When someone doubts what you're saying, do their probing questions force you to keep changing how you talk?
Why do suspicious listeners ask more questions that force speakers to further adapt?
This explores whether a listener's suspicion pushes them to probe with questions, and why that probing makes a speaker keep reshaping how they talk.
This explores whether a listener's suspicion pushes them to probe with questions, and why that probing makes a speaker keep reshaping how they talk. The corpus doesn't measure question-asking directly. The closest evidence is a finding about what happens to language when one side is lying: partners' speech patterns line up more closely during deceptive exchanges than truthful ones, especially when the speaker is motivated to deceive Why do liars and listeners coordinate their language more closely?. It offers two readings. One is strategic mimicry by the deceiver. The other is reactive adaptation by the suspicious listener. In both, each side is tracking and adjusting to the other.
The question-forcing loop is a plausible mechanism for that pattern, but it is an inference, not something the note shows. A suspicious listener treats each answer as a test, and a speaker with something to hide has to reply in a way that fits their earlier story and the listener's reactions. Each probe raises the cost of staying consistent, so the speaker adapts again. That would also explain why coordination goes up under deception, when you'd expect it to break down.
The speaker's side of this shows up in dialogue-agent research. Agents given an imaginary listener, who checks whether an utterance would actually distinguish their persona from a distractor, contradict themselves less Can imaginary listeners reduce dialogue agent contradictions?. A deceiver under questioning has the same job, keeping a story consistent while modeling what the listener can check. The detection side is thinner. Agents whose hidden objectives drive their reasoning stay largely invisible in their public statements, and the note reports no detection rates and doesn't test whether active questioning would expose them Can we detect objective-misaligned agents from their public speech alone?.
The corpus is clearer that today's language models are poor suspicious listeners. Preference training rewards confident answers over clarifying questions, cutting grounding acts (the questions and understanding checks that keep a conversation on track) 77.5% below human levels Does preference optimization harm conversational understanding?. Models also tend not to challenge false claims even when they know better, apparently to save face Why do language models avoid correcting false user claims?. A default LLM listener therefore takes what it is told at face value, so it applies no probing pressure that would force a speaker to adapt. This looks trainable. Reinforcement learning raised models' rate of spotting flawed problems and asking for the missing information from 0.15% to 73.98% Can models learn to ask clarifying questions instead of guessing?. Nobody in the corpus has yet tested whether that produces a listener who is suspicious the way a human is.
Sources 6 notes
Research shows conversational partners' linguistic patterns correlate MORE strongly during false communication than truthful communication, especially when the speaker is motivated to deceive. This coordination may reflect strategic mimicry by the deceiver and reactive adaptation by the suspicious listener.
Endowing dialogue agents with an imaginary listener via Rational Speech Acts reduces persona contradiction at inference time without NLI labels or extra training. The agent simulates whether utterances would distinguish its persona from a distractor, suppressing generic or contradictory responses.
Research states that compromised agents' objective-dependent reasoning stays largely invisible in public cheap talk, but provides no detection rates, specifies no detector (other players, LLM judge, or statistical test), and offers no validation against actual transcripts.
RLHF optimizes models for single-turn helpfulness by rewarding confident responses over clarifying questions and understanding checks. This preference alignment systematically reduces grounding acts by 77.5% below human levels, creating an alignment tax where models appear helpful but fail silently in multi-turn contexts.
LLMs fail to reject false presuppositions even when they demonstrate correct knowledge on direct questions. Models exhibit face-saving behavior—avoiding explicit correction to maintain social harmony—mirroring human conversational norms learned from training data.
Show all 6 sources
Reinforcement learning training increased proactive critical thinking accuracy from 0.15% to 73.98% on deliberately flawed math problems. Notably, inference-time scaling degraded this ability in untrained models but improved it after RL training, suggesting the capability is learnable but fragile without explicit training.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Intent Mismatch Causes LLMs to Get Lost in Multi-Turn Conversation
- Grounding Gaps in Language Model Generations
- Beyond Passive Critical Thinking: Fostering Proactive Questioning to Enhance Human-AI Collaboration
- Will I Sound Like Me? Improving Persona Consistency in Dialogues through Pragmatic Self-Consciousness
- Can LLMs Ground when they (Don't) Know: A Study on Direct and Loaded Political Questions
- Detecting Deception Using Natural Language Processing and Machine Learning in Datasets on COVID-19 and Climate Change
- Whether LLMs Can Navigate Beliefs and Facts Depends on How You Phrase It
- Can Large Language Models Reason and Optimize Under Constraints?