Lune

ACL Findings2025

Anchored Answers: Unravelling Positional Bias in GPT-2's Multiple-Choice Questions

Ruizhe Li, Yanjun Gao

Abstract

Large Language Models (LLMs), such as the GPT-4 and LLaMA families, have demonstrated considerable success across diverse tasks, including multiple-choice questions (MCQs). However, these models exhibit a positional bias, particularly an even worse "anchored bias" in the GPT-2 family, where they consistently favour the first choice 'A' in MCQs during inference. This anchored bias challenges the integrity of GPT-2's decisionmaking process, as it skews performance based on the position rather than the content of the choices in MCQs. In this study, we utilise the mechanistic interpretability approach to identify the internal modules within GPT-2 models responsible for this bias. We focus on the Multi-Layer Perceptron (MLP) layers and attention heads, using the "logit lens" method to trace and modify the specific value vectors that contribute to the bias. By updating these vectors within MLP and recalibrating attention patterns to neutralise the preference for the first choice 'A', we effectively mitigate the anchored bias. Our interventions not only mitigate the bias but also improve the overall MCQ prediction accuracy for the GPT-2 family across various datasets. This work represents the first comprehensive mechanistic analysis of anchored bias from the failing cases in MCQs within the GPT-2 models, introducing targeted, minimal-intervention strategies that significantly enhance GPT2 model robustness and accuracy in MCQs. Our code is available at https://github.com/ruizheliUOA/ Anchored_Bias_GPT2.

Ask Lune

Know this paper in minutes, not an afternoon.

Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.

9 top-tier papers have cited it since. Know what they found before your reviewers do.

Questions to start from

Lune can draw on

the full paper · the top-tier papers that cite it · related work

Get the answer

Free to start. No credit card required.

Lineage

  • Talking Heads: Understanding Inter-Layer Communication in Transformer Language Models cites Circuit Component Reuse Across Tasks in Transformer Language Models
  • Talking Heads: Understanding Inter-Layer Communication in Transformer Language Models cites How does GPT-2 compute greater-than?: Interpreting mathematical abilities in a pre-trained language model
  • Which of These Best Describes Multiple Choice Evaluation with LLMs? A) Forced B) Flawed C) Fixable D) All of the Above cites Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting
  • MIB: A Mechanistic Interpretability Benchmark cites Locating and Editing Factual Associations in GPT
  • MIB: A Mechanistic Interpretability Benchmark cites Circuit Component Reuse Across Tasks in Transformer Language Models
  • Attributing Response to Context: A Jensen–Shannon Divergence Driven Mechanistic Study of Context Attribution in Retrieval-Augmented Generation cites Locating and Editing Factual Associations in GPT
  • Attributing Response to Context: A Jensen–Shannon Divergence Driven Mechanistic Study of Context Attribution in Retrieval-Augmented Generation cites Transformer Feed-Forward Layers Build Predictions by Promoting Concepts in the Vocabulary Space
  • How does GPT-2 compute greater-than?: Interpreting mathematical abilities in a pre-trained language model cites Locating and Editing Factual Associations in GPT
Builds onCited byLeft to right by year; the more cited, the higher.

Cited by top-tier papers9

  1. Talking Heads: Understanding Inter-Layer Communication in Transformer Language ModelsJack Merullo, Carsten Eickhoff, Ellie PavlickNeurIPS 2024 · 49 citations
  2. Attributing Response to Context: A Jensen–Shannon Divergence Driven Mechanistic Study of Context Attribution in Retrieval-Augmented GenerationRuizhe Li, Chen Chen, Yuchen Hu, Yanjun Gao et al.ICLR 2026 · 11 citations
  3. Positional Overload: Positional Debiasing and Context Window Extension for Large Language Models using Set EncodingLukas Kinder, Lukas Edman, Alexander Fraser, Tobias KäferACL 2025
  4. Which of These Best Describes Multiple Choice Evaluation with LLMs? A) Forced B) Flawed C) Fixable D) All of the AboveNishant Balepur, Rachel Rudinger, Jordan Lee Boyd-GraberACL 2025
  5. MIB: A Mechanistic Interpretability BenchmarkAaron Mueller, Atticus Geiger, Sarah Wiegreffe, Dana Arad et al.ICML 2025

Builds on16

  1. Locating and Editing Factual Associations in GPTKevin Meng, David Bau, Alex Andonian, Yonatan BelinkovNeurIPS 2022 · 3,415 citations
  2. Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought PromptingMiles Turpin, Julian Michael, Ethan Perez, Samuel R. BowmanNeurIPS 2023 · 1,792 citations
  3. How does GPT-2 compute greater-than?: Interpreting mathematical abilities in a pre-trained language modelMichael Hanna, Ollie Liu, Alexandre VariengienNeurIPS 2023 · 251 citations
  4. Circuit Component Reuse Across Tasks in Transformer Language ModelsJack Merullo, Carsten Eickhoff, Ellie PavlickICLR 2024 · 108 citations
  5. Transformer Feed-Forward Layers Build Predictions by Promoting Concepts in the Vocabulary SpaceMor Geva, Avi Caciularu, Kevin Ro Wang, Yoav GoldbergEMNLP 2022 · 92 citations

Related papers

  1. Large Language Models Are Not Robust Multiple Choice SelectorsChujie Zheng, Hao Zhou, Fandong Meng, Jie Zhou et al.ICLR 2024 · 424 citations
  2. Benchmarking and Mitigating MCQA Selection Bias of Large Vision-Language ModelsMd. Atabuzzaman, Ali Asgarov, Christopher ThomasEMNLP 2025
  3. Exploring the Choice Behavior of Large Language ModelsWeidong Wu, Qinlin Zhao, Hao Chen, Lexin Zhou et al.ACL Findings 2025
  4. Characterizing Positional Bias in Large Language Models: A Multi-Model Evaluation of Prompt Order EffectsPatrick Schilcher, Dominik Karasin, Michael Schöpf, Haisam Saleh et al.EMNLP Findings 2025 · 1 citation
  5. LLM Agents Can Be Choice-Supportive Biased Evaluators: An Empirical StudyNan Zhuang, Boyu Cao, Yi Yang, Jing Xu et al.AAAI 2025 · 4 citations
Dusk over the sea between two cliffs drawn in fine vertical lines