Superposition Hypothesis

- The Superposition Hypothesis is a proposed explanation in Mechanistic Interpretability for why individual neurons in neural networks respond to multiple, apparently unrelated features — a phenomenon called Polysemanticity. The hypothesis holds that networks learn to represent more features than they have neurons by exploiting the approximate orthogonality of high-dimensional space: many sparse feature vectors can be packed into a smaller space with minimal interference, as long as the features rarely co-occur.

The Core Mechanism

The mathematical intuition behind superposition is simple but powerful. In high-dimensional spaces, randomly chosen vectors are approximately orthogonal. If features are sparse — meaning they are active only rarely and independently — then a network can represent N features using M neurons where N >> M by assigning each feature a direction in activation space. The interference between features is minimal because the probability of two features being simultaneously active (and thus their vector sum being needed) is low.

Elhage et al. (Anthropic, 2022) formalized this intuition in 'Toy Models of Superposition.' They trained small two-layer networks on synthetic tasks where the number of features exceeded the number of neurons, and showed that the networks learned superposed representations: each neuron responded to multiple features, but the features could be recovered by analyzing the geometry of the activation space. The key finding was that superposition is not a pathology but an efficient coding strategy — the network is doing the best it can with limited capacity.

Sparse Autoencoders and Feature Recovery

The primary empirical method for testing the superposition hypothesis is the sparse autoencoder. A sparse autoencoder is trained to reconstruct network activations while enforcing sparsity constraints (typically L1 regularization) on the hidden layer. The hypothesis predicts that:

1. Sparse autoencoders will discover more features than there are neurons in the original network 2. These features will be approximately orthogonal and sparse 3. The features will correspond to human-interpretable concepts

Results on small models have been promising. On larger models, the picture is more mixed. Sparse autoencoders do discover large numbers of features, but the features are not always interpretable, and the relationship between autoencoder features and the network's actual computations remains unclear. The method may be recovering superposed features, or it may be discovering statistical regularities that do not correspond to the network's causal structure.

Implications for AI Safety

If the superposition hypothesis is correct, it has profound implications for AI Safety. The most concerning is the possibility of latent misalignment: aligned and misaligned objectives could coexist in superposition, with misaligned features remaining dormant under normal operating conditions. A network might appear safe in standard testing while containing latent features that would activate under specific, rare inputs — inputs that an adversary could craft or that might occur naturally in deployment.

This possibility transforms the alignment problem from a question of 'does the network behave safely?' to 'does the network contain any unsafe features, even latent ones?' The second question is much harder to answer, because latent features by definition do not affect behavior under typical conditions. Sparse autoencoders might help detect them, but only if the autoencoder is itself aligned — a regress that has no obvious termination point.

Open Questions

The superposition hypothesis remains empirically contested. Key open questions include:

• Does superposition occur in frontier models (GPT-4 scale), or is it primarily a phenomenon of small, overparameterized toy models?
• Are sparse autoencoders recovering the 'true' features the network uses, or are they imposing a structure that is not causally relevant?
• Can we predict, from architecture and training data, which features will be superposed and which will be represented monosemantically?
• If latent misalignment is real, can it be detected and removed before deployment?

The empiricist position on the hypothesis demands testing it against frontier models, not just toy networks. The results from mechanistic interpretability work on large models remain preliminary — and the stakes of getting this wrong are high.