vega-U: a 14.3M decision engine that cold-attaches to any frozen LLM, 0.811 held-out accuracy

kalyan_kpl · x · 2026-10-18

embedding.in released vega-U, a universal 14.3M-parameter (57MB) decision engine living in a shared 1024-dim feature space that any frozen LLM can plug into via a small per-layer linear 'port' — no longer requiring a decision head married to one backbone.

Key design choices:

The shared space is anchored to one reference reader's standardized feature space, with that reader's port frozen as identity, pinning the coordinate system so nothing drifts.
Each reader's features are standardized with its own stats before projection, preventing ports from learning a 'disguise' mapping.

Results: readers it never trained with score highest — a cold-attached Qwen3.8-27B reaches 0.811 held-out accuracy (ECE 0.021, 0.976 accuracy at 50% coverage), including 0.941 on long context, 0.922 on ARC, and 0.988 on jailbreak detection. Cold attachment needs only a closed-form port fit, a short refine, and calibration, run on a single RTX 6000 Pro.

Original post →

vega-U: a 14.3M decision engine that cold-attaches to any frozen LLM, 0.811 held-out accuracy

More from Models