Paper review · October 15, 2026
Jeva: A System One Model Designed from First Principles of Decision Structure
Jeva: System One Models from First Principles
The authors designed a decision-specific architecture that encodes the request and the candidate options separately and updates only the request state against the candidate set. This gives order-invariant selection, an explicit 'none applies' verdict and caching of candidate representations at the same time.
Problem
System One models, which make short decisions directly, such as tool selection, document identification, condition checks and scoring, usually concatenate the request and the candidates into one text and encode it. This approach has four problems. The decision can change when the candidate order changes. The number of answer symbols is fixed, so it is hard to increase the number of candidates. It cannot distinguish 'no candidate fits' from 'candidates are competing with each other'. Computation for candidates that appear repeatedly has to be redone for every request.
Method
A shared language encoder (Qwen3.5-4B) vectorizes the request and each candidate independently. A lightweight decision block then reads the fixed candidate memory through attention and updates only the request state. Because it uses symmetric aggregation and a shared scoring function, reordering the candidates only reorders the probabilities in the same way, and the output parameters do not grow as the number of candidates grows. Separately from the competition distribution across candidates (softmax), a coverage gate estimates from the top score the 'probability that at least one valid candidate exists'. This gate produces a null result explicitly. Candidate vectors and per-layer key/value projections do not depend on the request, so they can be cached across requests. Training used rank-64 LoRA and a decision head (13.37M parameters).
Results
- 91.62% accuracy on BANKING77 (K=16), higher than Jev-1.13.0 (82.50%), the next-best of the compared models. However, on Typed Test it scored 78.30%, slightly below InternDecision-4B (78.70%)
- In the choice-reordering experiment, the share of decisions that changed (Flip) was 0.00%. The compared models ranged from 17 to 85%
- Noul Brier score of 0.0184, lower than InternDecision-4B (0.0340) (lower is better)
- The authors report higher accuracy than a recursive selection baseline with 2,048 candidates, and median latency more than 2x faster
- On newly assembled long-document candidate sets, they report that cache reuse across requests cut median latency by more than 20x with no change in decisions
Limitations
The main experiments are limited to a Qwen3.5-4B backbone with a single decision block. On some benchmarks (Typed Test) it is slightly below existing models. The encoder length limit for each text still applies. Order invariance is a property that assumes exact arithmetic, so tie handling and finite-precision effects need to be treated separately. The provided text was cut off partway, so the training data composition and detailed experimental conditions could not be checked.
Why it matters
For services that handle large volumes of short decisions such as tool routing, document selection and classification, order sensitivity, 'none applies' handling and latency are real operational problems. This paper proposes a design that addresses all three at the architecture level.
Key terms
- System One model
- A low-latency model that directly outputs a structured decision, such as option probabilities or scores, without generating its reasoning as text.
- Permutation equivariance
- The property that reordering the candidates only reorders the output probabilities in the same way, leaving the values unchanged. As a result, the same candidate is chosen whatever the input order.
- Coverage gate
- A component that, separately from the relative competition among candidates, estimates the probability that 'at least one suitable candidate exists'. It allows the model to explicitly output 'no answer (null)'.
- Option memory caching
- Because candidate representations are computed independently of the request, when the same candidate appears again in another request or another combination, the stored computation is reused as is.
Original
Authors: Qi Li, Xinchao Wang · Posted 2026-10-13 · License CC BY 4.0 · Based on the full text
This review is an AI-written summary of the paper. Please treat the original paper as the reference.