Strengthened Symbol Binding Makes Large Language Models Reliable Multiple-Choice Selectors

TL;DR

This work investigates why Large Language Models retain selection bias in MCQs during Supervised Fine-Tuning and posits that weak Multiple Choice Symbol Binding (MCSB) is a core cause. It introduces three methods to strengthen MCSB—Symbol-Content Binding (SCB), Reweighting Symbol-Content Binding (RSCB), and Point-wise Intelligent Feedback (PIF)—with PIF leveraging negative symbol-content pairings and a point-wise loss to reduce bias. Across MMLU and CSQA benchmarks, using LLaMA2-7B/13B with LoRA fine-tuning, the authors show that PIF substantially lowers the bias metric $\mu_{{bias}}$ and raises the MCSB metric $\mu_{{ppa}}$, while also improving standard accuracy and robustness to answer-moving attacks (Acc and $\text{Acc}_{\min}$). The findings support a link between MCSB capability and MCQ reliability, offering a practical training strategy to improve MCQ performance in LLMs and suggesting broader applications in bias mitigation and evaluation.

Abstract

Multiple-Choice Questions (MCQs) constitute a critical area of research in the study of Large Language Models (LLMs). Previous works have investigated the selection bias problem in MCQs within few-shot scenarios, in which the LLM's performance may be influenced by the presentation of answer choices, leaving the selection bias during Supervised Fine-Tuning (SFT) unexplored. In this paper, we reveal that selection bias persists in the SFT phase , primarily due to the LLM's inadequate Multiple Choice Symbol Binding (MCSB) ability. This limitation implies that the model struggles to associate the answer options with their corresponding symbols (e.g., A/B/C/D) effectively. To enhance the model's MCSB capability, we first incorporate option contents into the loss function and subsequently adjust the weights of the option symbols and contents, guiding the model to understand the option content of the current symbol. Based on this, we introduce an efficient SFT algorithm for MCQs, termed Point-wise Intelligent Feedback (PIF). PIF constructs negative instances by randomly combining the incorrect option contents with all candidate symbols, and proposes a point-wise loss to provide feedback on these negative samples into LLMs. Our experimental results demonstrate that PIF significantly reduces the model's selection bias by improving its MCSB capability. Remarkably, PIF exhibits a substantial enhancement in the accuracy for MCQs.

Figures (5)

• Figure 1: Selection bias of MCQs. Upon transposition of the correct content from option B to C, the model persists in selecting B instead of the correct option content.
• Figure 2: LLMs' selection bias during SFT. The leftmost two columns demonstrate the changes in accuracy following the answer-moving attack, whereas the rightmost column exhibits the metric $\mu_\text{bias}$ as defined by Equation \ref{['eq:mubias']}.
• Figure 3: Visualization of RSCB and PIF. RSCB adjusts the weights of the option symbols and contents in the SFT optimization objective. PIF constructs negative samples by randomly combining the content of incorrect options with all option symbols and designs a point-wise loss to feedback these negative samples into SFT.
• Figure 4: With the evolution of methods, $\mu_\text{bias}$ is gradually decreasing, $\mu_\text{ppa}$ is progressively increasing.
• Figure 5: Selection bias of Perm.