Jeva paper rethinks System 1 models: typed decisions over 2,048 candidates in a single pass

AndLukyane · x · 2026-10-29

Andrey Lukyanenko reviews the paper "Jeva: Rethinking System One Models from First Principles," on models that return typed decisions directly (tool choice, ticket routing, rubric scores) instead of generating rationales.

The paper identifies four shortcomings of existing open System 1 designs that stuff request and candidates into one prompt: null-awareness (distinguishing "no option applies" from "unsure"), permutation invariance, option scalability, and cacheability of recurring candidates.

Jeva's architecture: a shared Qwen3.5-4B backbone encodes the request and each option independently; a lightweight attention-based decision block refines only the request state (candidate vectors stay cacheable); a separate coverage gate outputs null probability. It scores up to 2,048 options in a single decision-block pass. On BANKING77, MASSIVE and MInd2Web it beats prior approaches (Jev, InterDec) on typed accuracy and null-detection AUROC, with zero flips under candidate permutations.

Original post →

More from Models

Models channel →