Toward Contemplative LLM: A Modular Framework for Evaluating and Enhancing LLM Alignment in Mental Health

Presentation slides of our paper "Toward Contemplative LLM: A Modular Framework for Evaluating and Enhancing LLM Alignment in Mental Health" presented at HARMONY 2026 in Pittsburgh, USA.

Transcript

1. Toward Contemplative LLM: A Modular Framework for Evaluating and Enhancing LLM Alignment in Mental Health
2. Toward Contemplative LLM: A Modular Framework for Evaluating and Enhancing LLM Alignment in Mental Health
3. Toward Contemplative LLM: A Modular Framework for Evaluating and Enhancing LLM Alignment in Mental Health
4. Toward Contemplative LLM: A Modular Framework for Evaluating and Enhancing LLM Alignment in Mental Health
5. Toward Contemplative LLM: A Modular Framework for Evaluating and Enhancing LLM Alignment in Mental Health
6. Toward Contemplative LLM: A Modular Framework for Evaluating and Enhancing LLM Alignment in Mental Health
7. Mental health problems are not new. Anxiety, grief, stress, and loneliness. Millions bring them to a chatbot.
8. So how do we know AI's answer is good?
9. But hold on — which human values are good? Alignment research is not new either. Contemplative traditions have asked exactly this question for millennia.
10. Contemplative principles may be an alternative paradigm for alignment.
11. And it already changes what models do. ↑ Cooperation ↓ Violations Incorporating contemplative principles into prompting strategies is promising!
12. Nobody has tested this systematically. It is hard. 01
13. Nobody has tested this systematically. It is hard. 02
14. Nobody has tested this systematically. It is hard. 03
15. So we built a modular framework. 01 02 03
16. Swap any piece. Keep the rest. + + Any combination
17. CARE-MH: modular inputs, modules, & outputs
18. Results that finally compare.
19. Taxonomy of mental health metrics
20. Define your own metrics — no technical skills needed. “ People should work on what they're best at.
21. Next: the prompting strategy module.
22. Lay the foundation: a modular, extensible infrastructure for evaluating and enhancing human-aligned AI.
23. Lay the foundation: a modular, extensible infrastructure for evaluating and enhancing human-aligned AI.
24. Lay the foundation: a modular, extensible infrastructure for evaluating and enhancing human-aligned AI.
26. My happiness tip: meditation (by far).
27. Toward Contemplative LLM Thank you. “ Alignment isn't a new AI problem. Contemplative practices have worked on it for millennia. The answers may be closer than we think.