Abstract
Large Language Models (LLMs) often struggle with hallucinations when multiple questions are aggregated into a single prompt—a common practice known as batch prompting. This paper introduces Multiple Answers and Confidence Stepwise Tuning (MAC-Tuning), a framework designed to improve a model's awareness of its own knowledge boundaries. By employing a two-step fine-tuning process that separates answer generation from linguistic confidence estimation, MAC-Tuning significantly reduces reasoning confusion. Our results show up to a 25% improvement in average precision and superior calibration compared to standard fine-tuning methods.
Introduction
The prevailing paradigm for evaluating LLM hallucinations focuses on the "single-problem setting," where a model responds to one question at a time. However, in production environments, efficiency and cost-optimization often drive the use of "multi-problem settings" or batch prompts. The researchers identify a critical phenomenon in these scenarios: reasoning confusion. When multiple questions are packed into a single context, the reasoning required for one sub-question can interfere with another, causing the model to confidently output false information that it might have correctly identified as unknown if queried in isolation.
Methodology: MAC-Tuning
To mitigate these errors, the authors propose MAC-Tuning, which teaches models to recognize the limits of their parametric memory. The process is divided into three distinct steps:
1. Multi-Problem Dataset Construction
The researchers synthesized multi-question prompts by combining individual QA pairs from diverse datasets, including CoQA (conversational), GSM (math), and MMLU (general knowledge). This creates a training environment that mirrors real-world batch prompting.
2. Knowledge Boundary Labeling
Instead of using static confidence scores, the framework uses the model's own performance to define its "knowledge boundary." Prompts are passed through a base model, and outputs are compared to ground-truth answers. Correct responses are labeled "I am sure," while incorrect ones are labeled "I am unsure."
3. Stepwise Fine-Tuning
A key contribution of this work is the discovery that "Stepwise Tuning" is superior to "Merged Tuning." In Merged Tuning, the model attempts to learn answers and confidence levels simultaneously, which often leads to task interference. MAC-Tuning employs a two-stage Supervised Fine-Tuning (SFT) approach:
- Stage 1: The model is trained to predict the set of multiple answers ($A$) for the given questions ($Q$).
- Stage 2: The model is trained to express its confidence ($C$) for those specific answers in a linguistic format (e.g., "1: I am sure... 3: I am unsure").
Experimental Results and Analysis
The framework was evaluated across six datasets categorized into independent and sequential settings.
- Precision Gains: MAC-Tuning achieved an Average Precision (AP) improvement of up to 25% over vanilla base models such as LLaMA-3-8B and Qwen2.
- Calibration and Honesty: The method significantly reduced the Expected Calibration Error (ECE). In human evaluations, there was a 48-percentage-point gap in accuracy between "sure" and "unsure" labels, demonstrating that the model became much more "honest" about its uncertainty.
- Robustness to Complexity: While base models typically show a performance decay as the number of questions ($n$) increases, MAC-Tuning remained effective even when scaled to 15 questions per prompt.
Conclusion
MAC-Tuning addresses a significant gap in LLM reliability for complex, multi-compositional tasks. By separating the tasks of information retrieval and self-reflection, the method enables models to better navigate their internal knowledge boundaries, leading to more trustworthy AI behavior in high-throughput applications.
Limitations: The authors note that while the method generalizes well, it remains sensitive to prompt phrasing and requires significant computational resources for fine-tuning. Future work will explore mixing question difficulties and broader generalization to open-ended scenarios.
Summary updated