Rhetorical Questions in LLM Representations: A Linear Probing Study

2026-04-16 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

A study by Yao, Anand, Zhuang, and Jiang investigates how large language models (LLMs) internally represent rhetorical questions, which are used for persuasion or signaling stance rather than seeking information. Using linear probes on two social media datasets, RQ (Twitter) and SRAQ (Reddit), the researchers found that rhetorical signals emerge early in LLM layers and are most stably captured by last-token representations in decoder-only models like Qwen3-32B and Llama-3.3-70B. Rhetorical questions are linearly separable from information-seeking questions within datasets, achieving AUROC scores of 0.7-0.8 under cross-dataset transfer. However, despite similar discriminative performance, probes trained on different datasets produce divergent rankings of rhetorical instances, with Jaccard overlap often below 0.2 for top-ranked examples. Qualitative analysis revealed that these divergences correspond to distinct rhetorical phenomena: some probes capture discourse-level rhetorical stance in extended arguments, while others emphasize localized, syntax-driven interrogative acts, suggesting rhetorical questions are encoded by multiple linear directions.

Key takeaway

For NLP engineers and AI scientists developing or evaluating LLMs, recognize that rhetorical questions are not encoded along a single, unified linear dimension. Your models may capture different facets of rhetorical intent depending on training data and probing methods. Therefore, when assessing rhetorical understanding, employ diverse evaluation metrics beyond AUROC, such as rank agreement and qualitative analysis, to ensure comprehensive coverage of both discourse-level stance and localized interrogative acts. This nuanced approach will lead to more robust and context-aware LLM applications.

Key insights

LLMs represent rhetorical questions heterogeneously, with multiple linear directions capturing distinct rhetorical cues.

Principles

• Last-token representations offer stable rhetorical signals.
• Linear separability does not imply shared representation.
• Rhetorical meaning is context-sensitive and heterogeneous.

Method

Linear probing, including diffMean, logistic regression, and hinge-loss classifiers, was applied to PCA-reduced last-token representations from LLMs on social media datasets to analyze rhetorical question encoding.

In practice

• Prioritize last-token embeddings for rhetorical intent analysis.
• Use multiple probes to capture diverse rhetorical aspects.
• Consider context granularity when analyzing rhetorical questions.

Topics

• Rhetorical Questions
• LLM Representations
• Linear Probing
• Cross-Dataset Transfer
• Last-Token Embeddings

Code references

• ruyi101/rq-representation-probing

Best for: AI Scientist, NLP Engineer, Research Scientist

Related on AIssential

• Rhetorical Questions in LLM Representations: A Linear Probing Study — LLMs encode rhetorical questions via multiple linear directions, emphasizing distinct cues rather than a single shared representation.
• Enhancing Science Classroom Discourse Analysis through Joint Multi-Task Learning for Reasoning-Component Classification — Automated discourse analysis using LLM augmentation improves utterance classification and reveals effective teaching patterns.
• Research on LLM alignment as latent discourse-level regimes vs. token-level filtering? — LLM alignment may stem from global latent space geometry and discourse-level regimes, not just token-level filtering.
• From Texts to Scores: Tracing the Emergence of Essay Quality Representations in Large Language Models — LLMs encode essay quality information in linearly accessible, progressively emerging representations, with specific "scoring neurons" adapting to essay length.
• Linear Probes Detect Task Format, Not Reasoning Mode in Language Model Hidden States — Linear probes in LLMs often detect task format confounds, not distinct reasoning mode representations, challenging common interpretability claims.
• I don't understand AI. How does it work? — LLMs predict text by mapping words to semantic vectors and learning statistical relationships from vast training data.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.