Angular Steering
Rotates activation vectors within a two-dimensional subspace to move the model’s representations toward or away from a target behavior. The method intervenes during inference without requiring the model to be retrained.
Research
NeurIPS 2025 Spotlight · Top ~3%
Exploring the
Exploring the
Foundations of
Artificial Intelligence
At Torilab, we study how behavior emerges inside large language models and develop methods to steer AI responses from within the models themselves.
Push the boundaries of emotion-aware AI.
NeurIPS 2025 Spotlight · Top ~3%
Geometric Steering
Geometric Steering
of AI Behavior
Whenever AI processes language, internal representations emerge within a high-dimensional space, resembling arrows pointing in different directions. Certain behaviors, such as refusing a request or expressing emotion, can be associated with particular directions in this space. Rather than adding to or removing a direction, Angular Steering rotates the arrow by an angle. This provides a way to adjust behavior continuously, much like turning a dial.
Adaptive Angular Steering
A selective variant that rotates only activation vectors aligned with the target feature, helping produce more stable and coherent outputs.
Results
Experiments across multiple model families on refusal/compliance and emotion steering tasks demonstrate stable behavioral control while maintaining general language capabilities. The work received NeurIPS 2025 Spotlight · Top ~3%; applying the methods to individual products depends on the technical evaluation of each system.
Research PaperAngular Steering: Behavior Control via Rotation in Activation Space
AuthorsMinh Hieu Vu (Torilab), Tan M. Nguyen (NUS)
RecognitionNeurIPS 2025 SpotlightTop ~3% of submissions
Core Research DirectionsExpanding the Boundaries
Expanding the Boundaries
of AI Behavior Steering
Torilab investigates how AI models form representations, generate responses, and change their behavior. Building on these underlying mechanisms, we develop steering methods aimed at enabling AI systems with deeper interactive and emotion-aware capabilities.
We study how behavioral features are represented in activation space and how these representations relate to AI-generated responses. This provides a foundation for intervening in model behavior from within.
Adjusting model representations toward a target behavior by rotating activation vectors within a subspace. This research received Spotlight recognition at NeurIPS 2025.
Input
Activation
Angular Steering
Behavior
Response
A selective steering method published alongside Angular Steering at NeurIPS 2025. It acts only on activation vectors aligned with the target feature, helping make the steering process more stable and coherent.
Extending behavioral steering research to emotion-related tasks and multimodal interaction, laying the groundwork for AI that responds more appropriately to context and human signals.
Research PrinciplesGrounded in Science.
Grounded in Science.
Open to Verification.
Torilab pursues foundational AI research by combining investigations into the internal mechanisms of models with systematic experimentation. We aim to develop methods grounded in clear scientific principles, evaluated under specific conditions, and refined through continued research.
Investigating Underlying Mechanisms
We study how internal model representations relate to observable behavior, rather than examining outputs alone.
Evaluating Through Experiments
We evaluate methods on specific tasks, considering both steering effectiveness and their impact on model performance.
Connecting Science with Technology
We clearly distinguish experimental findings from capabilities that have been deployed, while exploring how research can be translated into practical applications.
Research CollaborationResearch with
Research with
the Academic
Community
Torilab welcomes opportunities to collaborate with researchers, universities, and academic institutions interested in foundational AI challenges. Angular Steering emerged from collaboration between Torilab and a researcher affiliated with the National University of Singapore (NUS). We aim to connect knowledge, exchange methodologies, and explore new research directions in AI behavior steering, emotion-aware AI, and multimodal interaction.