Daily Paper Briefing

Abstract Overview

This paper introduces Jeva, a decision-native architecture for low-latency "system one" decision tasks that is designed around four structural requirements: null awareness, permutation invariance, variable-size candidate sets, and cacheable option computation. Jeva independently encodes the request and each candidate option with a shared backbone, then refines only the request state against immutable option memory using a compact decision stack. The paper argues theoretically that this design can produce set-dependent preference reversals while preserving candidate symmetry and exact reuse of cached option-side computation. Experiments evaluate typed decision quality, null detection, robustness to candidate reordering, scaling to large candidate sets, and latency benefits from caching.

Novelty

The distinctive contribution is an architecture that jointly targets null detection, permutation-equivariant decision probabilities, support for variable candidate counts, and cross-request candidate reuse within one decision framework. Unlike joint-encoding or fixed-label approaches, Jeva separates reusable option encoding from request-specific state refinement while still allowing the candidate set to influence the final choice.

Results

Across the reported evaluations, Jeva leads 7 of 8 full-coverage accuracy slices in Table 1, exceeds 90% accuracy on BANKING77 and MASSIVE, and attains 79.99%, 62.25%, and 66.23% on three Mind2Web selection slices. It reports near-perfect order stability with 0.00% decision flips under tested permutations and strong null detection, including 98.66% AUROC on a BANKING77 null-detection setup. The model also scales to 2,048 candidates, where it reaches 69.88% accuracy and shows latency reductions from batching and cross-request reuse, including over 20× median-latency speedups in the reuse setting described by the paper.

Key Points

1. Jeva uses independent request/option encoding plus request-only state refinement over immutable option memory, enabling reusable candidate representations without removing set-dependent interactions.
2. The architecture explicitly separates candidate competition from coverage estimation, which supports null outcomes rather than treating null as just another candidate in a flat normalization.
3. Empirically, the paper highlights strong typed decision performance, zero tested permutation-induced decision flips, effective scaling to thousands of candidates, and substantial latency gains from cache-based reuse.

References

• arXiv: https://arxiv.org/abs/2610.18773v1
• Fugu-MT: https://fugumt.com/fugumt/paper_check/2610.18773v1
• Hugging Face Papers: https://huggingface.co/papers/2610.18773
• GitHub: https://github.com/LiQiiiii/Jeva
• Hugging Face: https://huggingface.co/LIQIIIII/Jeva-4B
• Project: https://liqiiiii.github.io/Jeva/