Investigating Anchoring Bias and Prompt-Based Mitigation in Large Language Models

Large language models are being increasingly integrated into tasks involving numerical estimation and decision-making. However, due to their training on human-generated text, they may be susceptible to various cognitive biases. This includes anchoring bias, a well-documented phenomenon in humans where judgements are illogically influenced by an irrelevant reference number.

This study investigated whether three large language models (Mistral-Large-3, DeepSeek-V3.1, and Meta’s Llama-3.1-8B) exhibit anchoring bias when prompted with general-knowledge monetary estimation questions. We also tested whether two prompt-based mitigation strategies (explicit debias and estimate-first) can reduce the anchoring effect.

We hypothesized that larger models would show less susceptibility to anchoring bias as compared to smaller models and also respond better to mitigation strategies. Using a set of twenty-four monetary prompts, each with a high and low anchor variation, we quantified anchoring bias as the normalized percent difference between high and low anchor responses and determined statistical significance with paired t-tests.

We found that mean anchoring bias magnitudes differed across models, with Llama-3.1-8B showing the largest effect and Mistral-Large-3 the smallest. However, with high prompt-level variability, paired t-tests indicated that baseline differences were not statistically significant, despite the consistent directionality of responses. The effectiveness of mitigation strategies also differed across models, with explicit debias prompting significantly reducing anchoring bias for Mistral-Large-3, while having little effect on other models, and estimate-first prompting significantly reducing bias only for DeepSeek-V3.1.

These findings suggest that anchoring bias can influence LLM outputs but is inconsistent, while the effectiveness of prompt-based mitigation is model-dependent.