NUS Open-Sources Jeva: A Decision-Native Model That Never Flips on Candidate Reordering

Researchers at the National University of Singapore have released Jeva, an open-weight decision-native architecture that achieves zero decision flips under candidate reordering — a structural failure mode affecting 17 to 26 percent of queries in competing systems. The paper, arXiv:2610.18773, submitted October 13, 2026, reached number one on HuggingFace Papers the following day with more than 273 upvotes.

The release arrives one month after TypeSafe AI launched its commercial Jev model — a closed-weight System One service backed by $40 million in seed funding. Where Jev entered the market as a managed API, the NUS paper gives the research community the first peer-reviewed architectural treatment of what a decision model must structurally guarantee, with open weights for engineers to inspect and deploy.

What System One Decision Models Are and Why Ordering Bias Matters

A System One model — named after Kahneman's fast, automatic cognition — returns a typed decision (a candidate selection, a calibrated score, or an abstention) in a single forward pass rather than generating free-form text. Standard LLMs can be pressed into classification duty by placing candidate labels in the prompt, but positional encoding baked into the backbone's pretraining distribution leaks into candidate representations, meaning the position a candidate occupies can shift its probability. The paper measures this as flip rate — the fraction of test queries where reordering candidates changes the model's top-1 answer. Across models evaluated, flip rates run from 17 percent for InternDecision-4B to 26 percent for ReFlex-4B. Jeva's measured flip rate is zero.

The Architecture That Enforces the Guarantee

Jeva's design routes the request and each candidate option through a Qwen3.5-4B backbone independently, producing a pooled representation for each. Those vectors enter a compact decision stack: a single transformer block with eight attention heads, model dimension 768, and feed-forward width 3,072.

The structural constraint is in how that block operates. The request state attends to the option vectors, but option vectors do not attend back to the query. Option representations are immutable inside the decision stack. Because no query-dependent information can flow into the option side, the model cannot learn position-dependent patterns for candidates — removing the pathway through which ordering bias enters. Permutation invariance is enforced downstream by a symmetric aggregation function over candidate scores.

A separate sigmoid gate estimates whether any valid answer exists within the candidate set, using the maximum candidate score as input — independent of the softmax over candidates. This separation of "does a valid answer exist" from "which answer is best" is what the paper calls null awareness. It allows explicit abstention (a Noul output) without forcing probability mass onto the least-wrong option, a capability absent or only partially implemented in all competing models examined.

Training uses rank-64 LoRA adapters on the frozen Qwen3.5-4B backbone with a staged curriculum, combining BCEWithLogits loss for the coverage gate and cross-entropy for candidate selection. Adapters are merged at serving, eliminating inference overhead.

Benchmarks and the Caching Advantage

Jeva leads seven of eight full-coverage accuracy slices across BANKING77 (banking intent classification), MASSIVE (multilingual NLU, 18 languages), Mind2Web (web-agent action selection from HTML candidates), and typed decision routing tasks. All figures are lab-reported; no independent reproduction has been confirmed.

The economic case is driven by caching. Because option representations are query-independent and immutable, they can be pre-computed once and reused for every subsequent query against the same candidate set. The paper reports greater than two-times median latency reduction for within-request scoring and greater than twenty times for cross-request reuse. TypeSafe's Jev prices API access at $0.042 per million input tokens; at high throughput, Jeva's caching multiplier translates directly into cost reduction for equivalent throughput.

The Gap Jeva Opens in the Decision Model Ecosystem

The broader System One ecosystem arrived quickly: Kev (Jared Palmer, Apache-2.0, 0.8B through 9B), InternDecision, JevK5, and ReFlex-4B all preceded Jeva as open-weight entrants. What those models lacked was a formal derivation of the structural requirements any decision model must satisfy. Jeva's paper establishes four: null awareness, permutation invariance, option scalability to 2,048 or more candidates, and cross-request cacheability. A model that fails any of these is, in the paper's framing, structurally underspecified for the decision problem — regardless of its benchmark accuracy.

For teams building agent pipelines and routing systems, the result is practically significant: LLM-logit-readout classification carries a flip-rate risk that cannot be trained away, only architecturally removed. Jeva's independent encoder design is compatible with any backbone producing pooled embeddings, so the architecture is not permanently tied to Qwen3.5-4B.

Xinchao Wang's NUS group — known for MetaFormer, LLM-Pruner, and DeepCache — has produced a baseline that commercial and open-source decision model developers will now need to address. TypeSafe's $40 million seed implicitly assumed the absence of a credible open-weight alternative with formal guarantees. Jeva does not replicate Jev's production infrastructure, but it makes the architectural core of the problem open to scrutiny and extension.

Read more: vLLM-SR Decision 3.0 multimodal suite tops Jev

The next milestone is independent reproduction of Jeva's zero flip rate on standard benchmark splits. If confirmed, it would formally establish ordering bias as a structural disqualifier for logit-readout classifiers — with implications well beyond the current System One niche.