Qurated: the polysemanticity of polysemanticity in language models
The Polysemanticity of Polysemanticity in Language Models
One neuron can mean many things — and that's not a bug, it's the price of intelligence under scarcity.
To understand how neural networks think, you have to understand how they cheat. When a model has fewer neurons than concepts it needs to represent, it doesn't fail. It compresses. It teaches a single neuron to fire for multiple, unrelated inputs — Python syntax and Victorian prose. This is polysemanticity, and it sits at the heart of mechanistic interpretability.
The Core Idea
Polysemanticity arises from superposition: the representation of multiple states in the same space.
In an ideal world, each neuron maps cleanly to one concept — a "monosemantic" model we could read like a dictionary. But real networks don't have that luxury. There are always more features worth tracking than there are neurons to track them.
So the model makes a trade: it overloads neurons, packing many features into fewer dimensions. The same neuron activates on distinct, unrelated inputs. Interpretability gets harder — but capability gets cheaper.
The Example That Makes It Click
Imagine 25 Python snippets and 25 classic-novel snippets. Completely distinct domains — they almost never co-occur.
If you only have a handful of neurons, you can't give each concept its own home. But here's the insight: because Python code and Dickens rarely appear together, you can safely assign them to the same neuron. When that neuron fires, context resolves the ambiguity. Code-context means "Python." Prose-context means "novel."
The model exploits sparsity — the fact that most features are inactive most of the time — to store more concepts than it has neurons. This is superposition in action.
A Mental Model: The Overbooked Hotel
Think of a hotel with 10 rooms but 40 guests. Reckless — unless the guests rarely visit at the same time.
- Rooms = neurons. A fixed, scarce resource.
- Guests = features. Concepts the model wants to represent.
- Non-overlapping schedules = sparsity. The reason overbooking works.
The hotel stays functional as long as no two guests show up simultaneously. Collisions happen — that's interference, the noise polysemanticity introduces — but if features are sparse enough, the benefit of representing more concepts outweighs the occasional clash.
Why This Matters
1. Interpretability is hard by design. You can't just look at a neuron and read its meaning. A neuron is a smeared-together chord, not a single note. Tools like sparse autoencoders exist precisely to "unfold" superposition back into cleaner, monosemantic features.
2. Compression is intelligence. Polysemanticity isn't a defect to eliminate — it's evidence the model has learned to allocate scarce representational capacity efficiently. Understanding it tells us how networks generalize.
3. The analogy runs deep. Superposition in neural networks rhymes with superposition in quantum systems — multiple states coexisting in one substrate until context (or measurement) collapses them into meaning. Both are lessons in doing more with less.
Your Takeaway Framework
When you next probe a model — or design one — ask three questions:
- How scarce are the neurons relative to the features? Scarcity forces superposition.
- How sparse are the features? Sparsity is what makes superposition safe.
- Where does context resolve ambiguity? That's where the model recovers meaning from overloaded representations.
The lesson generalizes beyond machines: the most powerful systems aren't those with infinite capacity, but those that learn to share space intelligently. Constraint, handled well, becomes the engine of understanding.