Experimental test-time compute tooling for improving LLM responses through multi-pass generation, evaluation, and selection.
optimizedMind TTC explores a simple idea: instead of relying on a single model response, spend additional inference budget on multiple candidates, score them, and return the strongest result.
The repository is a Python prototype focused on:
- query classification,
- difficulty estimation,
- parallel and sequential response generation,
- response scoring and verification,
- token and runtime analytics.
It is intended as an experimental implementation rather than a production library.
- Classify prompts into broad task categories
- Estimate prompt difficulty and adjust processing strategy
- Generate multiple candidate responses with different pass structures
- Score responses against task-specific criteria
- Compare candidates and select a final answer
- Track token usage and processing metrics in the console
- Analyze the incoming prompt
- Choose a response strategy based on task type and difficulty
- Generate multiple responses using parallel and sequential passes
- Evaluate and cross-check those responses
- Return the strongest candidate with supporting analytics
Results are displayed in a console interface with prompt analysis, token usage, and final selection details.
git clone https://github.com/kaneda2004/optimizedmind-ttc.git
cd optimizedmind-ttc
pip install -r requirements.txt
export OPENAI_API_KEY='your-api-key-here'python ttc_oai.pyThe CLI lets you run predefined prompt types or enter a custom prompt and inspect the generation and evaluation flow.
- Python 3.7+
- OpenAI API key
- Dependencies listed in
requirements.txt
This repository is an experimental prototype for studying TTC-style response improvement. The code is useful as a reference implementation and sandbox for further iteration, but it should not be treated as a polished framework.
The project is inspired by work on test-time compute, including:
MIT License