Continual Learning in Transition
Zhiyan Hou, Dan Zhang, Tao Feng, Liyuan Wang, Wei Li, Xiangzhao Hao, Hongyan An, Junfeng Fang, Haokai Ma, Zhaohui Xu, Haiyun Guo, Jinqiao Wang, Tat-Seng Chua
cs.LG, cs.AI
2026-08-07
Reframes continual learning along When, Where, and How axes as it shifts from parameter-centric updates to system-level adaptation; argues frontier-model continual learning is the top priority toward AGI.
Continual learning spent years on one problem: catastrophic forgetting. Learn a sequence of tasks and the parameters that supported earlier tasks get overwritten. The field built four method families around it: replay, gradient projection, architecture isolation, and regularization, with EWC and LwF as familiar names.
These methods differ in mechanism but share three assumptions nobody stated: capability lives in the parameters, learning happens during a dedicated pre-deployment training stage, and updates are gradient steps on data collected beforehand. The authors' argument is that LLMs and agents have broken all three, and the old definition now excludes the most interesting developments: test-time training, on-policy reinforcement learning, external memory, and skill libraries. The paper sets out to give the shift a single coordinate system.
The three assumptions become three questions. Each used to have a fixed answer; now each is a variable.
Classical continual learning is one point in this space: post-training, parameters, off-policy gradients. LLM-era methods depart from it along one or more axes, and a method gets a When-Where-How profile rather than a grid cell. Voyager accumulates skills on the harness through on-policy exploration; TTRL updates parameters at deployment with verifiable rewards; MemoryBank adds an Ebbinghaus forgetting curve to external memory.
This is a position and survey paper with no experiments of its own and no benchmark numbers. Its payoff is analysis, and the most useful piece is the three-layer carrier hierarchy and the argument for why longer context cannot carry continual learning.
| Carrier | Write cost | Recall stability | Capacity ceiling |
| Parameters (innermost) | Highest; needs backprop and alignment data | Highest; genuinely internalized | Scales with model size |
| External memory (middle) | Low; no parameter updates | Medium; bounded by retrieval bandwidth | Storage and retrieval bound |
| Context window (outermost) | Lowest; one inference pass | Most volatile; hard length limit | Hard ceiling |
The central claim is that context has a structural ceiling. There is a hard limit, since even million-token windows are finite, and a soft limit, since the needle-in-a-haystack effect shows effective attention is uneven across positions and information density does not scale linearly with window length. Distilling rich tacit experience into a text summary is brittle outside text-centric domains and can silently discard the rationale behind a hard-won optimization. RAG is no escape either: as dialogue and task chains lengthen, retrieval gets dominated by stale documents, low-quality entries accumulate, and relevance scores drift, so external memory needs an active forgetting mechanism.
The paper also names a common confusion: the sense that engineering progress is approaching AGI rests on visible accumulation on the harness, but a large share of that capability lives in the scaffolding around the model rather than in the model.
For anyone building agents, the implication is that continual learning is now your problem too. Memory decay, skill-library bloat, and protocol drift are each continual-learning questions, not just the fine-tuning-forgetting kind. The three-axis frame at least gives a diagnostic map for which axis a failure sits on.
The practical hook is the cache-hierarchy analogy. What should live in context, what should be written to memory, what should be consolidated into a skill, and what should finally settle into parameters is a scheduling problem, structurally similar to cache-memory hierarchy management. The authors concede no formal framework exists yet for it, but as a thinking tool it already pays off.
The risk with a position paper is that its claims are themselves unverified. The bidirectional transfer the authors propose, where harness capability consolidates back into parameters and is retrieved on demand, plus a unified schedule across carriers, is described throughout as a suggestion, a direction, or an intuition, with no working instance.
Coverage is uneven. The authors' own work concentrates on memory and agents, so the Where axis (the harness) is the most developed, while When and How are thinner. Read as a full map, it overstates how mature the harness route is.
The AGI-priority judgment, that continual learning for the frontier model is the first priority and domain-specific continual learning is not, is a stance rather than a result. The most sober line in the paper is the honest distance estimate: the memory-and-harness route is fast approaching the capability ceiling of a general assistant at the engineering level, but genuine AGI remains far off, and the gap will not close simply through longer context, larger memory, or more skills.