Skip to the content.

Outlier mechanism digs

Targeted follow-ups on the two recurring outliers the cross-model dossier surfaced. Provisional.

Why is Llama-3.2-1B’s MLP0 context-determined?

MLP0’s input is the token embedding plus the layer-0 attention write. If the early attention injects a large, context-determined component, MLP0 processes a context-mixed residual and its output is no longer token-determined. Per model: the L0-attention write size relative to the embedding, and the token-determinism (η²) of the embedding, the L0 attention output, and MLP0’s output.

model ‖L0 attn‖ / ‖embedding‖ η² embedding η² L0 attn-out η² MLP0-out
gpt2 6.60 0.61 0.64 0.63
gpt2-large 3.14 0.66 0.72 0.74
gemma-2-2b 0.18 1.00 0.69 0.91
Qwen2.5-1.5B 11.88 1.00 0.84 0.64
Llama-3.2-1B 0.90 1.00 0.49 -0.14

Reading: it is the determinism of the L0 attention output, not its size, that distinguishes Llama. Qwen’s L0 attention is far *larger* (≈12× the embedding) yet token-determined (η² 0.84), so its MLP0 stays token-ish; Gemma’s L0 attention is tiny (0.18×) so MLP0 ≈ the pure embedding (0.91). Llama’s L0 attention is both comparable in size to the embedding and the most context-determined (η² 0.49) — its layer-0 induction-mass head cluster does genuine context-mixing — so MLP0 ingests a large context-laden component and its output carries ~no token-determinism (η²≈0; small negatives are estimation noise). The context-dependence is inherited from the early heads, not intrinsic to MLP0. (GPT-2’s embedding η²<1 is its absolute positional embedding adding position variance to the residual; the RoPE models read 1.00 — no positional component in the residual.)

Compensatory induction: which head triggers the recovery?

For each model’s top-k induction heads (by induction-mass), the leave-one-out marginal of head h = effect(ablate-all) − effect(ablate-all-except-h). A negative marginal means ablating h reduces induction damage — a net suppressor / self-repair trigger; a distributed op has all-positive marginals.

model induction top-k full ΔNLL most-negative LOO marginal (head) distributed?
gpt2 5.1, 6.9, 5.5, 7.10, 7.2 +5.27 +0.44 (none<0) yes (all marginals ≥0)
gpt2-large 16.0, 19.4, 16.9, 18.7, 20.1 +0.84 -0.66 (16.0) no — has a suppressor
gemma-2-2b 6.3, 6.2, 22.3, 4.4, 22.4 +3.77 -1.10 (4.4) no — has a suppressor

Per-head solo + LOO marginals are in the JSON. The suppressor head is the one whose removal lets a backup carry induction (the non-monotonic cumulative curve in the operator catalog). Caveat (see the next section): the apparent suppression is largely a *synthetic repeated-random probe artifact* — these heads have positive OV and are ~neutral on natural-text induction.

Is the suppressor a genuine negative head, or a synthetic-probe artifact?

For each identified suppressor (+ the workhorse for contrast): the OV copy-score sign (+ve copies the attended token → a real copy/induction head; −ve suppresses it → a copy-suppression / negative head), and the head’s ablation effect on induction measured over natural-repeated text (a real passage + itself) vs synthetic-repeated (random tokens + itself). A probe artifact would help natural induction (ablation ΔNLL > 0) but hurt synthetic (< 0); a genuine suppressor is −ve OV and helps both (ablation ΔNLL < 0).

model head role OV copy-score ablate ΔNLL natural ablate ΔNLL synthetic
gemma-2-2b 4.4 suppressor +0.036 -0.04 -0.60
gemma-2-2b 22.4 workhorse (contrast) +0.068 +0.07 +2.95
gpt2-large 16.0 suppressor +0.087 +0.11 -0.01
gpt2-large 16.9 workhorse (contrast) +0.039 +0.33 +0.70

+ve ablation ΔNLL = the head HELPS induction (removing it hurts); −ve = the head SUPPRESSES it (removing it helps). Finding: both suppressors have positive OV copy-scores — they are copy/induction heads, not copy-suppression / negative heads. The suppression shows up only on the synthetic repeated-random probe (Gemma 4.4: ΔNLL synthetic −0.60 but natural ≈0; gpt2-large 16.0 likewise marginal/positive on natural). So the *compensatory* redundancy is substantially a repeated-random probe artifact — these heads interfere with the degenerate synthetic-induction task but are ~neutral on real-text induction — not a genuine negative-head self-repair mechanism.

Do Llama’s layer-0 heads do single-layer (RoPE-enabled) induction?

GPT-2 needs a two-layer chain (a prev-token head feeds an induction head’s key). Llama has induction-load-bearing heads at layer 0 — where there is no prior layer to supply a prev-token signal. RoPE puts relative position in the key, so a single head can match token-after-previous-occurrence directly. If these heads carry induction-mass ≫ duplicate-mass and a +ve OV copy-score, they are single-layer inductors (no upstream writer needed).

model head induction-mass duplicate-mass OV copy-score single-layer inductor?
Llama-3.2-1B 0.31 0.034 0.021 +0.023 no — enabler, not inductor
Llama-3.2-1B 0.29 0.025 0.014 -0.060 no — enabler, not inductor
Llama-3.2-1B 0.13 0.025 0.028 -0.005 no — enabler, not inductor
Llama-3.2-1B 0.14 0.030 0.045 -0.009 no — enabler, not inductor
Llama-3.2-1B 1.31 0.033 0.033 -0.077 no — enabler, not inductor
Llama-3.2-1B 1.29 0.026 0.025 +0.017 no — enabler, not inductor

Finding (hypothesis not supported): these layer-0 heads do not behave as single-layer inductors — their induction-mass is weak (~0.03) and ≈ their duplicate-mass, even though 0.31 is strongly induction-*causal* (+7.99 when ablated, per the discovered candidates). So they are induction enablers, not inductors: they don’t attend induction-style themselves, but their early context-mixing (Dig 1 — Llama’s L0 attention is the most context-determined) sets up the residual that later heads read. Llama’s actual induction *reader* is a later head (10.23 in the dossier). A clean reminder that high causal effect ≠ doing the named operation.

Data: outlier_digs_summary.json. Regenerate: outlier_dig.py.