Back to AI Research

AI Research

Jeva builds explicit absence and reusable options into decision models

Key Takeaways

  • A decision architecture aims to preserve choices across candidate reordering while reusing representations between requests.
  • ## A decision has more structure than an answer string
  • [Jeva](https://arxiv.org/abs/2610.18773) studies models that return direct, low-latency decisions.
  • These requirements affect ordinary routing and selection problems.
  • Uncertainty between two plausible options is different from a case in which neither option is suitable.

A decision has more structure than an answer string

Jeva studies models that return direct, low-latency decisions. The authors argue that serializing a decision problem as text leaves several important requirements insufficiently addressed: detecting when no answer applies, preserving choices when options are reordered, handling different candidate counts and reusing computation for recurring alternatives.
These requirements affect ordinary routing and selection problems. Uncertainty between two plausible options is different from a case in which neither option is suitable. Likewise, changing the order of an otherwise identical candidate set should not change what the request means.

Represent requests and options independently

Jeva uses a shared encoder to represent the request and each option independently. A compact decision stack refines the request against immutable option memory. Symmetric aggregation supports probabilities that remain aligned with the options when their order changes.
Shared scoring accommodates variable-sized candidate sets, while a separate coverage estimate supports an explicit null outcome. Request-independent option representations and layer projections can be reused across newly assembled sets. The architecture supports Choice, Noul and Score outputs.
The paper reports strong typed-decision results, near-perfect null detection and no decision flips under its tested permutations. It also reports scaling to thousands of candidates and reduced latency through cross-request reuse. These are experimental findings for the evaluated tasks, rather than guarantees for arbitrary candidate descriptions.

Evaluate absence and ordering deliberately

Franklin analysis: a routing test should include requests for which none of the available destinations is appropriate. Shuffle the same options and check both the selected destination and its probability, then measure warm-cache and uncached latency separately.
This prevents a high overall accuracy score from hiding a model that always chooses something or changes its answer with option order. Jeva makes those structural properties part of the architecture and provides a concrete basis for testing them.

Comments