Atomic-SNLI: Fine-Grained Natural Language Inference through Atomic Fact Decomposition
Minghui Huang · 2026
Read Atomic-SNLI: Fine-Grained Natural Language Inference through Atomic Fact Decomposition by Minghui Huang (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-06528.txt.
Description
Judging whether a statement follows from a premise is usually done sentence by sentence, hiding the reasons. Minghui Huang proposes judging fact by fact in a 2026 paper on natural language inference. The atomic approach splits a hypothesis into individual facts and checks each, promising explainable verdicts.
Huang shows the naive rule, that a hypothesis holds only when every atomic fact holds, fails in practice because models reason much worse at the atomic level. The new Atomic-SNLI dataset decomposes the standard SNLI benchmark and enriches it with carefully generated atomic examples using linguistically informed strategies. Models fine-tuned on it improved significantly at atomic reasoning while keeping strong sentence-level performance. The result is inference that is both accurate and transparent at the level of individual facts.
Catalog record
| CREATOR | Minghui Huang |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 35 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2601-06528