What they found
The authors built ClosureBench, a test set of 960 instances. Each base scenario is posed under three rules for missing facts: OWA (missing means unknown), CWA (missing means false) and LCWA (a local version of the closed rule). Only the rule changes between variants. A model counts as correct on a scenario only if it is right in all of its variants. Overall accuracy ranged from 71.18% to 99.86% across replicated runs. Core switch accuracy, measured on scenarios where the right answer changes with the rule, ranged from 18.23% to 99.31%. The abstract says some agents answer unknown even when the prompt says missing information should be treated as false.
Why it matters
Many SI agents answer from incomplete records, such as a database that lacks an entry. Whether a missing entry means "unknown" or "no" can change the right answer. The authors argue that current tests mix this issue with other reasoning demands, so it goes unseen. Their benchmark isolates it and adds agent extensions for action choice, source scope and updates.
How it changes the race
This gives developers a narrow, checkable test of a failure that overall scores can hide. It is a measurement tool, not a new capability, so its effect depends on whether builders adopt it.
Safety and control
An agent that ignores the stated rule for missing information may act on wrong conclusions while looking accurate overall. Testing this directly helps check whether agents follow the instructions they are given.
The abstract, by the authors
AI agents built on large language models (LLMs) often answer questions from incomplete records. Missing information might be interpreted as unknown or false, depending on the completeness of the source. Current evaluations fail to distinguish these two aspects, as they are usually interwoven with other reasoning demands. We present ClosureBench, a benchmark that tests whether LLMs follow the stated rules for dealing with missing information. Every base scenario is tested under OWA, CWA, and LCWA, and only the treatment of missing information changes between variants. ClosureBench provides a verified base benchmark as well as agent extensions for action choice, source scope, and updates. A model is considered correct on a base scenario only if it provides correct answers in all of its closure variants. Across 960 instances, overall accuracy can mask substantial closure mistakes. Models achieve overall accuracy from 71.18% to 99.86% across replicated runs, but their core switch accuracy, measured on the scenarios whose gold label actually changes with the contract, varies between 18.23% and 99.31%. Large discrepancies emerge under LCWA. For example, some agents choose unknown despite the prompt indicating that missing information should be treated as false. Our findings demonstrate that evaluations of AI agents should not only measure whether agents can reason over existing facts but also whether they use the appropriate assumption when information is absent.
Quoted from arXiv:2611.26335. The research and its words belong to its authors.
