# From Weights to Features: SAE-Guided Activation Regularization for LLM Continual Learning

arXiv: https://arxiv.org/abs/2606.26629 · HTML: https://mlsanity.com/paper/2606.26629/

## Key facts

- Cited by 2 papers in the map (data through 14 Oct 2026).
- Placed in SAE features and their applications, one of 13 directions in Mechanistic interpretability.
- Builds on 17 of its 24 references that are in the map.
- First posted on arXiv 25 Jun 2026.

## Placed in

- Mechanistic interpretability / [SAE features and their applications](https://mlsanity.com/mechinterp/sae-features-and-their-applications/)

## Builds on (most cited first, 17 of 17)

- [Overcoming catastrophic forgetting in neural networks](https://mlsanity.com/paper/1612.00796/) (2016)
- [Learning without Forgetting](https://mlsanity.com/paper/1606.09282/) (2016)
- [Gradient Episodic Memory for Continual Learning](https://mlsanity.com/paper/1706.08840/) (2017)
- [Sparse Autoencoders Find Highly Interpretable Features in Language Models](https://mlsanity.com/paper/2309.08600/) (2023)
- [Progressive Neural Networks](https://mlsanity.com/paper/1606.04671/) (2016)
- [Memory Aware Synapses: Learning what (not) to forget](https://mlsanity.com/paper/1711.09601/) (2017)
- [Efficient Lifelong Learning with A-GEM](https://mlsanity.com/paper/1812.00420/) (2018)
- [Toy Models of Superposition](https://mlsanity.com/paper/2209.10652/) (2022)
- [Dark Experience for General Continual Learning: a Strong, Simple Baseline](https://mlsanity.com/paper/2004.07211/) (2020)
- [PackNet: Adding Multiple Tasks to a Single Network by Iterative Pruning](https://mlsanity.com/paper/1711.05769/) (2017)
- [Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2](https://mlsanity.com/paper/2408.05147/) (2024)
- [Orthogonal Subspace Learning for Language Model Continual Learning](https://mlsanity.com/paper/2310.14152/) (2023)
- [Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet](https://mlsanity.com/paper/2605.29358/) (2026)
- [TRACE: A Comprehensive Benchmark for Continual Learning in Large Language Models](https://mlsanity.com/paper/2310.06762/) (2023)
- [Natural continual learning: success is a journey, not (just) a destination](https://mlsanity.com/paper/2106.08085/) (2021)
- [MedCL-Bench: Benchmarking stability-efficiency trade-offs and scaling in biomedical continual learning](https://mlsanity.com/paper/2603.16738/) (2026)
- [SAE-FD: Sparse Autoencoder Feature Distillation for Continual Learning of Large Language Models](https://mlsanity.com/paper/2605.25525/) (2026)

## Built on by (most cited first, 2 of 2)

- [Omega-S: A Functional Resilience Index for LLM Fine-Tuning](https://mlsanity.com/paper/2608.03887/) (2026)
- [Decoupling Knowledge and Privacy: Post-Task Self-Distillation Replay for LLM Continual Learning](https://mlsanity.com/paper/2609.29711/) (2026)

## Abstract (arXiv)

Weight-space regularization methods such as Elastic Weight Consolidation (EWC) are the standard approach to catastrophic forgetting in continual learning. However, those methods tend to underperform when applied to large language models. We argue that such underperformance can be partly explained by the ``polysemantic'' nature of large language models: per-weight importance estimates utilized by EWC-style regularization are too coarse and cannot isolate the knowledge that needs protection. In this paper, we propose regularizing instead in the model's activation space, using pretrained Sparse Autoencoders (SAEs) as a monosemantic feature dictionary. From the perspective of constrained optimization, we derive a new loss function that uses the SAE feature dictionary to explicitly balance stability and plasticity, and show that EWC is a special case in the one-sided weight-space penalty setting. Unlike replay-based methods that store or revisit examples from earlier tasks, our method requires no previous-task data after mask construction: current-task data is used to compute a compact SAE feature mask, and only this mask is retained for later training. Further, since the feature space has significantly lower dimensionality than the parameter space, the proposed method is more memory efficient. On the TRACE and MedCL continual learning benchmarks, the method achieves the strongest result among approaches without introducing task-specific architectural components, also surpassing traditional weight-space regularization methods like EWC. Beyond performance comparisons, we provide empirical evidence for the polysemanticity thesis: task-relevant representations are linearly separable in the SAE feature basis but indistinguishable from chance in the weight basis, and weight-space protection is nearly non-selective at the concept level.