Jeva paper proposes a decision-native architecture with explicit null, order invariance and cached options
Qi Li and Xinchao Wang of the National University of Singapore posted “Jeva: System One Models from First Principles” on arXiv on 13 October 2026; Kalyan KS summarized the architecture on X. The paper argues that most open decision models put the request and the options into one text sequence and read the answer from token logits, which makes them sensitive to option order, caps the number of options and gives no clean way to say “none of these”.
The design. Jeva encodes the request and each option separately with a shared backbone. A small decision block then updates only the request state by reading the fixed option vectors. Because options are encoded on their own, their vectors can be cached and reused across requests. A separate coverage gate decides whether any option fits at all, which gives an explicit null answer. The same readout covers Choice, Noul and Score.
Results (authors’ runs, UnverifiedClaim). The released Jeva-4B is a Qwen3.5-4B LoRA plus a 13.4M-parameter decision head. Against Jev 1.13.0, JevK5, InternDecision, Kev-4B, Re·Flex, SemIf and others on the authors’ suite, it leads seven of eight accuracy slices. The exception is Typed Decisions, where it scores 78.30 against 78.70 for InternDecision-4B (Jev 1.13.0: 74.10). It shows zero decision flips under candidate reordering (Jev 1.13.0: 17%), 98.66% AUROC for null detection on BANKING77, and handles up to 2,048 candidates. Reusing cached options cut median latency on long-document sets from about 2.1 s to 75 ms without changing answers.
Why it matters. Order sensitivity and the lack of a real “none” answer are known weak spots of the category, and this is the first paper to attack them in the architecture rather than with prompt tricks or averaging over orders. The code is on GitHub under Apache-2.0. The weights got an Apache-2.0 licence on 18 Oct, and Jeva-4B now has a catalog page.
Licence update, 18 Oct 2026. > publicado (OK do Fabricio 18/10 BRT). At 07:52 BRT the Jeva-4B Hugging Face repo added a LICENSE file, a NOTICE file and license: apache-2.0 for the decision head and adapter; the bundled Qwen3.5-4B backbone keeps its own Apache-2.0 licence (1,246 downloads and 41 likes on 18 Oct). Our 17 Oct check had found no licence, so the catalog trigger (a stated weights licence plus a board result) was not met. The licence is now clear, but Jeva-4B is still not on the Decision Index, JevBench, ImageJevBench or the Artificial Analysis index, so the catalog page rests on the paper’s own numbers and says so.