Mental health is a salient and growing societal concern among college students. Social media platforms such as Reddit offer a rich source of data regarding how students talk about their mental health, and NLP tools may potentially assist in identifying when a student is struggling. In this paper, we investigate how different NLP tools can be used to extract context surrounding college students expressions of distress. We construct a novel dataset from Reddit posts (College Distress on Reddit, or CDR), and examine the "classical NLP pipeline", and modern generative LLMs on this data. Our dataset exploration is conducted in parallel with, and contrasted against the Dreaddit dataset to examine cross-domain variation. Results show that standard or "classical" NLP tools extract a limited number of concrete entities, whereas generative models can infer more nuanced causes. However, LLMs struggle with knowledge extraction in specific content areas. Our work shows how important it is to be wary of LLMs, especially in mental health contexts.

This project note describes challenges and procedures undertaken in annotating an audiovisual dataset capturing a multimodal situated collaborative construction task. In the task, all participants begin with different partial information, and must collaborate using speech, gesture, and action to arrive a solution that satisfies all individual pieces of private information. This rich data poses a number of annotation challenges, from small objects in a close space, to the implicit and multimodal fashion in which participants express agreement, disagreement, and beliefs. We discuss the data collection procedure, annotation schemas and tools, and future use cases.

AI support of collaborative interactions entails mediating potential misalignment between interlocutor beliefs. Common preference alignment methods like DPO excel in static settings, but struggle in dynamic collaborative tasks where the explicit signals of interlocutor beliefs are sparse and skewed. We propose the Frictional Agent Alignment Framework (FAAF), to generate precise, context-aware “friction” that prompts for deliberation and re-examination of existing evidence. FAAF’s two-player objective decouples from data skew: a frictive-state policy identifies belief misalignments, while an intervention policy crafts collaborator-preferred responses. We derive an analytical solution to this objective, enabling training a single policy via a simple supervised loss. Experiments on three benchmarks show FAAF outperforms competitors in producing concise, interpretable friction and in OOD generalization. By aligning LLMs to act as adaptive “thought partners”—not passive responders—FAAF advances scalable, dynamic human-AI collaboration. Our code and data can be found at https://github.com/csu-signal/FAAF_ACL.