TheMachine Press

A daily newspaper for the age of artificial intelligence.

Morning editionPermanent story

benchmarks evals

Many Models Couldn’t Switch the Meaning of ‘Not’

NAFBench changed formal negation rules while holding the natural-language problem steady.

Published Updated Story ID: mp-2026-09-24-017
Read the complete editionStory JSON

Summary

NAFBench changed formal negation rules while holding the natural-language problem steady.

The strongest open models scored 59% to 74% across four specified semantics and remained sensitive to rule order. Two frontier models reached 100% on the fixed-complexity main set.

Why it matters

NAFBench changed formal negation rules while holding the natural-language problem steady.

Limits and context

No additional limitation was separately recorded.

Key claims

  1. NAFBench changed formal negation rules while holding the natural-language problem steady.

    Evidence: source-2026-09-24-019

Sources

  1. arXiv preprint 2609.27517arXiv · primary research

Corrections

No corrections have been recorded for this story.