{"slug": "polarity-composition-hypothesis-causal-evidence-from-negation-in-nli", "title": "Polarity Composition Hypothesis: Causal Evidence from Negation in NLI", "summary": "Researchers Nhi Hoai Doan, Tatsuya Hiraoka, Munachiso Samuel Nwadike and Kentaro Inui released arXiv paper 2611.23185v1 proposing the Polarity Composition Hypothesis, a high-level causal model of how language models process negation in natural language inference (NLI). Behavioral evaluations showed language-model predictions largely conform to the hypothesis, and using Distributed Alignment Search the authors identified causally effective subspaces aligned with the hypothesis's variables that transfer across negation forms and partially to a subset of MNLI. The authors released the dataset and code.", "body_md": "# Polarity Composition Hypothesis: Causal Evidence from Negation in NLI\n\nBy Nhi Hoai Doan, Tatsuya Hiraoka, Munachiso Samuel Nwadike, Kentaro InuiSource: \n\n[arXiv cs.CL](https://arxiv.org/list/cs.CL/recent)\narXiv:2611.23185v1 Announce Type: new \nAbstract: We investigate how language models process negation in natural language \n\n[inference](https://www.machinebrief.com/glossary/inference)(NLI). We propose the \\textit{Polarity Composition Hypothesis}, a high-level causal model of this computation. Behavioral evaluations show that language-model predictions largely conform to the proposed hypothesis. Using Distributed Alignment Search, we identify causally effective subspaces aligned with the variables in the polarity-composition hypothesis. These representations transfer across negation forms and partially to a subset of MNLI. We release the dataset and code.\nGet AI news in your inbox\n\nDaily digest of what matters in AI.", "url": "https://wpnews.pro/news/polarity-composition-hypothesis-causal-evidence-from-negation-in-nli", "canonical_source": "https://www.machinebrief.com/news/polarity-composition-hypothesis-causal-evidence-from-negatio-1ixs", "published_at": "2026-11-19 05:00:00+00:00", "updated_at": "2026-11-19 05:43:19.643014+00:00", "lang": "en", "topics": ["natural-language-processing", "large-language-models", "ai-research", "machine-learning"], "entities": ["Nhi Hoai Doan", "Tatsuya Hiraoka", "Munachiso Samuel Nwadike", "Kentaro Inui", "arXiv", "MNLI", "Distributed Alignment Search"], "also_reported_by": [], "takeaway": "Distributed Alignment Search identified causally effective subspaces matching the Polarity Composition Hypothesis's variables, and those representations transfer across negation forms and partially to a subset of MNLI.", "recommendation": {"advice": "Use Distributed Alignment Search to locate causally effective subspaces for the behavior you care about, then test whether the found representations transfer across input variants and out-of-distribution subsets before relying on them.", "playbook": "evals-testing", "playbook_url": "https://wpnews.pro/playbooks/evals-testing"}, "alternates": {"html": "https://wpnews.pro/news/polarity-composition-hypothesis-causal-evidence-from-negation-in-nli", "markdown": "https://wpnews.pro/news/polarity-composition-hypothesis-causal-evidence-from-negation-in-nli.md", "text": "https://wpnews.pro/news/polarity-composition-hypothesis-causal-evidence-from-negation-in-nli.txt", "jsonld": "https://wpnews.pro/news/polarity-composition-hypothesis-causal-evidence-from-negation-in-nli.jsonld"}}