@inproceedings{li-etal-2026-propositional,
    title = "Propositional compositionality in neural language models",
    author = "Li, Jane  and
      Patil, Abhinav  and
      Rawlins, Kyle",
    editor = "Voigt, Rob  and
      Warstadt, Alex  and
      Feldman, Naomi  and
      Linzen, Tal",
    booktitle = "Proceedings of the Society for Computation in Linguistics 2026",
    month = jul,
    year = "2026",
    address = "San Diego, CA",
    publisher = "Association for Computational Linguistics",
    url = "https://preview.aclanthology.org/ingest-acl-workshops/2026.scil-main.34/",
    pages = "365--368",
    ISBN = "979-8-89176-412-5",
    abstract = "One of the most fundamental representations in linguistic semantics is that of the proposition (McGrath and Frank, 2005), standardly taken as the carrier of truth-conditions. Recent work shows that some form of truth can be decoded from language models (Azaria and Mitchell, 2023; Li et al., 2023), and strikingly, that for some models, truth is even represented linearly in intermediate layers (Marks and Tegmark, 2024, GoT). We take this line of work a step further and argue that neural language models can use propositional representations compositionally (Janssen 2010; Pickel and Szab{\'o} 2025 a.o.), drawing from evidence of the behaviour of logical connectives: the linear compositionality hypothesis. Specifically, we show (a) that the truth values of individual conjuncts can be decoded independently of the truth value of a complex conjunction, and (b) that we can causally intervene on individual conjuncts in a way that affects the truth value of the whole."
}