Conference Papers Year : 2024

Strong hallucinations from negation and how to fix them

Swarnadeep Bhar
  • Function : Author
  • PersonId : 1487638

Abstract

Despite great performance on many tasks, language models (LMs) still struggle with reasoning, sometimes providing responses that cannot possibly be true because they stem from logical incoherence. We call such responses strong hallucinations and prove that they follow from an LM's computation of its internal representations for logical operators and outputs from those representations. Focusing on negation, we provide a novel solution in which negation is treated not as another element of a latent representation, but as an operation over an LM's latent representations that constrains how they may evolve. We show that our approach improves model performance in cloze prompting and natural language inference tasks with negation without requiring training on sparse negative data.
Embargoed file
Embargoed file
0 1 28
Year Month Jours
Avant la publication
Thursday, April 10, 2025
Embargoed file
Thursday, April 10, 2025
Please log in to request access to the document

Dates and versions

hal-04878406 , version 1 (10-01-2025)

Licence

Identifiers

  • HAL Id : hal-04878406 , version 1

Cite

Nicholas Asher, Swarnadeep Bhar. Strong hallucinations from negation and how to fix them. Association for Computational Linguistics (ACL), Association for Computational Linguistics, Aug 2024, Bangkok / Thailand, Thailand. pp.12670-12687. ⟨hal-04878406⟩
0 View
1 Download

Share

More