本文由 AI 分析生成
建立時間： 2026-04-07 來源： https://x.com/carnot_cyclist/status/2041479655035679163
Summary
A principled definition of continual learning (CL) in LLMs grounded in classical ML literature, arguing that most existing approaches (self-distillation, memory scaffolds, replay, regularization) are solving the wrong problem. The author proposes five desiderata for genuine CL: preserve general performance, sequential learning, different-distribution data, efficiency, and compositionality across time-separated skills. The article then makes the case for parametric (in-weights) CL over harness-based memory, with two arguments: scaling limits and automaticity.
本文提出 LLM 持續學習的五項原則性標準（保留通用能力、序列學習、分布外數據、效率、跨時間組合性），並論證為何 harness 式記憶（RAG、向量庫、skill 檔案）在規模擴大時達到天花板。作者主張參數化（權重內）持續學習是唯一具有複利潛力的路徑。
Key Points
- Five CL desiderata: (1) preserve general performance, (2) sequential (not simultaneous multitask) learning, (3) different-distribution data, (4) efficiency (no replaying all prior data), (5) compositionality (skills from different time-points transfer and combine)
- Multi-task learning avoids catastrophic forgetting by training on all distributions simultaneously — CL must transcend this without it
- Harness-based memory (RAG, vector stores, skill .md files) hits hard ceiling: context rot and retrieval difficulties as knowledge bank grows to hundreds/thousands of items
- Parametric knowledge changes intelligence per forward pass — compounding potential; harness memory keeps same intrinsic intelligence operating over growing bank
- Automaticity argument: neural pretrained knowledge enables generalizable cross-domain recombination that explicit retrieval cannot replicate
- Cartridges (corpus-specific KV caches) noted as principled but has unknown compositional properties at scale (how to use 1000 KV caches?)
Insights
The automaticity argument is the strongest part: the example of how “pretrained coding knowledge aids general reasoning” while “retrieving code snippets would not exhibit the same effect” demonstrates that parametric knowledge creates representations that automatically generalize, bypassing the retrieval bottleneck. This directly challenges harness-first architectures as a permanent solution to continual learning. The five desiderata provide a clean test for evaluating whether any proposed CL method is actually solving the real problem or just a convenient subset of it.
Connections
Raw Excerpt
“TLDR: we are interested in an LLM being able to efficiently and compositionally learn new capabilities during sequential exposure to new, differently-distributed data, while at least preserving general capabilities.”