Abstract
Mechanistic interpretability aims to reverse-engineer the algorithms implemented by neural networks by analyzing their internal representations. A central obstacle is polysemanticity: individual neurons in large language models activate for multiple unrelated concepts, making it hard to assign clean interpretations. Sparse autoencoders (SAEs) have emerged as a practical tool for decomposing these entangled representations into a larger set of monosemantic features, each of which activates sparsely and corresponds more cleanly to a human-interpretable concept. Anthropic’s large-scale SAE work on Claude, and parallel efforts from EleutherAI and independent researchers, have brought SAEs from a niche technique to a core interpretability primitive.
Why Polysemanticity Occurs and What SAEs Do About It
The superposition hypothesis, developed by Elhage et al. at Anthropic, argues that networks represent more features than they have neurons by encoding multiple features in overlapping directions in activation space, relying on sparsity of real-world inputs to keep interference manageable. This is computationally efficient but renders individual neurons uninterpretable. A sparse autoencoder is trained on a model’s residual stream or MLP activations: it learns to reconstruct the activation vector from a sparse combination of a larger dictionary of learned features, with an L1 penalty encouraging sparsity. If trained successfully, the learned features correspond to concepts the network actually uses, rather than arbitrary linear combinations imposed by the neuron basis.
What Has Been Found
Anthropic’s release of SAE features from Claude Sonnet internal layers demonstrated that interpretable features exist at scale and are not just researcher artifacts: features for concepts like “the Golden Gate Bridge,” “base64 encoding,” and specific emotional valences were identified and found to causally influence model behavior when artificially activated. This causal test, steering the model by directly modifying feature activations, distinguishes features that are genuinely part of the model’s computation from features that are merely correlated with certain outputs. EleutherAI’s work on GPT-2 and Pythia models has replicated the core finding at smaller scale with open weights, enabling community auditing of the learned dictionaries.
Current Limitations and Open Problems
SAEs are trained post-hoc on frozen models, which means they may not recover all features the model uses, particularly those that only emerge in context across many layers. Dictionary learning is also sensitive to hyperparameters: the sparsity coefficient determines how many features activate per token, and too-aggressive sparsity can fragment concepts across multiple features while too-weak sparsity produces polysemantic features that defeat the purpose. Scaling laws for SAEs are only beginning to be characterized. The most significant open question is whether the features SAEs recover at one layer compose predictably with features at adjacent layers, which would make circuit-level analysis tractable. Early results from Anthropic’s “Scaling Monosemanticity” line of work suggest partial composability, but a complete picture is not yet in hand.