Composing Continual Learning Mechanisms Boosts 100-Task Memorization Retention 28x to 34.9%
DanielKhashabi · x · 2026-09-09
- New research on "long-horizon memorization": models learn 100 query-answer tasks via sequential SFT without retaining earlier examples or task IDs at inference.
- Naive sequential fine-tuning causes catastrophic forgetting (1.2% retention after 100 tasks); no single continual learning mechanism maintained strong retention at this horizon.
- The authors organize compositions along two dimensions: data/function/weight anchors (what prior information each update preserves) and low-rank allocation rules (where successive updates live).
- They use task-level successive halving to search the combinatorial space plus a factorial experiment to measure individual and interaction effects, validated on three distinct 100-task datasets.
- Best method combines all three anchors with merged LoRA: top-3 on all datasets, raising average final retention from 1.2% to 34.9% (28x). Data anchors and merged LoRA give the largest gains and interact super-additively across all three datasets.
More from Research
- Deep Learning Course Adds Mechanistic Interpretability Lecture This Winter — CSProfKGD · 2026-10-26
- INCLUDE-2.0 benchmark evaluates regional knowledge across 91 languages and 86 countries — ABosselut · 2026-10-26
- CoRL paper: flow-matching VLAs can run 1-step Euler inference with no retraining — kaixhin · 2026-10-26
- CoVer at EMNLP 2026: Training Code LLMs to Generate and Self-Verify Simultaneously — algo_diver · 2026-10-26
- Robotic dexterity should be judged on speed, force control and tool use, not copying humans — DJiafei · 2026-10-26
- Turing-complete malware can run in PyTorch's backward pass via a zero-valued loss — retr0jirachi · 2026-10-26