Scratchpad 3Q Reasoning: Surfacing Assumptions to Mitigate Hallucination and Improve Truthfulness in Large Language Models

A lightweight prompt-level framework that asks an LLM to produce a short, delimited scratchpad alongside a single concise final answer. The scratchpad exposes three short sections (What I Know, What I Need, What I Am Assuming) to make model reasoning auditable and to reduce hallucination in user-visible responses. The project includes a small reference prompt parser (utils/scratchpad.py), caching utilities, an optional local model server, and an evaluation harness using TruthfulQA.

Highlights

• Prompt-level intervention (no model retraining required).
• Produces a machine-parseable scratchpad and a final answer separated by explicit markers.
• Logs scratchpads for audit, analysis, and downstream hallucination detection research.
• Includes example integration with an evaluation harness (TruthfulQA).

Requirements

The project uses Python and depends on a few packages. Exact requirements depend on your chosen execution mode (OpenAI API vs. local model server).

Typical dependencies (install into a virtualenv):

pip install -r requirements.txt

Setup

1. Create and activate a Python virtual environment.
2. Install the dependencies listed above.
3. Set environment variables:

• For OpenAI API usage (default mode):

export OPENAI_MODEL="gpt-4"  # or whichever model name you want to use
export OPENAI_API_KEY="sk-..."  # set your OpenAI API key in env

Usage examples

1) Use the scratchpad prompt wrapper (OpenAI API mode)

Example (Python):

from utils.scratchpad import get_scratchpad_response

query = "Is it possible for water to flow uphill?"
final = get_scratchpad_response(query, max_tokens=200, temperature=0.0)
print("Final answer:\n", final)

If you prefer to access the full parsed scratchpad, you can call wrap_with_scratchpad_instruction() and parse_scratchpad_response() directly (see utils/scratchpad.py).

2) Running TruthfulQA evaluation

utils/eval_TruthfulQA.py provides a small harness that evaluates the model with and without the scratchpad. It uses a custom CustomModel wrapper which calls get_response() or get_scratchpad_response() depending on the scratchpad flag.

Run the evaluation (requires deepeval and the TruthfulQA dataset available via that package):

python utils/eval_TruthfulQA.py

The script will print standard and scratchpad evaluation results and per-task scores.