A learned memory gate decided which incoming examples needed supervision and which historical examples to replay, reaching 93.66 percent on WikiBigEdit after only 3K facts
AspireAI community · arXiv (cs.CL)
Continual learning for language models has a bookkeeping problem: given a new fact, should you train on it directly, and which older examples should you pull back in to avoid forgetting? Most replay systems answer with a fixed recipe. CAMR instead learns the answer. It combines compatibility checks with a learned memory gate that drives an online policy over three actions - learn the incoming example with general replay, jointly reinforce compatible related examples, or skip direct supervision of an example already verified as mastered - and the policy is trained on acquisition, retention and training-cost feedback under a bounded per-arrival replay budget.
Evaluated with Qwen3-8B-Base, CAMR leads the evaluated methods on WikiBigEdit's Original Match metric from 3K up to 100K facts, hitting 93.66 percent after just 3K facts and overtaking Full-history Replay with LoRA only after 100K facts. It also posts the highest final scores on all four selected TRACE tasks and holds earlier-task performance at or above post-acquisition levels across a sequence spanning financial question answering, mathematical reasoning, code generation and summarization.
The scope is worth stating plainly: one base model and four selected TRACE tasks, so the ordering against other replay methods is not shown to hold at larger scale or across model families. The results also cover factual updating and a mixed-task sequence rather than open-ended continual learning, and the replay budget is bounded by the method's own design.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.