Focus particles and scalar inferences across humans and language models

What it is about

The paper asks whether humans and a language model give stable scalar judgments for sentences with the focus particles "even" and "only" (for example, "Mike can bake a cake"). The data are 60 ability sentences, 30 with each particle, rated on a 5-point scale by 108 human participants in four response-scale layouts (horizontal or vertical, standard or reversed labels). The model is Llama 3.3 70B, prompted 20 times per item at temperature 1.0, giving 4,800 observations. Both humans and the model rate "only" sentences higher than "even" sentences. The means are 4.31 vs 2.07 for humans and 5.00 vs 1.41 for the model. The model is more extreme, and its answers barely vary: a temperature sweep from 0.0 to 2.0 left them almost unchanged. The authors read the stable particle effect as evidence for an evaluative account, not a spatial one.

Ideas

• I1. Make the layout test a rule-based invariance check in prag/src/infer.rs: for each particle sentence, the implicature set must not change when the scale orientation is flipped. Log the result per item, run it twice and compare.
• I2. Report the spread of Mova's own judgments per item, not only the mean. The paper's main worry is zero variability. Mova is deterministic, so its spread should be exactly zero. Keep that as an explicit, stated limit, not as a success.
• I3. Build a small held-out set of about 100 "even"/"only" ability sentences, with human ratings collected once and kept out of development, to test whether Mova's ordering of particles matches people.
• I4. Use the paper's design as a negative control: shuffle the particle labels, and a working rule must score worse.

For Mova

The chapters are 33. Implicature and presupposition and 32. Politeness and hedging. The code is prag/src/infer.rs (implicatures) and prag/src/roles.rs. The measure is the particle-ordering agreement with the human means above, plus the invariance check across scale layouts. The LLM part is not for Mova. Only the design and the question are worth copying.

Cite

The paper itself, not this card. BibTeX from the library's record of the work. arXiv 2608.08227

@misc{brousse2026focus,
  title = {{Focus particles and scalar inferences across humans and language models}},
  author = {Brousse, Catherine M. and Radpour, Nelu D.},
  year = {2026},
  eprint = {2608.08227},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2608.08227}
}

References

9 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Dehaene, S., Bossini, S., & Giraux, P. (1993). The mental representation of parity and number magnitude. Journal of experimental psychology: General, 122(3), 371. ↗
2. Du, W., Yang, Y., & Welleck, S. (2025). Optimizing temperature for language models with multi-sample inference. arXiv preprint arXiv:2502.05234. ↗
3. Kryvosheieva, D., de Varda, A., Fedorenko, E., & Tuckute, G. (2025). Different types of syntactic agreement recruit the same units within large language models. arXiv preprint arXiv:2512.03676. ↗
4. Li, L., Sleem, L., Xu, Y., Song, Y., Jia, A., Francois, J., & State, R. (2026). The necessity of setting temperature in llm-as-a-judge. arXiv preprint arXiv:2603.28304. ↗
5. Proctor, R. W., & Cho, Y. S. (2006). Polarity correspondence: A general principle for performance of speeded binary classification tasks. Psychological bulletin, 132(3), 416. ↗
6. Qiu, M., Brisebois, Z., & Sun, S. (2025). Can llms simulate human behavioral variability? a case study in the phonemic fluency task. arXiv preprint arXiv:2505.16164. ↗
7. Rodriguez, M. A., Candito, M., & Huyghe, R. (2026). Modeling the human lexicon under temperature variations: Linguistic factors, diversity and typicality in llm word associations. arXiv preprint arXiv:2603.18171. ↗
8. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.-A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al. (2023). Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971. ↗
9. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., & Polosukhin, I. (2017). Attention is all you need. Advances in neural information processing systems, 30. find

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/brousse-2026-focus-particles-scalar-inferences-across/