NUS Builds Jeva-4B to Pick From 2,048 Options Without Order Bias
A 4B decision-native model from NUS handles 2,048 candidates, zero decision flips under reordering, and 20x speedups via cross-request caching.
Takeaways
- NUS researchers released Jeva-4B, a decision-native System One model built on Qwen3.5-4B
- Zero decision flips under tested candidate permutations, versus 17-23% for comparable open-weight models
- Scales to 2,048 candidates in a single decision-head pass with 78.4% aggregate accuracy on ToolRet
- Hybrid caching of option representations delivers 27.86x median latency speedup on MLDR
- Unified Choice, Noul, and Score outputs with explicit null detection via a separate coverage gate
- Apache 2.0 weights and code at github.com/LiQiiiii/Jeva, paper at arXiv:2610.18773
Language models usually handle routing, tool selection, and scoring by serializing candidates into a prompt and generating an answer token. That shortcut imposes sequence semantics on an unordered set: reordering options can change the result, fixed answer tokens cap the set size, and the model lacks an independent signal for “none of the above.” Adding or replacing one candidate can also force the pipeline to encode the full prompt again.
Researchers at the National University of Singapore built Jeva-4B, a 4B-parameter model for fast, typed decisions over candidate sets. It targets workloads such as request routing, tool selection, response evaluation, Boolean verification, and document filtering. The architecture independently encodes requests and candidates, allowing candidate representations to be cached, recombined, and scored across many requests.
Why serialized choices drift
The paper uses “System 1” to describe fast decisions that do not require generated reasoning. It identifies four properties such models should provide:
- Explicit absence: distinguish an unsuitable candidate set from uncertainty among valid candidates.
- Order stability: preserve the decision when candidates are reordered.
- Variable cardinality: support changing set sizes without redefining the output vocabulary.
- Reusable computation: cache recurring candidate representations across requests.
Many decision models concatenate the request and every candidate into one sequence, then read probabilities from designated answer tokens. Candidate position can leak into predictions, the answer-token inventory limits set size, and every change to a tool registry or document set triggers another full-context encoding pass.
Stable options, adaptive requests
Jeva uses a shared Qwen3.5-4B encoder to process the request and each candidate independently. Pooling produces one vector for the request and one for every option. Candidate vectors see neither the request nor one another during encoding, so the same vectors can be reused with new requests or assembled into new candidate sets.