Propositional compositionality in neural language models

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, quick

Summary

Research by Jane Li, Abhinav Patil, and Kyle Rawlins investigates propositional compositionality in neural language models, building on prior work demonstrating that truth conditions can be decoded from LMs and are linearly represented in intermediate layers. Their study argues that LMs can use propositional representations compositionally, supporting the "linear compositionality hypothesis." Specifically, the authors show two key findings: first, the truth values of individual conjuncts within a complex conjunction can be decoded independently of the overall conjunction's truth value. Second, they demonstrate that causally intervening on individual conjuncts directly impacts the truth value of the entire complex proposition. This work extends the understanding of how LMs process and represent logical structures.

Key takeaway

For AI Scientists developing or analyzing neural language models, this research suggests LMs possess a deeper, compositional understanding of propositional logic than previously assumed. You should consider these findings when designing models for tasks requiring robust logical reasoning or truth-condition sensitivity. This implies new avenues for improving model interpretability and reliability by exploring and manipulating these internal compositional representations.

Key insights

Neural language models exhibit propositional compositionality, allowing independent decoding and causal manipulation of conjunct truth values.

Principles

• LMs represent propositions compositionally.
• Conjunct truth values are independently decodable.
• Intervening on conjuncts causally impacts whole truth values.

Topics

• Neural Language Models
• Propositional Compositionality
• Linguistic Semantics
• Truth Conditions
• Causal Intervention
• Logical Connectives

Best for: Research Scientist, AI Scientist

Related on AIssential

• Leveraging LLM Parametric Knowledge for Fact Checking without Retrieval — Fact-checking without retrieval leverages LLM internal representations for robust, scalable claim verification.
• Compositionality and the lexicon in evolutionary semantics — The framework integrates formal semantics with evolutionary modeling to explain semantic universals like conservativity.
• Structural Preservation and the Logical Expressiveness of Graph Neural Networks — GNN classifier expressiveness is characterized by graded modal logic fragments based on structural preservation properties.
• Relational Linear Properties in Language Models: An Empirical Investigation — A new Kullback-Leibler divergence probing method efficiently tests relational linearity in language models, showing variations across models and sensitivity to phrasing.
• Beyond Surface Statistics: Robust Conformal Prediction for LLMs via Internal Representations — Internal LLM representations offer more robust uncertainty signals for conformal prediction, especially under distribution shifts.
• Beyond representational alignment with brain-guided language models for robust reasoning — Large language models can be directly enhanced by task-evoked brain signals to improve reasoning performance.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.