Give a model four bad choices and it may still confidently choose the least bad. That confidence says little about whether any choice belongs on the list.
Jeva, described in an October 13 paper, treats this as an architectural problem. It independently encodes requests and options, then uses a separate coverage estimate to support an explicit null outcome. Its authors report no decision flips under the option permutations they tested, alongside reuse of option representations across requests. Read the research paper.
The option list is part of the uncertainty
The distinction matters for an agent selecting tools. A fifth tool could resolve a task the original four cannot. A conventional confidence score over those four does not express that missing capability; it redistributes probability inside an inadequate menu.
A useful deployment test would therefore remove the correct tool deliberately, then expand the menu with distractors. Does abstention survive both interventions? This tests coverage rather than merely ranking.
Jeva's evaluated results do not establish that arbitrary real-world option descriptions are represented correctly. The encouraging change is narrower: a decision system can expose two different reasons for hesitation. Operators can then answer them differently—gather more evidence when candidates compete, or find another capability when none fits.
Questions, news tips or corrections? Contact [email protected].
