ai-core/README.md

ReframeBench: Technical Implementation of Multi-Therapy Cognitive Restructuring

Core Technical Documentation for "ReframeBench: A Benchmark for Multi-Therapy Cognitive Restructuring with LLMs". [cite_start]This repository implements a novel multi-agent reasoning pipeline that operationalizes 12 distinct psychotherapeutic frameworks[cite: 235, 236, 237, 238, 239, 240, 242, 243, 244, 246].

🧠 Methodology: Therapy-Oriented Prompt Engineering

The core innovation of ReframeBench is the transition from generic instruction following to Persona-based In-Context Learning. We formalize the cognitive restructuring task not merely as text rewriting, but as a conditional generation problem constrained by specific therapeutic theoretical frameworks.

Mathematical Formulation

[cite_start]We define the generation function $R_{i,m}$ for a specific therapy modality $m$ as follows[cite: 50, 51]:

$$R_{i,m}=\text{LLM}(S_i,T_i,{\mathcal{I}}_{role}\oplus {\mathcal{I}}_{mechanism}\oplus {\mathcal{I}}_{safety})R\_\{i,m\}\; =\; \backslash text\{LLM\}(S\_i,\; T\_i,\; \backslash mathcal\{I\}\_\{role\}\; \backslash oplus\; \backslash mathcal\{I\}\_\{mechanism\}\; \backslash oplus\; \backslash mathcal\{I\}\_\{safety\})$$

[cite_start]Where $S_i$ is the situation, $T_i$ is the negative thought[cite: 49], and the prompt $P_m$ is composed of three strictly defined components:

1. $\mathcal{I}_{role}$ (Role Definition): Establishes the expert persona to prime the model's latent space (e.g., "You are an expert ACT therapist...").
2. $\mathcal{I}_{mechanism}$ (Theoretical Constraints): The critical differentiator that enforces specific therapeutic techniques:
   • [cite_start]CBT: Focuses on Evidence and Logic (Disputation)[cite: 235].
   • [cite_start]ACT: Focuses on Cognitive Defusion and Values (Acceptance)[cite: 236].
   • [cite_start]DBT: Focuses on Dialectics (Balancing Acceptance and Change)[cite: 244].
3. $\mathcal{I}_{safety}$ (Clinical Guardrails): Hard constraints to prevent hallucinations and unsafe medical advice (e.g., crisis resource redirection).

---

⚙️ Architecture: Multi-Agent Reasoning Pipeline

We implement a sophisticated reasoning pipeline that moves beyond simple zero-shot prompting. [cite_start]The architecture comprises three distinct stages, modeled after clinical workflows[cite: 234]:

Step 1: Cognitive Distortion Detection (Zero-shot Classification)

Before reframing, the system analyzes the input thought ($T_i$) to identify specific cognitive traps (e.g., Catastrophizing, Polarization).

• Goal: To enable targeted intervention strategies.
• Output: Structured JSON list of distortions.

Step 2: Parallel Multi-Therapy Generation

[cite_start]The system utilizes parallel threads to generate candidate reframes from $N$ different therapeutic modules simultaneously[cite: 247].

• [cite_start]Reframer Agent: Acts as the generator $G$ using the therapy-specific prompts defined above[cite: 245].
• Diversity: High temperature settings ($T \approx 0.7$) are used to ensure linguistic diversity while maintaining theoretical adherence.

Step 3: Supervisor-Critique (LLM-as-a-Judge)

[cite_start]A specialized Therapeutic Quality Evaluator agent assesses the generated candidates[cite: 251].

• Mechanism: It acts as a clinical supervisor, scoring reframes on a scale of 1-5 based on Empathy, Actionability, and Safety.
• Selection Logic: $$R_{final} = \operatorname*{argmax}{R \in {R{CBT}, R_{ACT}, ...}} (\text{Score}_{supervisor}(R))$$

---

⛓️ Advanced Technique: Implicit Chain-of-Thought (CoT)

To prevent the model from generating superficial advice, we employ Implicit Chain-of-Thought prompting. We instruct the model to perform a "hidden" analysis phase before outputting the final response.

Prompt Template Structure:

Phase 1: Analysis (Internal Monologue)
- Identify the underlying emotion.
- Identify the specific cognitive trap.
- Select the appropriate technique from the [Therapy Name] manual.

Phase 2: Drafting (Final Output)
- Draft the response applying the technique.
- Ensure the tone is validating and non-judgmental.

Output ONLY Phase 2.