GEOSTEER: Geodesic Optimization for Activation Steering in Large Language Models
GeoSteer introduces a Riemannian optimization framework for norm-preserving activation steering in LLMs, replacing fixed one-step edits with adaptive geodesic steps on the representation manifold The method learns a nonlinear activation-space objective that distinguishes desired from undesired activations, enabling geometry-aware, multistep steering updates Norm preservation is maintained throughout, mitigating risks of representation collapse and degradation common in prior steering approaches
Analysis
TL;DR
- GeoSteer introduces a Riemannian optimization framework for norm-preserving activation steering in LLMs, replacing fixed one-step edits with adaptive geodesic steps on the representation manifold
- The method learns a nonlinear activation-space objective that distinguishes desired from undesired activations, enabling geometry-aware, multistep steering updates
- Norm preservation is maintained throughout, mitigating risks of representation collapse and degradation common in prior steering approaches
- GeoSteer outperforms state-of-the-art activation steering baselines across TruthfulQA, RealToxicityPrompts, and UltraFeedback benchmarks
- The work demonstrates that adaptive, geometry-aware optimization is a more effective paradigm for inference-time model steering than predefined trajectories
Why It Matters
Activation steering is a critical tool for aligning LLM behavior at inference time without expensive fine-tuning, and norm-preserving methods are particularly valuable for avoiding representation degradation. GeoSteer's shift from one-step, fixed-direction updates to multistep geodesic optimization addresses a fundamental limitation in the field, offering a more principled way to navigate the complex geometry of activation distributions. This has direct implications for safer, more controllable LLM deployment in production environments.
Technical Details
- Riemannian optimization formulation: GeoSteer treats activation steering as an optimization problem on a Riemannian manifold, updating hidden activations through a sequence of small geodesic steps rather than single linear projections
- Nonlinear objective learning: A learned nonlinear function in activation space distinguishes desired from undesired activations, replacing fixed steering directions with adaptive, context-sensitive guidance at each step
- Norm preservation guarantee: The geodesic update mechanism inherently preserves activation norms, preventing the representation collapse that plagues additive steering methods
- Benchmark evaluation: Evaluated on TruthfulQA (truthfulness), RealToxicityPrompts (safety), and UltraFeedback (instruction following), consistently surpassing existing norm-preserving steering baselines
- Inference-time only: No model retraining or fine-tuning required; steering is applied purely at inference by modifying hidden activations
Industry Insight
- The geodesic optimization perspective opens a new research direction for activation-space interventions, suggesting that manifold-aware methods could outperform linear approximations across a range of steering and control tasks
- Norm-preserving multistep steering offers a practical path toward safer LLM deployment, as it reduces the risk of catastrophic representation degradation while still enabling meaningful behavioral control
- As inference-time steering becomes more competitive with fine-tuning approaches, tools like GeoSteer could reduce the cost and carbon footprint of model alignment by eliminating the need for repeated retraining cycles
Disclaimer: The above content is generated by AI and is for reference only.