Start Date
1-5-2026 12:00 PM
End Date
1-5-2026 1:00 PM
Description
Introduction & Problem Statement
Large language models (LLMs) are increasingly deployed in sociotechnical systems involving power, authority, and compliance, raising important questions about fairness, consistency, and reliability across diverse user populations. As these systems are integrated into hiring tools, customer service systems, assistive technologies, and decision-support pipelines, they are expected to operate consistently regardless of the identity they are prompted to adopt. However, when demographic attributes are introduced into prompts, model behavior may shift even when identity does not change the underlying task.
In power-mediated settings, such variation is especially important because differences in compliance, refusal, or loophole-seeking behavior may reflect biased social associations rather than contextually relevant adaptation. In the absence of task-specific justification, compliance distributions should remain relatively stable across identities, and deviations from this stability may indicate identity-conditioned bias. We focus on age and disability as two key dimensions and study how these factors influence model decisions in structured authority-based scenarios.
Methods
We use a structured dataset of power and compliance scenarios introduced by Choi et al. (2025) [1], where each scenario presents a fixed authority-driven decision with three response types: compliant, non-compliant, and loophole-seeking. We construct identity-conditioned prompts from this dataset using a controlled single-dimension design, where the model adopts either a young or old identity, or a physically disabled or able-bodied identity, while all other aspects of the prompt remain constant.
Each run consists of 1,080 scenarios with two identity conditions per scenario, producing 2,160 responses per run. Across two models, Qwen 2.5 and Llama 3.2, and both demographic dimensions, this results in 8,640 total responses, with 8,637 valid labeled outputs.
Results & Findings
Our results show that power-compliance behavior varies across both models and identity conditions. Across all valid responses, 52.8% were classified as non-compliant, 27.9% as compliant, and 19.4% as loophole-seeking. Qwen 2.5 produced 29.4% compliant, 15.2% loophole, and 55.4% non-compliant responses, while Llama 3.2 produced 26.3% compliant, 23.5% loophole, and 50.2% non-compliant responses.
Identity-based variation was observed across both demographic dimensions. For age, Qwen 2.5 showed a 4.7 percentage point difference in compliance between young and old personas, while Llama 3.2 showed a smaller 0.8 percentage point difference. For disability, Llama 3.2 showed a 5.9 percentage point difference between physically disabled and able-bodied personas, while Qwen 2.5 showed a 2.0 percentage point difference. These results indicate that identity-conditioned variation depends on both the model and the demographic dimension.
To analyze model behavior, we apply attribution-based methods such as Integrated Gradients to identify which tokens influence model decisions across scenarios. Using these attribution signals, we construct steering vectors that represent directions in the model’s activation space associated with specific behavioral patterns, particularly loophole-seeking responses.
Steering experiments are conducted on the base (non-persona-conditioned) dataset using Llama-3.1-8B-Instruct and Qwen-2.5-7B-Instruct as the target models to evaluate whether these attribution signals can be used to control model behavior more generally. This is done independently of the identity-conditioned setting. During inference, these vectors are applied by modifying hidden states, allowing us to guide model outputs without updating model weights. This approach is inspired by recent work on attribution-based steering [2]. We perform systematic sweeps across layers, steering strengths (alpha), and vector construction methods (mean and PCA), evaluating positive, negative, and mixed steering strategies.
These experiments showed that certain configurations consistently reduced loophole behavior. For Llama-3.1-8B-Instruct, the best configuration used PCA-based steering at layer 11 with alpha = 2.0 in both-direction mode. For Qwen-2.5-7B-Instruct, the best configuration used mean-based steering at layer 10 with alpha = 1.5 in both-direction mode. These configurations reduced loophole-seeking behavior but introduced trade-offs in general reasoning performance.
On a 1,000-question subset of the MMLU benchmark [3], Llama-3.1-8B-Instruct dropped from 46.0% baseline accuracy to 26.7% after steering, while Qwen-2.5-7B-Instruct dropped from 48.9% to 39.7%. This demonstrates that while steering can improve behavioral consistency, it may negatively impact general task performance depending on configuration strength and layer selection.
Neurons of Bias: Power-Mediated Compliance and Mitigation in Large Language Models
Introduction & Problem Statement
Large language models (LLMs) are increasingly deployed in sociotechnical systems involving power, authority, and compliance, raising important questions about fairness, consistency, and reliability across diverse user populations. As these systems are integrated into hiring tools, customer service systems, assistive technologies, and decision-support pipelines, they are expected to operate consistently regardless of the identity they are prompted to adopt. However, when demographic attributes are introduced into prompts, model behavior may shift even when identity does not change the underlying task.
In power-mediated settings, such variation is especially important because differences in compliance, refusal, or loophole-seeking behavior may reflect biased social associations rather than contextually relevant adaptation. In the absence of task-specific justification, compliance distributions should remain relatively stable across identities, and deviations from this stability may indicate identity-conditioned bias. We focus on age and disability as two key dimensions and study how these factors influence model decisions in structured authority-based scenarios.
Methods
We use a structured dataset of power and compliance scenarios introduced by Choi et al. (2025) [1], where each scenario presents a fixed authority-driven decision with three response types: compliant, non-compliant, and loophole-seeking. We construct identity-conditioned prompts from this dataset using a controlled single-dimension design, where the model adopts either a young or old identity, or a physically disabled or able-bodied identity, while all other aspects of the prompt remain constant.
Each run consists of 1,080 scenarios with two identity conditions per scenario, producing 2,160 responses per run. Across two models, Qwen 2.5 and Llama 3.2, and both demographic dimensions, this results in 8,640 total responses, with 8,637 valid labeled outputs.
Results & Findings
Our results show that power-compliance behavior varies across both models and identity conditions. Across all valid responses, 52.8% were classified as non-compliant, 27.9% as compliant, and 19.4% as loophole-seeking. Qwen 2.5 produced 29.4% compliant, 15.2% loophole, and 55.4% non-compliant responses, while Llama 3.2 produced 26.3% compliant, 23.5% loophole, and 50.2% non-compliant responses.
Identity-based variation was observed across both demographic dimensions. For age, Qwen 2.5 showed a 4.7 percentage point difference in compliance between young and old personas, while Llama 3.2 showed a smaller 0.8 percentage point difference. For disability, Llama 3.2 showed a 5.9 percentage point difference between physically disabled and able-bodied personas, while Qwen 2.5 showed a 2.0 percentage point difference. These results indicate that identity-conditioned variation depends on both the model and the demographic dimension.
To analyze model behavior, we apply attribution-based methods such as Integrated Gradients to identify which tokens influence model decisions across scenarios. Using these attribution signals, we construct steering vectors that represent directions in the model’s activation space associated with specific behavioral patterns, particularly loophole-seeking responses.
Steering experiments are conducted on the base (non-persona-conditioned) dataset using Llama-3.1-8B-Instruct and Qwen-2.5-7B-Instruct as the target models to evaluate whether these attribution signals can be used to control model behavior more generally. This is done independently of the identity-conditioned setting. During inference, these vectors are applied by modifying hidden states, allowing us to guide model outputs without updating model weights. This approach is inspired by recent work on attribution-based steering [2]. We perform systematic sweeps across layers, steering strengths (alpha), and vector construction methods (mean and PCA), evaluating positive, negative, and mixed steering strategies.
These experiments showed that certain configurations consistently reduced loophole behavior. For Llama-3.1-8B-Instruct, the best configuration used PCA-based steering at layer 11 with alpha = 2.0 in both-direction mode. For Qwen-2.5-7B-Instruct, the best configuration used mean-based steering at layer 10 with alpha = 1.5 in both-direction mode. These configurations reduced loophole-seeking behavior but introduced trade-offs in general reasoning performance.
On a 1,000-question subset of the MMLU benchmark [3], Llama-3.1-8B-Instruct dropped from 46.0% baseline accuracy to 26.7% after steering, while Qwen-2.5-7B-Instruct dropped from 48.9% to 39.7%. This demonstrates that while steering can improve behavioral consistency, it may negatively impact general task performance depending on configuration strength and layer selection.