When Absence Is Evidence: Evaluating Completeness-Sensitive Negative Reasoning in Large Language Models

2026-08-05

AI Summary

This work addresses the tendency of large language models to conflate “unobserved” with “nonexistent” when performing negation reasoning under incomplete information, often failing to assess whether available evidence fully covers a given query. The study formally introduces the problem of completeness-aware negation reasoning and proposes CROWN-QA, a novel evaluation framework comprising the synthetic dataset CROWN-Synth and the real-document contrast set CROWN-Real. It incorporates paired controlled-variable designs and structured evidence coverage certificates to rigorously probe model behavior. Experiments reveal a pervasive over-closure bias across mainstream models, which systematically struggle to distinguish between falsifiable negations and genuinely unknown statements. Furthermore, error patterns exhibit consistent asymmetries across different models and prompting strategies, highlighting fundamental limitations in current approaches to reasoning under uncertainty.

Abstract

Large language models (LLMs) are often asked whether something is absent from a record, list, or retrieved context. Yet non-observation licenses a negative answer only when evidence completely covers the query scope; otherwise, the answer should remain unknown. We call this completeness-sensitive negative reasoning. We introduce CROWN-QA, comprising CROWN-Synth, a controlled paired core that fixes the question and observed facts while varying only query-relative coverage, and CROWN-Real, a real-document contrast-set evaluation with controlled coverage variants. Across three LLM families, models show unstable closure judgments and substantial over-closure, failing to reliably distinguish a justified negative answer (Certified-Negative) from insufficient evidence (Unknown). The dominant CROWN-Synth failure is asymmetric: models often recognize implicitly complete evidence yet treat implicitly partial evidence as query-covering. Prompting redistributes errors between over- and under-closure rather than consistently resolving them. Structured certificate elicitation traces many errors to evidence-coverage mischaracterization. CROWN-Real shows that the core partial-coverage asymmetry persists on real-document content, while its strength and the balance between over- and under-closure vary by model, prompt, and source.

Problem

• completeness-sensitive reasoning
• negative reasoning
• evidence coverage
• large language models
• absence as evidence

Innovation

• completeness-sensitive reasoning
• negative reasoning
• evidence coverage
• CROWN-QA
• large language models