Sashrika Pandey and Zhiyang He and Mariah Schrum and Anca Dragan

EECS Department, University of California, Berkeley

Technical Report No. UCB/EECS-2024-99

May 13, 2024

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2024/EECS-2024-99.pdf

When querying a large language model (LLM), the context, i.e. personal, demographic, and cultural information specific to an end-user, can significantly shape the response of the LLM. For example, asking the model to explain Newton's second law with the context "I am a toddler" yields a different answer compared to the context "I am a physics professor." Proper usage of the context enables the LLM to generate personalized responses, whereas inappropriate contextual influence can lead to stereotypical and potentially harmful generations (e.g. associating "female" with "housekeeper"). In practice, striking the right balance when leveraging context is a nuanced and challenging problem that is often situation-dependent. One common approach to address this challenge is to fine-tune LLMs on contextually appropriate responses. However, this approach is expensive, time-consuming, and not controllable for end-users in different situations. In this work, we propose Context Steering (CoS) - a simple training-free method that can be easily applied to autoregressive LLMs at inference time. By measuring the contextual influence in terms of token prediction likelihood and modulating it, our method enables practitioners to determine the appropriate level of contextual influence based on their specific use case and end-user base. We showcase a variety of applications of CoS including amplifying the contextual influence to achieve better personalization and mitigating unwanted influence for reducing model bias. In addition, we show that we can combine CoS with Bayesian Inference to quantify the extent of hate speech on the internet. We demonstrate the effectiveness of CoS on state-of-the-art LLMs and benchmarks.

Advisors: Anca Dragan

---

BibTeX citation:

@mastersthesis{Pandey:EECS-2024-99,
    Author= {Pandey, Sashrika and He, Zhiyang and Schrum, Mariah and Dragan, Anca},
    Title= {CoS: Enhancing Personalization and Mitigating Bias with Context Steering},
    School= {EECS Department, University of California, Berkeley},
    Year= {2024},
    Month= {May},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2024/EECS-2024-99.html},
    Number= {UCB/EECS-2024-99},
    Abstract= {When querying a large language model (LLM), the context, i.e. personal, demographic, and cultural information specific to an end-user, can significantly shape the response of the LLM. For example, asking the model to explain Newton's second law with the context "I am a toddler" yields a different answer compared to the context "I am a physics professor." Proper usage of the context enables the LLM to generate personalized responses, whereas inappropriate contextual influence can lead to stereotypical and potentially harmful generations (e.g. associating "female" with "housekeeper"). In practice, striking the right balance when leveraging context is a nuanced and challenging problem that is often situation-dependent. One common approach to address this challenge is to fine-tune LLMs on contextually appropriate responses. However, this approach is expensive, time-consuming, and not controllable for end-users in different situations. In this work, we propose Context Steering (CoS) - a simple training-free method that can be easily applied to autoregressive LLMs at inference time. By measuring the contextual influence in terms of token prediction likelihood and modulating it, our method enables practitioners to determine the appropriate level of contextual influence based on their specific use case and end-user base. We showcase a variety of applications of CoS including amplifying the contextual influence to achieve better personalization and mitigating unwanted influence for reducing model bias. In addition, we show that we can combine CoS with Bayesian Inference to quantify the extent of hate speech on the internet. We demonstrate the effectiveness of CoS on state-of-the-art LLMs and benchmarks.},
}

EndNote citation:

%0 Thesis
%A Pandey, Sashrika 
%A He, Zhiyang 
%A Schrum, Mariah 
%A Dragan, Anca 
%T CoS: Enhancing Personalization and Mitigating Bias with Context Steering
%I EECS Department, University of California, Berkeley
%D 2024
%8 May 13
%@ UCB/EECS-2024-99
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2024/EECS-2024-99.html
%F Pandey:EECS-2024-99