If you want to understand what a neural network has learned, one of the most promising approaches is to decompose its dense internal activations into sparse, human-interpretable features. Sparse autoencoders do exactly this — and in 2025, they have become the workhorse of mechanistic interpretability. But a fundamental scaling problem has threatened their utility: as you train larger SAEs to capture more concepts, the very features you are trying to understand begin to fragment, absorb each other, or merge into uninterpretable composites.
The Dictionary Size Dilemma
Sparse autoencoders work by training an encoder-decoder pair to map a model's activations into a sparse, overcomplete representation — a "dictionary" of features where each active feature ideally corresponds to a single human-interpretable concept. The appeal is clear: instead of trying to make sense of thousands of entangled neurons, researchers can examine a set of cleanly separated features that the model actually uses.
But scaling reveals three pathological behaviors. Feature splitting occurs when a general concept (punctuation marks) fragments into overly specific variants (question marks, commas, periods), losing the unified concept that the model functionally uses. Feature absorption happens when a specialized feature develops blind spots in the general feature it was split from — a latent for female names that activates on Mary, Jane, and Sarah might stop activating on Lily once a Lily-specific feature emerges. Feature composition merges conceptually independent features (red and triangle) into a single composite (red triangle) to minimize active latents.
These problems become worse at larger dictionary sizes, creating a painful tension: bigger dictionaries capture more concepts but produce less reliable features. The standard workaround — training multiple SAEs at different sizes and comparing them — is expensive and provides no guarantee of consistent feature identity across scales.
Matryoshka Sparse Autoencoders: Nested Dictionaries
Bussmann, Nabeshima, Karvonen, and Nanda (2025) propose an elegant solution inspired by Matryoshka Representation Learning — the technique of training nested representations within a single embedding vector. Matryoshka SAEs simultaneously train multiple nested sub-SAEs of increasing dictionary size, where each sub-SAE must reconstruct the input using only its subset of the total latents.
This nested structure creates a natural hierarchy: early latents in the smaller sub-SAEs are pressured to learn general, high-level concepts (since they must reconstruct inputs with fewer features), while later latents in the larger sub-SAEs specialize in fine-grained distinctions. Crucially, the nesting prevents later specialized features from absorbing or fragmenting earlier general ones — the general features must remain intact because they carry the reconstruction burden for the smaller sub-SAEs.
Experiments on Gemma-2-2B and TinyStories demonstrate that Matryoshka SAEs produce more disentangled concept representations (measured by maximum decoder cosine similarity), reduce feature absorption rates, and improve performance on sparse probing and targeted concept erasure tasks compared to standard SAEs at equivalent dictionary sizes.
The trade-off is a modest increase in reconstruction error — the nested constraints slightly limit the autoencoder's ability to minimize loss. But the authors argue that the improved quality of learned latents more than compensates, particularly for practical applications where feature interpretability matters more than reconstruction fidelity.
The Geometry of Concept Representations
Complementary work by Hindupur, Lubana, and Fel (2025) explores the theoretical foundations of SAE-based interpretability, examining the duality between sparse autoencoders and the underlying geometry of concept representations. Their analysis reveals that assumptions embedded in SAE training — particularly about the linear separability of features — have direct consequences for what kinds of concepts the autoencoder can discover and how faithfully they are represented.
Meanwhile, Kulkarni, Weng, and Narayanaswamy (2025) bridge two traditions by combining concept bottleneck models with sparse autoencoders, creating systems that are both mechanistically interpretable (through sparse decomposition) and conceptually steerable (through bottleneck-layer intervention). This synthesis suggests that the historically separate communities of mechanistic interpretability and concept-based explainability are converging on shared tools and representations.
Why Sparse Autoencoders Matter Now
The practical importance of SAEs extends beyond academic interpretability research. As frontier AI labs face increasing pressure to understand what their models know and how they reason, SAEs provide one of the few scalable tools for concept-level analysis. They have been used to study attention mechanisms in GPT-2, analyze feature circuits, and perform model comparison across architectures.
For AI safety specifically, SAEs enable a form of targeted oversight: identify the features associated with deceptive reasoning or harmful knowledge, then monitor or modify those features during deployment. The ability to precisely erase specific concepts (as demonstrated by Matryoshka SAEs' improved concept erasure performance) has direct applications in content moderation, bias mitigation, and compliance with regulations that require models to "forget" certain information.
Open Questions
The core question remains: do SAE features correspond to the features that the model actually uses during computation, or are they artifacts of the autoencoder's training objective? The distinction matters because interventions based on spurious features would be unreliable. The non-monotonic relationship between dictionary size and feature quality — where bigger is not always better — suggests that our understanding of the SAE-model relationship remains incomplete.
How should the research community standardize SAE evaluation? Current metrics (reconstruction loss, sparsity, downstream probing accuracy) capture different aspects of quality, and optimizing one can degrade others. A unified evaluation framework that balances interpretability, faithfulness, and computational cost is still missing.
Looking Forward
The Matryoshka approach represents a conceptual shift in how we think about interpretability tools: rather than training separate models at different granularities, build a single tool that naturally organizes features across levels of abstraction. As interpretability transitions from research curiosity to regulatory requirement, this kind of efficient, multi-scale analysis will become increasingly valuable.
References
Bussmann, B., Nabeshima, N., Karvonen, A., & Nanda, N. (2025). Learning multi-level features with Matryoshka sparse autoencoders. arXiv preprint, arXiv:2503.17547.
Hindupur, S. S. R., Lubana, E., & Fel, T. (2025). Projecting assumptions: The duality between sparse autoencoders and concept geometry. arXiv preprint, arXiv:2503.01822.