Anthropic Evaluating Feature Steering: A Case Study in Mitigating Social Biases

Anthropic has conducted a quantitative evaluation of feature steering in Claude 3 Sonnet to determine if modifying internal model features can reliably mitigate social biases. The research reveals that while steering can influence specific behaviors and reduce certain biases, it often produces unpredictable "off-target effects" and can severely degrade general model capabilities if applied outside a specific range.

The Feature Steering "Sweet Spot"

Feature steering involves using dictionary learning to identify interpretable directions (features) in a model's residual stream and adding a constant in that direction to modify outputs. Anthropic identified a "steering sweet spot"—a steering factor between -5 and 5—where model capabilities remain stable.

Within this range, general knowledge and reasoning abilities, measured by MMLU and PubMedQA benchmarks, are preserved. However, once the steering factor exceeds these bounds, model capabilities drop sharply, potentially rendering the model unusable.

Impact on Social and Political Biases

Using the BBQ (Bias Benchmark for QA) and a model-written evals dataset, Anthropic tested 29 features related to social biases and political ideologies. The results indicate that steering can be both targeted and unpredictable:

Targeted Bias Influence

• Bias Reduction: Positively steering the "Multiple perspectives and balance" feature reduced overall BBQ bias scores by approximately 3%, with significant decreases in Age bias (17%) and Disability Status bias (7%).
• Political Stances: Steering the "Pro-life and anti-abortion stance" feature increased anti-abortion selections by 50%, while steering "Left-wing political ideologies" decreased them by 47%.

Off-Target Effects

Anthropic observed that steering a feature often affects unrelated domains, suggesting that feature activation context does not always map directly to resulting behavior:

• Gender and Age Bias: Steering the "Gender bias awareness" feature increased gender bias scores by 10%, but unexpectedly increased age bias scores by 13%.
• Cross-Domain Political Effects: Steering the "pro-life" feature had a larger impact on anti-immigration selections (21.60% increase) than the feature specifically designed for immigration concerns (3.90% change).

Promising Neutrality Features

Researchers identified two specific features—"Neutrality and Impartiality" and "Multiple Perspectives"—that consistently reduced bias scores across all nine BBQ benchmark dimensions within the steering sweet spot. While steering "Neutrality and Impartiality" caused a slight decrease in BBQ accuracy, the "Multiple Perspectives" feature maintained accuracy across the steering range, suggesting a potential path for mitigating social biases without sacrificing general utility.

Technical Limitations and Lessons Learned

Anthropic noted several constraints and findings that impact the reliability of feature steering as a safety tool:

• Evaluation Noise: Accuracy was estimated by sampling 10 responses per question, which introduced noise into the results.
• Context vs. Behavior Disconnect: Features were identified by where they activate, not the behaviors they produce. Consequently, steering a feature does not always lead to predictable changes in model outputs.
• Scope of Study: The study analyzed only 29 features out of millions and used five evaluations, limiting the ability to generalize findings.

Future Research Directions

To improve the precision of feature steering, Anthropic proposes several future avenues:

• Scaling SAEs: Investigating if larger Sparse Autoencoders (SAEs) result in more sensitive on-target features and fewer off-target effects.
• Refined Steering Implementation: Applying steering only to the assistant's response tokens rather than the entire prompt to avoid confounding the model's processing of human input.
• Alternative Steering Methods: Exploring multiplicative, projective, or conditional steering instead of the current additive approach to avoid creating "ungrammatical" internal states.
• Circuit Analysis: Moving beyond individual features to study "circuits"—interconnected groups of neurons—to understand how features work together to perform specific functions.