TLDR: A new study reveals that Large Language Models (LLMs) exhibit a ‘primacy effect,’ favoring options presented first, a bias amplified by fine-tuning. Researchers propose a novel, training-free method that reorders multiple-choice answer options based on their semantic similarity to the query. By placing the most relevant options first, this approach strategically exploits the LLM’s inherent bias, leading to significant improvements in accuracy across various models and datasets without needing prior knowledge of the correct answer.
Large Language Models (LLMs) have become indispensable tools in various Natural Language Processing (NLP) tasks, demonstrating remarkable accuracy through extensive pre-training and fine-tuning. However, much like humans, these advanced AI models can exhibit certain cognitive biases, particularly positional biases such as the primacy and recency effects.
The primacy effect, a key focus of a recent study, describes the tendency for items presented first to be more readily remembered or selected. In the context of Multiple Choice Question Answering (MCQA), this means that the order in which answer options are presented can significantly influence an LLM’s prediction outcomes.
Understanding the Primacy Bias in LLMs
Researchers Bianca Raimondi and Maurizio Gabbrielli from the University of Bologna, Italy, delved into this primacy bias, especially in fine-tuned LLMs. Their findings indicate that the process of fine-tuning, which exposes LLMs to human-like patterns, actually amplifies this positional bias. This means that models trained with specific instructions or human feedback tend to show an even stronger preference for options appearing early in a list.
Traditionally, such biases are viewed as limitations that need to be mitigated. However, this study takes a novel approach: instead of trying to eliminate the bias, it strategically leverages it to enhance performance.
A Smart Reordering Strategy
The core of their proposed technique involves reordering response options based on their semantic similarity to the original query. The intuition is straightforward: if an LLM is more likely to select options presented first, then placing the most semantically relevant candidates at the beginning of the list can guide the model toward more accurate predictions. This method is particularly innovative because it doesn’t require prior knowledge of the correct answer, making it applicable even in scenarios with unlabeled data.
The reordering process involves computing the mean cosine similarity between the embeddings (numerical representations) of the query and each answer option. Options are then ranked in descending order of similarity, ensuring that those most closely related to the query appear first. While this doesn’t guarantee the correct answer is always at the very top, it significantly moves relevant options closer to the beginning of the list, increasing their chances of being selected.
Significant Performance Improvements
The experimental results are compelling. The researchers tested their approach on several LLMs, including versions of Llama and Mistral, across different MCQA datasets like CLINC, BANKING, and HWU. They consistently observed that this reordering strategy significantly improved the models’ accuracy, especially in fine-tuned versions where the primacy bias was more pronounced.
For instance, on the CLINC dataset, the “Sort” technique (their reordering method) notably increased model accuracy compared to the “NoSort” baseline. The study also revealed that the primacy effect intensifies as the number of answer options in the prompt increases, further highlighting the value of their reordering method in complex scenarios.
This research underscores the dual nature of biases in AI: they can be both challenges and opportunities. By embracing and exploiting the primacy effect, this study offers valuable insights for designing more bias-aware models and improving NLP applications. For more details, you can read the full research paper: Exploiting Primacy Effect To Improve Large Language Models.
Also Read:
- Navigating the Ideaverse: How AI Explores Latent Spaces for Breakthrough Creativity
- DUALRec: A New Approach to Movie Recommendations with AI
Future Directions
The authors suggest future work could involve refining this method, extending it to other types of biases (like emotional biases), and exploring adaptive reordering techniques. While the approach is robust, they acknowledge limitations such as potential variations across LLM architectures and the assumption that embedding relevance aligns perfectly with semantic correctness.
