Large Language Models Do Multi-Label Classification Differently

Multi-label classification is prevalent in real-world settings, but the behavior of Large Language Models (LLMs) in this setting is understudied. We investigate how autoregressive LLMs perform multi-label classification, with a focus on subjective tasks, by analyzing the output distributions of the models in each generation step. We find that their predictive behavior reflects the multiple steps in the underlying language modeling required to generate all relevant labels as they tend to suppress all but one label at each step. We further observe that as model scale increases, their token distributions exhibit lower entropy, yet the internal ranking of the labels improves. Finetuning methods such as supervised finetuning and reinforcement learning amplify this phenomenon. To further study this issue, we introduce the task of distribution alignment for multi-label settings: aligning LLM-derived label distributions with empirical distributions estimated from annotator responses in subjective tasks. We propose both zero-shot and supervised methods which improve both alignment and predictive performance over existing approaches.

Humans Hallucinate Too: Language Models Identify and Correct Subjective Annotation Errors With Label-in-a-Haystack Prompts

Modeling complex subjective tasks in Natural Language Processing, such as recognizing emotion and morality, is considerably challenging due to significant variation in human annotations. This variation often reflects reasonable differences in semantic interpretations rather than mere noise, necessitating methods to distinguish between legitimate subjectivity and error. We address this challenge by exploring label verification in these contexts using Large Language Models (LLMs). First, we propose a simple In-Context Learning binary filtering baseline that estimates the reasonableness of a document-label pair. We then introduce the Label-in-a-Haystack setting: the query and its label(s) are included in the demonstrations shown to LLMs, which are prompted to predict the label(s) again, while receiving task-specific instructions (e.g., emotion recognition) rather than label copying. We show how the failure to copy the label(s) to the output of the LLM are task-relevant and informative. Building on this, we propose the Label-in-a-Haystack Rectification (LiaHR) framework for subjective label correction: when the model outputs diverge from the reference gold labels, we assign the generated labels to the example instead of discarding it. This approach can be integrated into annotation pipelines to enhance signal-to-noise ratios. Comprehensive analyses, human evaluations, and ecological validity studies verify the utility of LiaHR for label correction. Code is available at https://github.com/gchochla/LiaHR.

Deep Learning Characterizes Depression and Suicidal Ideation from Eye Movements

Identifying physiological and behavioral markers for mental health conditions is a longstanding challenge in psychiatry. Depression and suicidal ideation, in particular, lack objective biomarkers, with screening and diagnosis primarily relying on self-reports and clinical interviews. Here, we investigate eye tracking as a potential marker modality for screening purposes. Eye movements are directly modulated by neuronal networks and have been associated with attentional and mood-related patterns; however, their predictive value for depression and suicidality remains unclear. We recorded eye-tracking sequences from 126 young adults as they read and responded to affective sentences, and subsequently developed a deep learning framework to predict their clinical status. The proposed model included separate branches for trials of positive and negative sentiment, and used 2D time-series representations to account for both intra-trial and inter-trial variations. We were able to identify depression and suicidal ideation with an area under the receiver operating curve (AUC) of 0.793 (95% CI: 0.765-0.819) against healthy controls, and suicidality specifically with 0.826 AUC (95% CI: 0.797-0.852). The model also exhibited moderate, yet significant, accuracy in differentiating depressed from suicidal participants, with 0.609 AUC (95% CI 0.571-0.646). Discriminative patterns emerge more strongly when assessing the data relative to response generation than relative to the onset time of the final word of the sentences. The most pronounced effects were observed for negative-sentiment sentences, that are congruent to depressed and suicidal participants. Our findings highlight eye tracking as an objective tool for mental health assessment and underscore the modulatory impact of emotional stimuli on cognitive processes affecting oculomotor control.