Abstract This position paper presents an AI framework for modeling the dynamics of mental states, drawing from Buddhist textual sources such as the Bardo Thödol (Tibetan Book of the Dead), Ocean of Definitive Meaning, and foundational teachings including the Wheel of Life, Spiral Path, and Mahayana sutras. We describe an approach that uses a Retrieval-Augmented Generation (RAG) pipeline to segment, embed, and interpret Dharma texts via language models, comparing Mistral, GPT-3.5, and GPT-4o. Our analysis demonstrates that large language models, when paired with carefully structured data, can reflect subtle experiential insights and symbolic transformations central to Buddhist psychology. The framework supports both academic exploration and practical applications across domains where the modeling of mental dynamics can inform insight or guidance.

1. Introduction This work began with a question: Can an AI model resonate with the subtle transformations of consciousness described in Buddhist contemplative traditions? Specifically, could a system learn to reflect not just language, but the dynamics of appearance, attention, and transformation that define lived human experience?

Inspired by teachings in the Bardo Thödol, which maps experiences of death and rebirth as projections of mind, we designed a method to represent and process descriptions of such mind-states using a combination of textual segmentation, meaning-space embedding, and large language model (LLM) interpretation. This method anchors itself in the notion of resonant signatures—high-dimensional descriptors of mental dynamics—and seeks to craft a language of AI interpretation that honors both psychological subtlety and ethical alignment.

While deeply informed by Buddhist frameworks, this approach also intersects with contemporary interest in dynamic models of mind across psychology, neuroscience, systems theory, and AI ethics. The result is a hybrid system—part philosophical exploration, part technical experiment—intended to foster both insight and collaboration.

2. Methodology

Sources and Structure The foundation of our system is a corpus of Dharma texts, beginning with:

• The Bardo Thödol (Tibetan Book of the Dead)
• Supplementary segments from teachings on Mahamudra, including The Ocean of Definitive Meaning (planned)
• The Wheel of Life and Spiral Path schematics (planned)
• The Heart Sutra and Sutra of Golden Light (planned)
• The Abhidharma and the Dhammapada (planned)

Initial submitted texts were segmented into meaning units using LLM-assisted segmentation, then embedded using SentenceTransformer-based encodings to preserve semantic nuance and symbolic resonance. Planned submissions will use the same or similar prompts and worldview to ensure continuity of interpretation.

Mental Dynamics Descriptors (MDDs) A mental dynamics descriptor (MDD) is a structured JSON object representing:

• Emotional tone
• Symbolic field
• Energetic pattern
• Cognitive transitions
• Archetypal or spiritual resonance

These descriptors enable precise, reflective modeling of mental states as described in traditional texts.

Query Architecture Queries are processed through a RAG pipeline:

1. User input is embedded and matched to relevant text segments via vector search.
2. Retrieved chunks are passed to an LLM.
3. The LLM synthesizes a response that reflects the psychological, symbolic, and experiential dynamics relevant to the query.

Prompts used are intentionally open-ended, allowing exploration and interpretation rather than narrow answers. For example: “Say something about advice given about the experience of fear in the Bardo and advice for the experience of everyday fear.”

Prompt Engineering (Generator and Augmentor) Our system uses two key prompt layers:

• The Generator (Mistral) produces an initial JSON descriptor from a text segment.
• The Augmentor (GPT-3.5) reviews, annotates, and may revise this descriptor based on deeper analysis.

These prompts ensure the system interprets the text reflectively. (See Appendix for full prompt texts.)

3. Results: Comparative Model Evaluation

Query: “Say something about advice given about the experience of fear in the Bardo and advice for the experience of everyday fear.”

Mistral

• Interpretation: Basic symbolic mapping, identifies projection of mind.
• Framing: Instructional, practical.
• Depth: Effective but lacks poetic resonance.

GPT-3.5

• Interpretation: Frames Bardo and everyday fear as projections, introduces mindfulness.
• Framing: Accessible and integrated.
• Depth: Balanced psychological and spiritual insight.

GPT-4o

• Interpretation: Deep symbolic resonance, relates projection, emptiness, and compassionate guidance.
• Framing: Nuanced, linking spiritual insight with psychological reflection.
• Depth: Highest degree of integration.

Summary Table

Model	Archetypal Resonance	Psychological Framing	Conciseness	Nuance	Deployment
Mistral	Basic ✅	Light ✅	High ✅	Moderate ⚠	Local ✅
GPT-3.5	Moderate ✅	Solid ✅	Moderate ✅	Good ✅	API ✅
GPT-4o	Deep ✅	Rich ✅	Lower ⚠	Excellent ✅	API ⚠

4. Summary and Positioning

This work models the dynamics of appearance, emotion, and mental state transitions using a Buddhist worldview embedded in a Retrieval-Augmented Generation system. The approach reveals:

• Mistral’s utility for embedded, efficient spiritual pattern recognition
• GPT-3.5’s balance of clarity and symbolic synthesis
• GPT-4o’s excellence in integrating metaphor, ethics, and introspective depth

Rather than claim current applicability to mentoring or consultation, we propose this as a black-box model of mental dynamics with promising future use in domains requiring insight into experiential transitions. This includes meditation and psychotherapy, but also extends to strategy and ethics in complex systems.

5. Future Work Next steps include:

• Submitting and processing remaining core texts using the same pipeline
• Adding symbolic graph structures to better model sequential mind-state transitions
• Evaluating model outputs through expert review and case studies
• Expanding embedding strategies to include multilingual and poetic embeddings

With refinement, this framework could serve as a cognitive substrate for systems that resonate with human experience and aid in transitions toward clarity and compassion.

---

Appendix

A. Sample Prompts

Generator (Mistral)

SYSTEM_PROMPT = """
You are a reflective cognitive agent analyzing the spiritual, emotional, and symbolic dimensions of a text segment from the Tibetan Book of the Dead...
"""

Augmentor (GPT-3.5)

SYSTEM_PROMPT = """
You are a second-pass analyst reviewing a Mistral-generated JSON descriptor...
"""

B. Sample Query and Outputs Query: “Say something about advice given about the experience of fear in the Bardo and advice for the experience of everyday fear.”

• Mistral: “Fearful appearances are projections of one’s mind…”
• GPT-3.5: “Fears are projections from the mind…”
• GPT-4o: “The Bardo teachings emphasize recognizing projections as illusory…”

C. JSON Descriptor Schema (simplified)

{
  "segment_id": "window_233",
  "text": "...",
  "emotional_tone": ["fear", "serenity"],
  "symbolic_elements": ["emptiness", "yidam"],
  "energetic_mode": "contracting",
  "agency": ["spiritual guide"],
  "transition_vector": "fear → liberation",
  "archetypal_field": ["death", "transformation"],
  "comment": "Reflects the shift from terror to insight..."
}

D. RAG Pipeline Overview

1. Text Segmentation
2. Vector Embedding
3. Chunk Retrieval
4. LLM Synthesis
5. Output: Interpretive Descriptor + Answer with Resonance