Preference for Additive Over Subtractive Solution Strategies in Humans and the GPT-4 Language Model

Core Concepts

Humans and the GPT-4 language model exhibit a systematic bias towards additive solution strategies over subtractive ones, even when subtraction is more efficient.

Abstract

This study explored the addition bias, a cognitive tendency to prefer adding elements over removing them to alter an initial state or structure, by conducting four preregistered experiments examining the problem-solving behavior of both humans and OpenAI's GPT-4 large language model. The key findings are: Overall, additive solution strategies were chosen significantly more often than subtractive strategies, indicating an addition bias in both humans and GPT-4. The influence of solution efficiency on the addition bias worked in opposite directions for humans and GPT-4. Humans were more likely to choose additive strategies when addition and subtraction were equally efficient, but GPT-4 showed the opposite pattern, favoring additive strategies more when subtraction was more efficient. The valence of the instruction (neutral vs. positive) affected the addition bias in GPT-4 in the linguistic task, with more additive strategies used when the instruction was to "improve" rather than "edit". Humans were not influenced by the valence of the instruction. GPT-4 exhibited a stronger addition bias compared to humans across the different conditions, suggesting the language model may magnify certain biases. The findings highlight the importance of considering subtractive solutions and carefully evaluating the outputs of large language models, as they may exhibit biases that differ from human problem-solving tendencies.

Stats

"Additive solution strategies (64.4%; 816/1268) were observed more frequently than subtractive strategies (35.6%; 452/1268)." "Humans being relatively less likely to choose additive strategies (57.0%; 335/588) compared to GPT-4 (70.7%, 481/680)."

Quotes

"Additive ideas have the potential to enhance recognition by others, e.g., policy initiatives to "plant a billion trees" might sound more appealing than "allowing forests to regenerate"." "If the decision is made to mitigate the effects of the addition bias replicated in this study, several strategies can be employed. One way to mitigate bias in decision-making is to increase awareness of its potential impact."

Key Insights Distilled From

Influence of Solution Efficiency and Valence of Instruction on Additive and Subtractive Solution Strategies in Humans and GPT-4

by Lydi... at arxiv.org 04-26-2024

https://arxiv.org/pdf/2404.16692.pdf

Deeper Inquiries

What cognitive processes or heuristics underlie the addition bias in both humans and language models?

The addition bias observed in both humans and language models can be attributed to several cognitive processes and heuristics. One key factor is the tendency to prioritize additive solutions over subtractive ones, which can be linked to the "what can we add here?" heuristic. This heuristic simplifies decision-making by focusing on adding elements rather than removing them. Additionally, cognitive biases such as the base rate fallacy, sunk cost fallacy, and loss aversion may also contribute to the preference for additive solutions. These biases overlook probabilities, investments, and the aversion to losses, respectively, leading individuals to default to additive strategies. In the case of language models, the addition bias can be a result of the training data and algorithms used during the model's development. Language models like GPT-4 are pre-trained on vast amounts of human-generated text, which may contain inherent biases present in human language. The model's learning process and the subjective judgments of human annotators during training can further reinforce these biases. As a result, the language model may replicate and amplify the addition bias observed in human decision-making processes.

How can the tendency towards additive solutions be reduced in practical problem-solving scenarios, beyond just increasing awareness?

Reducing the tendency towards additive solutions in practical problem-solving scenarios requires a multi-faceted approach beyond just increasing awareness. One strategy is to actively prompt for and consider subtractive solutions during problem-solving tasks. By explicitly requesting the exploration of subtractive options, individuals can overcome the default "additive" heuristic and broaden their problem-solving strategies. Another approach is to provide examples and training on efficient subtractive strategies. By familiarizing individuals with the benefits and effectiveness of subtractive solutions, they can develop a more balanced approach to problem-solving. Additionally, incorporating feedback mechanisms that highlight the efficiency and success of subtractive solutions can reinforce the value of considering both additive and subtractive strategies. In the case of language models, developers can fine-tune the models on datasets that emphasize balanced solutions and provide diverse examples of both additive and subtractive transformations. By training the model to recognize and generate subtractive solutions effectively, the tendency towards additive bias can be mitigated in the model's outputs.

What other types of biases might be present in the outputs of large language models, and how can they be systematically identified and addressed?

In addition to the addition bias, outputs of large language models may exhibit various other biases, such as confirmation bias, availability bias, and stereotype bias. Confirmation bias can lead to the reinforcement of existing beliefs or preferences in the model's outputs. Availability bias may result in the overrepresentation of easily accessible information in the generated text. Stereotype bias can manifest as the perpetuation of societal stereotypes and prejudices in the language model's responses. To systematically identify and address these biases in the outputs of large language models, several approaches can be employed. One method is to conduct bias audits and evaluations of the model's outputs using diverse datasets and evaluation metrics. By analyzing the model's responses across different demographic groups and contexts, researchers can identify and quantify the presence of biases. Furthermore, bias mitigation techniques such as debiasing algorithms, adversarial training, and dataset augmentation can be implemented to reduce biases in the language model's outputs. These techniques involve modifying the training data, adjusting the model's architecture, and incorporating fairness constraints to promote more equitable and unbiased responses. Regular monitoring and auditing of the model's performance can help track the effectiveness of these mitigation strategies and ensure ongoing bias reduction.