← News

ResearchSources: arXiv · 18 Nov 2026

Researchers propose causal account of how language models handle negation in NLI

A new arXiv paper introduces the Polarity Composition Hypothesis, a causal model of negation processing in natural language inference, supported by behavioral tests and representation search.

In plain words

Natural language inference is the task of deciding whether one sentence entails another, and negation (words like "not") often flips that answer. The authors propose a hypothesis describing how models combine polarity information to make such decisions. Their tests suggest language models behave as the hypothesis predicts.

What was announced

The paper, released on arXiv, examines how language models process negation when performing natural language inference. The authors propose the Polarity Composition Hypothesis, which they describe as a high-level causal model of that computation. Behavioral evaluations reported in the abstract indicate that language-model predictions largely conform to the hypothesis.

Using a technique called Distributed Alignment Search, the researchers identified subspaces inside the models that causally correspond to the variables in their hypothesis. According to the paper, these representations transfer across different forms of negation and only partially to a subset of MNLI, a standard inference dataset. The authors state they are releasing the dataset and code alongside the paper.

What it does not show

  • The evidence is limited to the paper's abstract; full experimental details, model sizes and datasets are not available in the supplied passage.
  • The abstract does not name which language models were tested or how large the behavioral and causal effects are beyond the word "largely".
  • Transfer to MNLI is described as partial and limited to a subset, so generalization to broader inference tasks is not established.

Sources

Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 19 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.