Elamur Research⌕
☰
home / papers / 2026 / Analysis of the Neglect-Zero Effect in Large Language Models

Analysis of the Neglect-Zero Effect in Large Language Models

Paper · 2026 · Jin Tanaka, Daiki Matsuoka, Ryoma Kumon et al.

Summary

This study asks whether large language models share a human bias called the neglect-zero effect. Humans tend to disregard "zero-models": situations where a statement is only trivially true because some set is empty. The authors adapt a human experiment (Klochowicz et al., 2025) for LLMs and combine it with methodology from Tsvilodub et al. (2024). The design uses structural priming inside a two-choice picture-matching task, the "better-picture paradigm." In each prime trial, the model is pushed to accept a zero-model. The authors then check whether this acceptance carries over to a target trial. Primes cover two inferences thought to arise from neglect-zero, ESQ and DIS, plus upper-bounded scalar implicature (UPP) as a control driven by a different mechanism. Targets always involve ESQ. Model responses are analyzed with GLMMs. No tested model behaves like humans, since priming effects are either negative or non-significant. Some models appear to rely mostly on literal meaning, while others react to zero-models in a non-human way. The findings add evidence to the debate on how human-like LLM language processing is.

Key points

  • The three inference types are ESQ (non-empty-scope strengthening with superlative quantifiers, e.g. "at most two"), DIS (distributive inference with "each … or …") and UPP (some → not all). Prime trials use one of the three, and targets are fixed to ESQ.
  • Six models were tested, including the Gemma-3 series, Gemma-3-27B, Llama-4-Scout and GPT-5 nano.
  • For Gemma-3-27B and Llama-4, every sub-experiment showed a significant priming effect in the negative direction. This points to some sensitivity to zero-models, though not in the human direction. In Llama-4, UPP also behaved differently from ESQ and DIS. The authors flag their own GLMM fit for Gemma-3-27B as inappropriate.
  • For GPT-5 nano, priming was non-significant everywhere. Its very high accuracy, like that of Gemma-3-27B, suggests these models mostly process literal meaning.
  • The code is released at github.com/ynklab/neglect_zero.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓The University of TokyoTokyo · Japan90 works · 9,156 citations🔬RIKENWako · Japan26 works · 526 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 works

🧭 Topics 1Research subjects and languages the work is about.Language models & pretraining · 3,320 works

Cite

Cite as Jin Tanaka, Daiki Matsuoka, Ryoma Kumon and Hitomi Yanaka. 2026. Analysis of the Neglect-Zero Effect in Large Language Models. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.acl-srw.91

@inproceedings{tanaka2026analysis,
  title = {{Analysis of the Neglect-Zero Effect in Large Language Models}},
  author = {Jin Tanaka, Daiki Matsuoka, Ryoma Kumon and Hitomi Yanaka},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-srw.91},
  url = {https://doi.org/10.18653/v1/2026.acl-srw.91}
}

References

12 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Maria Aloni. 2022. Logic and conversation: The case of free choice. Semantics and Pragmatics, 15(5):1–60.
  2. Kathryn Bock. 1986. Syntactic persistence in language production. Cognitive Psychology, 18(3):355–387. find
  3. Lewis Bott and Emmanuel Chemla. 2016. Shared and distinct mechanisms in deriving linguistic enrichment. Journal of Memory and Language, 91:117–140. find
  4. Laurence Horn. 1984. Towards a new taxonomy for pragmatic inference: Q-based and R-based implicature. In Deborah Schiffrin, editor, Meaning, Form and Use in Context, pages 11–42. find
  5. Jaap Jumelet, Willem Zuidema, and Arabella Sinclair. 2024. Do language models exhibit human-like structural priming effects? In Findings of the Association for Computational Linguistics: ACL 2024, pages 14727–14742, Bangkok, Thailand. Association for Computational Linguistics. find
  6. Tomasz Klochowicz, Fabian Schlotterbeck, Sonia Ramotowska, Oliver Bott, and Maria Aloni. 2025. Neglect zero: Evidence from priming across constructions. In Proceedings of the Annual Meeting of the Cognitive Science Society, volume 47, pages 5954–5960. find
  7. Kate McCurdy, Sharon Goldwater, and Adam Lopez. 2020. Inflecting when there’s no majority: Limitations of encoder-decoder neural networks as cognitive models for German plurals. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1745–1756, Online. Assoc… find
  8. Marie-Christine Meyer and Roman Feiman. 2021. Priming reveals similarities and differences between three purported cases of implicature: Some, number and free choice disjunctions. Journal of Memory and Language, 120:104206. find
  9. Qian Niu, Junyu Liu, Ziqian Bi, Pohsun Feng, Benji Peng, Keyu Chen, Ming Li, Lawrence KQ Yan, Yichao Zhang, Caitlyn Heqi Yin, Cheng Fei, Tianyang Wang, Yunze Wang, Silin Chen, Ming Liu, Ziyuan Qin, Riyang Bao, Xinyuan Song, and Zekun Jiang. 2025. Large language models and cognitive science: A compre… — [arXiv:2409.02387](https://arxiv.org/abs/2409.02387) find
  10. Martin Pickering and Victor Ferreira. 2008. Structural priming: A critical review. Psychological Bulletin, 134(3):427–459. find
  11. Polina Tsvilodub, Paul Marty, Sonia Ramotowska, Jacopo Romoli, and Michael Franke. 2024. Experimental pragmatics with machines: Testing LLM predictions for the inferences of plain and embedded disjunctions. In Proceedings of the Annual Meeting of the Cognitive Science Society, volume 46, pages 3960–…
  12. Akhila Yerukola, Saujas Vaduguru, Daniel Fried, and Maarten Sap. 2024. Is the pope catholic? yes, the pope is catholic. generative evaluation of non-literal intent resolution in LLMs. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers),… find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/tanaka-2026-analysis-neglect-zero-effect-large/