Skip to the content.

Scaling synthesis — what tracks scale, not just architecture

The single clearest cross-cutting finding: several properties usually attributed to architecture (absolute-position vs RoPE) actually track scale. This table lines up the scale-varying quantities from across the catalog (assembled by scaling_synthesis.py from the committed result JSONs — no model run). Read the GPT-2 ladder (124M → 355M → 774M) top-to-bottom.

model params induction key-collapse induction redundancy recon. coverage (mean / resample) MLP0 token-determinism succession MLP depth knowledge trace peak depth
gpt2 124M +39% distributed +17% / +31% 0.63 0.09 0.00
gpt2-medium 355M +8% distributed +7% / +24% 0.61 0.00 0.00
gpt2-large 774M +1% compensatory +0% / +5% 0.75 0.20 0.03
gpt2-xl 1.5B +0% distributed +1% / +7% 0.80
gemma-2-2b 2.6B +3% compensatory +14% / +7% 0.91
Llama-3.2-1B 1.2B +2% distributed +10% / +10% 0.01 0.00
Qwen2.5-1.5B 1.5B +0% distributed -4% / +0% 0.65 0.00

What the columns show

The thesis: as models scale, the same named circuits become more distributed — single dominant writers give way to populations, compact circuits stop being sufficient, and the load-bearing MLP sites broaden and deepen. Absolute-vs-RoPE is a real axis (the sink, positional broadcast), but much of what looks architectural is the small models being unusually localized. See Cross-model findings.

The controlled ladder — Pythia (architecture held fixed)

The table above mixes the GPT-2 ladder with heterogeneous RoPE models, so architecture and scale are confounded. The Pythia ladder (one GPT-NeoX architecture, the same training data, 14M→1.4B) is the clean control (scaling_laws.py, arch-generic block-level + logit-lens — no head resolution needed). Three quantities turn into monotone laws with architecture fixed:

pythia d×L induction-NLL all-block-ablated Δ capital table capital read-out depth language read-out depth
pythia-14m 128×6 2.09 +8.3 58% 91% 89%
pythia-70m 512×6 2.17 +7.1 83% 78% 82%
pythia-160m 768×12 0.99 +9.2 100% 68% 67%
pythia-410m 1024×24 0.54 +11.0 100% 57% 53%
pythia-1b 2048×16 0.45 +10.9 100% 61% 57%
pythia-1.4b 2048×24 0.48 +11.3 100% 52% 69%

This is the thesis on a clean axis: with architecture fixed, induction appears and sharpens, the fact table fills, and retrieval moves earlier — all monotone in size. Scale, not architecture.

Assembled from the committed runs/disassembly/** summaries. Regenerate: scaling_synthesis.py.