15 : 12 : 34 : 30

15 : 12 : 34 : 30

15 : 12 : 34 : 30

15 : 12 : 34 : 30

Keep Apart Research Going: Donate Today

Jun 2, 2025

Manipulating Self-Preference for Large Language Models

Matthew Nguyen, Simon Fu, Matthew Bozoukov, Dani Roytburg, Jou Barzdukas

Details

Details

Arrow
Arrow
Arrow
Arrow
Arrow
Arrow

Large language models (LLMs) carry great value as evaluators of synthetic data for research and production settings. However, recent research shows that language models exhibit bias towards their own responses in blind model-to-model evaluation settings. Self-preference bias shows clear negative effects on effective Judge Model Development, our chosen research track. We first identify instances of this phenomenon at the behavioral level, robustly demonstrating that two models – Llama 3.1-8b-Instruct and DeepSeek V3 – exhibit this behavior. Then, focusing on Llama, we construct a steering vector from the residual stream of the

model that represents the ``self-preference'' direction. Then, we demonstrate that the vector causally impacts a model’s ability to assert self-preference by applying the vector to the model’s output as it generates it (steering). When steered in the positive self-preference direction , we find that the model asserts self-preference on 85\% of the examples where it previously did not. Conversely, when steered in the negative direction, we find that the model asserts non-self-preference on 25\% of the examples where it previously did not. These findings suggest potential approaches to de-biasing expert orchestrators such as judges and routers, potentially enabling a fair allocation of responses. Further research is necessary to scale this approach to larger models, as well as to determine the impact on orchestration systems in production. This aligns with the Judge Model Development track because the prevalence of self-preference in judges is not negligible especially when it is compared to responses from other models (not humans). One of the key components of the Expert Orchestration Architecture is the judges, which help the router make a better educated decision on which model to direct the query to. One of the main criteria users care about is bias. We show that this bias exists in a specific judge model when given a choice between another model's response and further mitigate this.

Cite this work:

@misc {

title={

},

author={

Matthew Nguyen, Simon Fu, Matthew Bozoukov, Dani Roytburg, Jou Barzdukas

},

date={

6/2/25

},

organization={Apart Research},

note={Research submission to the research sprint hosted by Apart.},

howpublished={https://apartresearch.com}

}

Reviewer's Comments

Reviewer's Comments

Arrow
Arrow
Arrow
Arrow
Arrow

No reviews are available yet

Apr 14, 2025

Read More

Jan 24, 2025

Safe ai

The rapid adoption of AI in critical industries like healthcare and legal services has highlighted the urgent need for robust risk mitigation mechanisms. While domain-specific AI agents offer efficiency, they often lack transparency and accountability, raising concerns about safety, reliability, and compliance. The stakes are high, as AI failures in these sectors can lead to catastrophic outcomes, including loss of life, legal repercussions, and significant financial and reputational damage. Current solutions, such as regulatory frameworks and quality assurance protocols, provide only partial protection against the multifaceted risks associated with AI deployment. This situation underscores the necessity for an innovative approach that combines comprehensive risk assessment with financial safeguards to ensure the responsible and secure implementation of AI technologies across high-stakes industries.

Read More

Jan 24, 2025

CoTEP: A Multi-Modal Chain of Thought Evaluation Platform for the Next Generation of SOTA AI Models

As advanced state-of-the-art models like OpenAI's o-1 series, the upcoming o-3 family, Gemini 2.0 Flash Thinking and DeepSeek display increasingly sophisticated chain-of-thought (CoT) capabilities, our safety evaluations have not yet caught up. We propose building a platform that allows us to gather systematic evaluations of AI reasoning processes to create comprehensive safety benchmarks. Our Chain of Thought Evaluation Platform (CoTEP) will help establish standards for assessing AI reasoning and ensure development of more robust, trustworthy AI systems through industry and government collaboration.

Read More

This work was done during one weekend by research workshop participants and does not represent the work of Apart Research.
This work was done during one weekend by research workshop participants and does not represent the work of Apart Research.