AI Framework Improves Negative Thought Identification in Emotional Support

Hainiu Xu, Zhaoyue Sun, Hanqi Yan, Jinhua Du, Caroline Catmur, Yulan He· August 3, 2026 View original

Key takeaways

  • LLMs for emotional support need to identify salient cognitive appraisal dimensions, not just all possible ones.
  • The AppraiSal benchmark provides human-annotated data for this specific task.
  • PRISM, a multi-agent probabilistic framework, significantly improves LLMs' ability to infer context-specific appraisals.
  • Better identification of appraisal dimensions leads to more effective negative thought reframing.

Who benefits

Healthcare (Mental Health)Customer ServiceAI DevelopmentSocial Work

Summary

This research introduces PRISM, a multi-agent probabilistic framework that enhances LLMs' ability to identify salient cognitive appraisal dimensions in emotional support conversations. It addresses the limitation of current models that exhaustively evaluate all dimensions without considering context-specific saliency.

Large Language Models are increasingly being deployed for sensitive tasks like emotional support, particularly in reframing negative thoughts. This process relies on understanding cognitive appraisals—how individuals interpret events that trigger negative emotions. Existing AI frameworks often evaluate every possible appraisal dimension, overlooking the fact that only a few are truly salient in any given context. To address this, researchers developed the AppraiSal benchmark, a dataset of nearly a thousand emotional support conversations with human-annotated salient cognitive appraisal dimensions. They also propose PRISM, a multi-agent probabilistic framework grounded in Bayesian Inverse Planning. PRISM significantly improves LLMs' capacity to infer these context-specific, salient appraisal dimensions, leading to more targeted and effective emotional support interventions across various LLM sizes.

Why it matters

For professionals developing AI for mental health support, customer service, or empathetic communication, this research offers a path to building more nuanced and effective systems that can better understand and respond to human emotional states.

How to implement this in your domain

  1. 1Integrate context-aware appraisal identification mechanisms, like PRISM, into AI models designed for emotional support or empathetic dialogue.
  2. 2Develop training datasets that include human annotations of salient cognitive appraisal dimensions to improve model performance.
  3. 3Design multi-agent AI architectures for emotional support systems to leverage probabilistic reasoning for better context understanding.
  4. 4Apply insights from this research to refine prompt engineering strategies for LLMs in sensitive conversational AI applications.

Original post by Hainiu Xu, Zhaoyue Sun, Hanqi Yan, Jinhua Du, Caroline Catmur, Yulan He

"arXiv:2607.28648v1 Announce Type: cross Abstract: Large Language Models (LLMs) are increasingly used for emotional support tasks, such as negative thought reframing. This task relies on modifying cognitive appraisals, the subjective interpretation of events that elicit negative e…"

View on X

Originally posted by Hainiu Xu, Zhaoyue Sun, Hanqi Yan, Jinhua Du, Caroline Catmur, Yulan He on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses