Hugging Face has published a research paper titled "Jeva: System One Models from First Principles," introducing a decision-native architecture designed for direct, low-latency decision-making. The authors argue that predicting an answer from serialized text does not inherently respect the structure of a decision problem, and that current designs underemphasize four structural requirements.
The paper identifies these requirements as: null-aware (distinguishing answer absence from uncertainty among candidates), permutation-invariant (reordering alternatives should not change the decision), option-scalable (growing candidate sets should not require redefining the output space), and cacheable (request-independent computation over recurring alternatives should be reusable across requests).
Jeva addresses these jointly through a shared encoder that independently represents the request and each option, while a compact decision stack refines only the request state against immutable option memory. Symmetric aggregation yields permutation-equivariant probabilities, and shared scoring accommodates variable-size candidate sets. A separate coverage estimate distinguishes answer absence from candidate competition, enabling explicit null outcomes.
According to the paper, request-independent option representations and layer projections support reuse across newly assembled sets. These mechanisms unify Choice, Noul, and Score outputs. The authors report that across evaluations, Jeva combines strong typed-decision performance with near-perfect null detection and no decision flips under any tested permutations. It scales to thousands of candidates and substantially reduces latency through cross-request reuse without changing decisions.
The paper states that the 4B model supports Choice, Noul, and Score, handles 2,048 candidates, and shows zero flips under tested permutations. Full reuse of pre-cached long-document candidates delivers over 20× median-latency speedups, according to the authors. Code and model weights are available.