LLMs Detect Rhetorical Questions Early: Linear Probe Study

A new arXiv preprint from April 2026 reveals that large language models represent rhetorical questions as linearly separable from information-seeking questions, with signals emerging in early layers and most stable at the last token. This finding challenges the notion that pragmatic intent is a high-level reasoning task and opens the door to simple, probe-based detection of persuasive language.

Linear probes can distinguish rhetorical from information-seeking questions in LLM hidden states with high accuracy.
Rhetorical signals emerge in early transformer layers (layers 2-4) and are most stably captured by the last token representation.
The findings suggest that pragmatic intent is not a high-level cognitive abstraction but a low-level, linearly separable feature in LLMs.

What makes a rhetorical question different in an LLM's internal representation?

According to the study published on arXiv (April 15, 2026), the authors used two social-media datasets—one from Reddit and one from Twitter—to compare how LLMs represent rhetorical questions versus information-seeking questions. They trained linear probes on the hidden states of models like Llama-2-7B and Mistral-7B. The probes achieved over 85% accuracy in distinguishing the two question types within the same dataset, indicating that the models encode a clear, linearly separable feature for rhetorical intent.

The key insight is that this distinction is not a byproduct of surface-level cues like question length or punctuation. The probes generalized across different discourse contexts (e.g., debate vs. casual chat) but showed a drop in cross-dataset transfer, suggesting that rhetorical style is partly dataset-specific.

Why does the early emergence of rhetorical signals matter?

The authors reported that rhetorical signals appear as early as layer 2 in a 7B-parameter model. This is significant because it implies that LLMs do not need to reason deeply about context to detect rhetorical intent—the feature is encoded almost immediately. This challenges the prevailing view that pragmatic understanding requires multi-step reasoning or world knowledge.

For safety researchers, this is a double-edged sword. On one hand, it means that lightweight probes could be used to flag potentially manipulative or persuasive language in real-time. On the other hand, it suggests that LLMs may be more susceptible to adversarial exploitation of this early encoding, where a simple prompt tweak could flip the rhetorical signal.

LLMs Detect Rhetorical Questions Early: A Linear Probe Success

How do these findings compare to existing work on LLM pragmatics?

Prior work, such as the 2023 study by Hu et al. on 'Pragmatic Reasoning in LLMs,' argued that understanding rhetorical questions required chain-of-thought reasoning or explicit context. The current study directly contradicts that, showing that a linear probe—a simple classifier—can match or exceed the performance of much larger models on this task without any reasoning steps.

This suggests that the 'reasoning' we observe in LLMs may sometimes be a post-hoc rationalization of a pre-existing, linearly separable feature. The implications for interpretability are profound: if intent is encoded early and linearly, then we can build simpler, more transparent detectors for harmful language.

What are the key limitations of this study?

The study explicitly notes that its datasets are limited to English social-media text, and the rhetorical questions are annotated by crowdworkers, which may introduce bias. Furthermore, the linear probes achieve lower accuracy (around 70%) when tested cross-dataset, indicating that the representation of rhetorical intent is not universal but is shaped by the specific discourse community.

Another limitation is that the study only tested on open-source models (Llama-2, Mistral). It remains unclear whether proprietary models like GPT-4 or Claude exhibit the same early, linear separability. The authors call for replication on larger, more diverse models and datasets.

Who benefits most from this line of research?

Stakeholder	Gain	Risk
Safety teams	Simple, fast detection of manipulative language	May over-rely on probes, ignoring context
Adversarial users	Could craft prompts that evade probe detection	Detection is still imperfect
Interpretability researchers	Evidence for early, linear encoding of intent	May overgeneralize to other pragmatic phenomena
Model developers	Can use probes to audit training data for bias	May need to retrain probes for each new domain
Verdict	Safety teams gain the most immediate, practical benefit, but adversarial users are a close second if probes become widely deployed without safeguards.

My thesis is that this study is a wake-up call for the AI safety community: the battle against manipulative language may be fought at the feature level, not the reasoning level. In the short term, we will see a rush to deploy linear probes for content moderation, but this is a fragile solution. In the long term, the real winner is interpretability science, as we now have a concrete, testable hypothesis about how LLMs encode speaker intent. The loser is the 'black box' narrative—if intent is linearly separable, then we can build transparent, auditable detectors. My prediction: within 12 months, at least one major social media platform will deploy a linear-probe-based filter for rhetorical questions, and it will be bypassed within 3 months by adversarial prompts that shift the early-layer representation.

Predictions

By Q3 2027, a major social media platform (e.g., Reddit or Twitter/X) will announce a linear-probe-based content moderation system for detecting rhetorical questions used in harassment or misinformation.
By Q1 2028, at least one academic paper will demonstrate a successful adversarial attack on such a probe, using minimal token perturbations to flip the rhetorical signal.
By 2029, the finding that pragmatic intent is linearly separable will be replicated on at least three different model architectures, leading to a new subfield of 'feature-level pragmatics' in NLP.

Article Summary

LLMs encode rhetorical questions as a linearly separable feature, not a high-level reasoning product.
The signal emerges in early layers (2-4), challenging the assumption that pragmatic understanding requires deep context.
Cross-dataset transfer is limited, indicating that rhetorical style is partly community-specific.
Safety teams gain a lightweight detection tool, but adversarial robustness remains a critical open problem.
This research opens a new path for interpretability: probing for speaker intent at the feature level.