NUS Unveils Jeva: Unified Architecture Solves Four System One Decision Challenges

Researchers from the National University of Singapore (NUS), Qi Li and Xinchao Wang, introduce Jeva ( Jeva: System One Models from First Principles ), a novel architecture designed to address four fundamental challenges in System One decision-making models: determining whether a valid answer exists ( null-aware ), ensuring decisions remain consistent regardless of candidate order ( permutation-invariant ), scaling to large candidate sets without interface changes ( option-scalable ), and reusing computation when candidate materials reappear across requests ( cacheable ).

Architecture Overview

Jeva employs a shared language encoder to independently represent the current request and each candidate option. A lightweight decision module then reads the complete option memory while updating only the request state. The final judgment is produced jointly by the updated request state and the candidate representations. The evaluated Jeva-4B model uses a Qwen3.5-4B backbone with a single decision block.

Four distinct mechanisms support the four capabilities within one unified decision flow:

Coverage estimation judges whether any acceptable answer exists.

Symmetric aggregation reads options without treating permutation as a signal.

Shared scoring rules adapt to varying candidate counts.

Independent option representations enable cross-request reuse.

Null-Aware: Explicitly Detecting Absent Answers

Standard models forced to choose among provided candidates often pick the "most similar" option even when none is appropriate. Jeva separates candidate competition (relative preference among options) from coverage estimation (whether the set contains a valid answer). The output includes both candidates and a null token, making answer absence a first-class result.

On the BANKING77 dataset (3,080 requests, 16 candidates each, 588 answer-absent cases), Jeva achieves a null detection AUROC of 98.66% , surpassing KEV-BF16 (93.61%) , InternDecision-4B (84.27%) , and JevK5 (79.44%) . At a ~5% false-negative rate on answer-present requests, Jeva identifies ~94% of answer-absent requests.

Permutation-Invariant: Ignoring Order, Preserving Content Sensitivity

Jeva uses symmetric aggregation to read the option set, then applies shared scoring rules. Mathematically, reordering candidates only reorders output probabilities; aligned by candidate identity, the probability distribution remains unchanged. Crucially, the request state reads the entire collection, so changing a third candidate can alter the relative preference between the first two — content matters, position does not.

On 600 Choice samples tested under original, reversed, and up to three random permutations, Jeva exhibits zero decision flips (original accuracy and all-sequences-correct accuracy both 76.67% ). In contrast, Jev flips 17.00%, InternDecision-4B 22.67%, and JevK5 26.17%. A constructive proof with a single decision block shows that independent per-candidate scoring cannot express such context-dependent preference reversals.

Option-Scalable: Same Scoring Rules, Growing Candidate Sets

Jeva applies the same scoring rule to every candidate, so adding candidates does not require expanding output parameters. Text can be encoded in batches; once prepared, the full candidate set enters the decision module together. With fixed decision width and depth, post-encoding computation scales linearly with candidate count.

In six ToolRet scale experiments, Jeva supports all 500 requests across all scales. On a single H200 with 2,048 candidates , Jeva achieves 69.88% accuracy at 12.04s median latency . Recursive JevK5 reaches 67.47% accuracy at 28.35s — Jeva improves accuracy by 2.41 percentage points and speeds up inference by ~2.35× . A further controlled test fixing 83 questions and growing candidates from 32 to 2,048 shows Jeva retains correct answers on 58 questions across all scales, versus 48 for JevK5 and 7 for CLM.

Cacheable: Reusing Candidate-Side Computation

Option representations in Jeva are request-independent, allowing caching of semantic vectors and the projections used by the decision module. New requests still encode the query, update state, and recompute outputs, but the candidate-side work is reused. Reuse is granular: if a previous candidate set was {A, B, C} and the new set is {B, C, D}, B and C are fetched from cache; only D is newly encoded.

On the MLDR long-document benchmark (single H200), when all candidates in a new combination are cached, request median latency drops from ~2.10s to 75.1ms — a 27.98× speedup (cache warm-up timed separately). No decision changes were observed under paired cache conditions.

Unified Decision Interface

Jeva exposes three output modes from the same architecture: Choice (selection + candidate probabilities), Noul (proposition truth probability, with null meaning no suitable answer), and Score (expected value of a numeric rating scale). A single service fault report, for example, can route to a handling team (Choice), estimate outage likelihood (Noul), and assess severity level (Score) without retraining.

The paper concludes that effective direct decision-making requires not only how answers are emitted, but also how the model handles absent answers, irrelevant order, variable candidate counts, and repeated information — all four requirements integrated into one computational framework.