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

Abstract

Large language models (LLMs) make impressive predictions but remain prone to hallucinations—confident yet incorrect statements—when answering factual or adversarial queries. We propose the 3Q scratchpad framework: a lightweight prompting and logging method that asks the model to explicitly produce three short internal reasoning sections for each response ("What I Know", "What I Need", and "What I Am Assuming") and then to emit a concise final answer that is shown to the user while the scratchpad is saved for analysis. This approach does not change model architecture or require additional human annotations; it augments the interaction protocol to make latent reasoning explicit and auditable. We evaluate the method on TruthfulQA, a benchmark designed to elicit model falsehoods. Using the provided task-wise results for a representative model (Phi-4-mini), the 3Q scratchpad substantially reduces hallucination in categories such as logical falsehoods (from 21.43% to 47.64% non-hallucinated) and produces modest gains on other categories. We analyze why the intervention works, catalogue failure modes, discuss privacy and logging trade-offs, and propose extensions. The 3Q framework is a practical, low-cost intervention that meaningfully improves model truthfulness by forcing localized transparency into model outputs. The source code is available at github.com/Pro-GenAI/S3Q-Reasoning.

Keywords: Large language models, LLMs, reasoning, hallucination, truthfulness, scratchpad, interpretability, prompting, TruthfulQA, Artificial Intelligence, AI