In standard neural networks, individual neurons encode multiple unrelated features in superposition — a consequence of networks learning more features than they have dimensions. This polysemanticity makes interpretability difficult: when a single neuron fires for "Paris," "the color blue," and "monetary transactions," no clean causal story exists.
The phenomenon arises from a simple capacity argument. A network with d dimensions can represent at most d orthogonal features perfectly — but the world contains far more features than any practical hidden dimension can accommodate. Networks resolve this by superimposing features at angles less than 90°, accepting interference as the cost of compression.
Why This Matters for Alignment
Polysemanticity is not merely an interpretability inconvenience. It means that interventions on individual neurons are unreliable — activating or suppressing a neuron affects multiple concepts simultaneously. Circuit-level analysis becomes a tangle of overlapping influences rather than a clean causal graph.
Scalable oversight requires the ability to inspect and verify internal reasoning chains. A polysemantic model resists this inspection at a fundamental level. Before we can trust that a model is reasoning honestly about a concept, we need to know which neurons are that concept — cleanly, with no crosstalk.
Our approach scales the FFN hidden dimension massively — e.g., 32,000 neurons vs. the standard 768 — while enforcing extreme sparsity: only 30–50 neurons activate per token. With enough neurons to represent features without overlap, superposition becomes unnecessary. The network no longer needs to pack multiple concepts into a single unit.
The routing mechanism is learned end-to-end. A lightweight gating network scores all hidden units for each token and selects the top-k by magnitude. Gradients flow only through active neurons, keeping compute proportional to the activation budget rather than the total hidden dimension.
Compute Profile
The expanded-but-sparse architecture sits on the Pareto frontier of interpretability and efficiency. Parameter count scales with the hidden dimension; compute scales with the activation budget per token. We can increase the former without touching the latter — a property unavailable in any dense architecture.
Preliminary benchmarks show FLOPs comparable to a dense baseline at 1/40th the hidden dimension, with the full 32k-unit model offering 10× more expressive capacity at equivalent inference cost.
Relation to SAEs
Research from Anthropic demonstrates that Sparse Autoencoders (SAEs) can extract monosemantic features post-hoc from a trained polysemantic model. Our approach differs in intent: rather than decomposing polysemanticity after it forms, we prevent it from forming at all. The routing constraint is part of training, not an analysis tool applied afterward. This means the monosemantic structure is native to the model — available for inspection, intervention, and verification without an external decomposition step.
