Grounded Dual-Stream World Models at Scale
The perception stream stays load-bearing.
Justin Hill, Independent Research. A follow-up to the July 2026 preprint Grounded Dual-Stream Language Models: it carries the founding perception ablation across three larger from-scratch runs and continues the lineage to 1.939B.
Held-out authored-token accuracy at 4.79M parameters, with the perception stream zeroed at inference. Bits per byte splits the same way: 1.5239 with the read intact against 2.7513 with it severed. Remove the read, and the author loses more than half of its held-out accuracy.
The training pipeline
Four control layers sit between the corpus and the optimizer, and they all read one signal, the canary. Each turn the mixture is re-cut toward whatever the world model is thinnest on: grounded reads where perception is starved, plain text where it is already fluent.
The lineage
If grounding were decoration, the gap should shrink with scale, since a larger model could carry the answers in its weights and stop consulting the read. It stayed between +0.44 and +0.60 and never fell back toward chance.
| Run | Params | Dates | Ablation Δ | Standing |
|---|---|---|---|---|
| Preprint (v1) | 4.79M | Jul 2026 | +0.575 | founding |
| 526M CPT | 526M | Jul 17–18 | +0.444 | replicated · 800k steps, zero NaN |
| 124M pair | 124M | Jul 19 | +0.603 | replicated after 29GB injection |
| 1.939B run | 1.94B | Jul 20–21 | n/a | diagnostic-only · uncontrolled, not failed |
The 1.939B run trains cleanly at scale (bits per byte 3.517 to 1.114, zero NaN) and carries no capability claim: its curriculum controller steered on a drive that was roughly 78% phantom need, and its two grounding readouts were built but never wired into the loop. The correct response is a controlled re-run under a frozen configuration and a versioned mixture manifest.
