Instructions.md

MiCRo Expert Routing Visualizer (Gradio)

This demo visualizes how modular language models allocate computation across specialized experts—Language, Logic, Social, and World—when processing a given prompt. Each expert corresponds to a cognitive domain inspired by human brain networks. Enter a prompt to see how tokens are dynamically routed across modules, revealing the model’s internal reasoning structure.

How it works

• Choose a model (dropdown) or type a custom model id.

• Enter a User prompt. Optionally add an Assistant prompt; if provided, the app concatenates them as:

  User: <user text>
  Assistant: <assistant text>
  

• When the prompt fails, the demo falls back to "mock data", which generates deterministic, pseudo-random percentages from the prompt.

Backend contract

get_expert_routing(model_id: str, prompt: str) must return 4 values (percentages) for the experts in this fixed order: ["Language", "Logic", "Social", "World"] or a dict with those exact keys.