention from 1.2% to 34.9%
DanielKhashabi · x · 2026-09-09
Johns Hopkins researchers release ComposeCL, showing that simply composing existing continual learning mechanisms dramatically extends long-horizon memorization.
- Combines weight anchoring (Synaptic Intelligence), function anchoring (self-distillation), data anchoring (generative replay), and merged LoRA in a full 2⁴ factorial experiment across 100 sequential tasks per dataset, 3 seeds
- Best composition raises average final retention from 1.2% (naive sequential SFT) to 34.9% — a 28-fold improvement — ranking top-3 on all three datasets
- Replay and merged LoRA give the largest gains, with joint benefit exceeding the sum of individual gains on all datasets
- Paper, code, and datasets are open-sourced (arXiv, GitHub, alphaxiv)
More from Research
- DeepMind paper: LLMs can't sample randomly on their own, only convert given seeds — rohanpaul_ai · 2026-10-27
- DeepMind paper: LLMs can't sample randomly on their own — pass random numbers from code instead — rohanpaul_ai · 2026-10-27
- World Model Assigns Weight and Friction to Objects, Then Simulates with a Physics Engine — jonstephens85 · 2026-10-27
- Token Collapse: Study Finds Graph Soft Prompting Tokens in LLMs Fail to Encode Structure — Sapienza · 2026-10-27
- Mathathon 2.0 launches with Terence Tao's SAIR, $1.5M in compute for teams — _sathvikr · 2026-10-27
- The 'sweet lesson' is real: two kinds of ML problems need opposite strategies — mike64_t · 2026-10-27