P20_react.md

P20 — ReAct

Field	Value
ID	P20
Name	ReAct
Category	Reasoning Augmentation
Origin	Yao et al. (2022)
PromptPrism Level	Structural / Semantic
Volatility Class	III — Model-Specific
Transfer	0.59 [.53,.66]
Restricted Transfer	0.74 [.67,.80]
PSI	0.21
Core Mechanism	Interleaved reasoning and action

Per-Model PTS

Model	PTS	±CI
GPT-4o (M1)	0.81	±.03
Claude 3 Opus (M2)	0.79	±.04
Gemini 1.5 Pro (M3)	0.77	±.04
LLaMA-3-70B-Instruct (M4)	0.69	±.05
Mistral-7B-Instruct (M5)	0.60	±.07
MedLLaMA-2-7B (M6)	0.48	±.08

Canonical Specification (V-A)

Solve [TASK] using the ReAct format. Alternate between Thought and Action steps:
Thought: [reason about current state]
Action: [specify what you would do or look up]
Observation: [what the action yields]
...repeat until...
Final Answer: [conclusion]

Deployment Guidance

The Thought-Action-Observation loop requires schema maintenance that degrades significantly in 7B models (Mistral: 0.60, near Class III boundary). MedLLaMA-2 approaches Class IV (0.48). Reliable only on frontier and 70B+ models. Best suited for agentic task categories not included in this evaluation set.