Learning, Fast and Slow: Towards LLMs That Adapt Continually

The framing of optimized prompts as 'fast weights' is doing real conceptual work here: it borrows from neuroscience-adjacent ML theory (Hinton's fast weights, complementary learning systems) to argue that in-context learning and gradient-based learning are not competing strategies but complementary timescales. The practical implication is that practitioners may not need to choose between prompt engineering and fine-tuning workflows at all.

This connects directly to the test-time adaptation thread running through recent coverage. 'Task-Adaptive Embedding Refinement via Test-time LLM Guidance' (also from May 12) demonstrated that inference-time adjustment can substitute for retraining in embedding models, and this paper extends a similar intuition to the full LLM training loop. Where that work treated test-time adaptation as a practical workaround, this paper attempts to give it theoretical standing within a unified learning framework. The LongMemEval-V2 benchmark coverage is also relevant context: the memory competencies it measures (persistence, accumulated experience) are precisely what catastrophic forgetting destroys, so any architecture that credibly addresses forgetting has direct implications for agent evaluation.

The critical test is whether the dual-timescale framework holds up on continual learning benchmarks like Split-CIFAR or Permuted MNIST variants adapted for language, where catastrophic forgetting is measured directly rather than inferred. If independent replications show retained plasticity after 10-plus sequential tasks without performance degradation on earlier tasks, the theoretical framing earns its weight.