How do discourse-level patterns reveal cognitive distortions better than individual statements?
This explores why looking at how statements connect to each other—the reasoning that runs across a whole passage—catches distorted thinking that you'd miss by scoring words or sentences one at a time.
This explores why looking at how statements connect to each other—the reasoning that runs across a whole passage—catches distorted thinking that you'd miss by scoring words or sentences one at a time. The clearest evidence in the collection is on anxiety: the strongest predictor isn't anxious vocabulary but the *causal explanations a person stretches across multiple statements* Why do discourse patterns predict anxiety better than single words?. Anxious thinking works by overgeneralization—'this happened, so that will happen, so everything is at risk'—and that chain only becomes visible when you read statement-to-statement, not word-to-word. Tellingly, a model that combines both levels beats either alone, so the lexical signal isn't useless; it's just incomplete without the connective tissue.
Why would the connective tissue matter so much? Because a cognitive distortion is a *shape of reasoning*, not a flagged word. The work on detecting distortions directly leans into this: the best results come from separating the steps—judging subjectivity, reasoning by contrast, then mapping the underlying schema—rather than asking for a one-shot label, which lifts accuracy more than ten percent and produces explanations clinicians find usable Can structured prompting improve cognitive distortion detection?. The gain comes from forcing the analysis to operate at the level of how claims relate, which is exactly where the distortion lives.
The corpus suggests this 'patterns over points' principle generalizes well beyond clinical distress. Deception is another case where the single statement lies but the *interaction* gives it away: linguistic style matching rises during false communication, so the tell shows up in how a speaker and listener coordinate over a conversation, not in any one sentence Do liars and listeners coordinate their language during deception?. In all three cases—anxiety, distortion, deception—the meaningful signal is relational.
There's a sharp cautionary thread here too, and it's the part you might not expect: models are seductively good at the surface and weak at the structure. They default to surface strategies instead of genuinely tracking what someone believes Do large language models genuinely simulate mental states?, they miscalibrate how often a pattern like irony actually occurs because salient examples dominate training Do language models overestimate how often irony appears?, and their own reasoning traces turn out to be persuasive style rather than verified logic—invalid steps perform almost as well as valid ones Do reasoning traces show how models actually think?. So 'read the discourse' isn't only advice for human analysts; it's a warning that a system scoring local cues can look fluent while completely missing the cross-statement reasoning that defines the distortion.
The thing worth carrying away: the unit of analysis is the argument, not the utterance. Whether you're spotting an anxious overgeneralization, a clinical schema, or a lie, the diagnostic information is stored in the *relationships between statements*—and any method, human or machine, that flattens text into isolated points is structurally blind to it.
Sources 6 notes
Causal explanations across statements—not individual words—are the strongest predictor of anxiety because anxious thinking involves overgeneralization through inter-statement reasoning. A dual model combining both representation levels outperforms either alone.
DoT prompting separates subjectivity assessment, contrastive reasoning, and schema analysis to achieve 10%+ improvement over zero-shot ChatGPT. Expert evaluators rated the resulting explanations as clinically useful for case formulation.
Research shows interlocutors' linguistic styles correlate more during false communication than truthful communication, especially when the speaker is motivated to deceive. This coordination serves as a detectable deception signal through the listener's adaptive behavior, not just the liar's language.
ChangeMyView and FANTOM benchmarks show LLMs fail at authentic perspective-taking in open-ended scenarios, despite succeeding on structured tasks. Hybrid Bayesian architectures that force explicit belief tracking outperform LLM-alone approaches, suggesting the gap is architectural rather than merely training-based.
GPT-4o assigns significantly higher irony scores than humans (p < .001), revealing that LLMs detect irony as a pattern but miscalibrate its prevalence because ironic examples are more salient in training data than in actual use.
LLM reasoning traces perform as persuasive appearances rather than reliable explanations of computation. Invalid logical steps perform nearly as well as valid ones, and corrupted traces generalize comparably, showing that semantic correctness is not what produces the performance gains.