Lab notes: lesioning the model

A second follow-up, July 2026 — same rules as the paper: every number recomputable from persisted artifacts, every claim next to the control that could kill it. The full narrative, in Spanish, is on the research blog; the full technical detail is in the findings notebook, Part 3.

Neuropsychology learned the brain, in large part, by breaking it: a stroke takes one region and a patient can no longer name objects though speech is fluent, or speaks in telegram though comprehension is intact. Each selective deficit maps what a piece does. A language model can be lesioned with a precision no surgeon ever had — silence one exact neuron, at every position, and measure what breaks. So we did.

The question comes from the paper: the thought splits into the thing (Italy) and the operation (capital-of). If those are real pieces, they should live in different tissue. Is there a set of neurons that does "capital-of", removable separately from the ones that handle "Italy"? And if you lesion it, does the model suffer a specific aphasia?

The method (borrowed, not the fMRI data — every claim is about the model): the paper's balanced 12×5 grid is an fMRI localizer — decompose each neuron's activity into how much it tracks the operation vs the entity, and rank. Then remove the tissue (silence it everywhere, not steer it) and demand the deficit be selective: the targeted lesion breaks its function and spares the rest, while controls of the same size, in the same layers, at the same activation magnitude, do not.

What we found

A tiny brainstem. Ablating each of Qwen3-1.7B's 448 attention heads alone: 2 are load-bearing (one, L1H5, raises perplexity ×19.7 by itself), the other 446 cost almost nothing to lose. And it is a small-model phenomenon — at Qwen3-8B, zero of 1152 heads are critical. The small model keeps a head it cannot survive; the large one keeps none.

The method is validated. Localize and lesion the 48 induction heads (the field's most universal circuit) and in-context copying collapses −79% while perplexity, arithmetic and factual retrieval are spared.

The operation has an organ; the entity does not. Lesioning operator-selective neurons makes the model answer the right country under the wrong relation — at every lesion size, dose- ordered, in both models, with perplexity spared (Fisher-exact p down to 5.5e-09 against a pooled control null). But the operand network has no removable tissue: the predicted "says another country's capital" essentially never happens — not at the query position, and not at the entity token (the obvious mechanistic guess, tested and refuted). A factorization symmetric in the algebra is not symmetric in the tissue. It is a paradigm of operators — and the operand's mechanism stayed open until the follow-up below.

(An honesty note I like to tell: reading the raw generations under the operand lesion, I thought I saw a beautiful anomia — "it's a city… I can't say which". The rigorous scoring showed "degraded" is the sink every failure falls into (~37% even for controls), so it can't be a signature. I was reading signal in the pit. The result survives without the ornament.)

Areas or networks? It depends on the function. The operator and operand networks are spatially concentrated, stacked at 84–93% depth — the paper's causal lever, found again by a method that knows nothing about the paper. The text-vs-symbolic distinction is fully distributed (indistinguishable from a random-unit null). Localized areas and diffuse networks, side by side — like a brain.

Why does the small model need a head the large one does not?

Two ways the large model could "not need" that head: the function is distributed across a set (no single head critical, but a group is), or it simply isn't needed. A greedy search — remove the head whose ablation costs the most, repeat — tells them apart.

At 1.7B the damage compounds catastrophically: L1H5 alone ×19.7, then ×270, ×23,000, up to ×2.8 million at 8 heads — a tiny but interdependent critical core. At 8B the same procedure over 12 heads reaches only ×4.3. No catastrophic core: not one head, not a small set. Six orders of magnitude apart. The answer is the second — the large model concentrates its load-bearing computation in no small set of heads. The small model stacks work it cannot lose into a handful; the large one spreads it so thoroughly no small group matters.

And what does L1H5 do? The obvious guess — an attention sink, a head dumping almost all its attention on the first token (a documented load-bearing pattern) — is wrong. Sinks exist: L19H4 sends 97.4% of its attention to token 0. But the critical heads are not sinks: L1H5 sends only 1.3% (rank 430 of 448). It is load-bearing for some other reason, still unidentified — an honest loose end.

The operand is in the wiring (update)

Why did lesioning neurons never find the operand? Because a lesion removes stored computation, and the entity is not stored at the query position — it is moved there, by attention, from the entity token. The operand may have no removable tissue for the simple reason that it is a wire, not a store. And you cut a wire with an attention knockout (Geva et al., 2304.14767): block the query→entity attention edge and see what breaks.

It breaks exactly as predicted. Cutting query→entity attention in a mid-late band (~76% depth at 1.7B, ~70% at 8B) destroys entity retrieval — dose-dependently in how many layers you cut (1.7B accuracy 63% → 36% → 15%) — while the model keeps the relation: "the city of Rome" for Egypt's capital, "a city that is known for…" for France's. Meanwhile cutting attention to the operator word ("capital", "currency") is inert (61–80% accuracy, barely moved): the operator is not read from its word, it is constructed. entity-vs-operator: Δaccuracy −25% to −48%, McNemar p down to 1.3e-07, replicated at 8B.

So the two halves have two substrates: the operator is stored MLP tissue (removable by lesion), the operand is attention routing (severed by cutting the wire). The symmetry of the algebra does not imply symmetry of the implementation — one factor is a store, the other is a wire. Full detail: findings §3.9 and the research blog.

Fine print

• One and a half models (Qwen3-1.7B, replicated on 8B), 12 countries, 5 relations, English. The operator result and the geometry replicate across scale; the head arm is unreadable at 1.7B (the sink) and readable-but-silent at 8B (the operation is not a small head set).
• Methodological inspiration, not a literal brain claim — no fMRI comparison, no Brodmann area.
• Reproduce it: scripts/lesion_*.py with committed summary JSONs, checked by verify_numbers.py.

Background first? The paper, explained simply · the operator-vector note · what's next.