Angular Steering is a method for adjusting the response behavior of large language models by intervening in the model’s internal representations.
Torilab researches the foundational mechanisms within AI models and develops methods for steering AI behavior with precision and flexibility — advancing toward emotion-aware AI systems that interact and adapt more deeply to humans.
Research. Engineer. Deploy.
Selected among the Top ~3% at NeurIPS 2025, one of the world’s leading AI conferences. Torilab’s “Angular Steering” research, conducted with a researcher from the National University of Singapore (NUS), officially received Spotlight recognition. We steer LLM behavior in geometric space, enabling continuous, stable, and precise adjustment of AI responses.
These four technology pillars form the foundation of Torilab’s research: from understanding how behavior is represented within models, to developing behavior-steering methods and expanding into emotion-aware AI.
Researching how behavioral features are represented in activation space and how those representations can be intervened upon to steer model behavior. This provides a foundation for understanding and controlling AI from within, rather than relying solely on prompts.
A method for modulating language model behavior by rotating activation vectors within a subspace, moving the model’s representation toward or away from a target behavior direction. The method allows behaviors such as refusal or compliance with requests to be increased or decreased during inference.
View Research →A method that selects activation vectors meeting conditions related to a target feature before applying steering, enabling more selective, stable, and flexible behavior control. This helps responses remain coherent during steering, making the method suitable for applications requiring a high degree of stability.
Extending behavior-steering research to emotion-steering tasks and multimodal interactive AI systems, with the aim of producing responses better suited to people’s states and contexts. The goal is to create AI avatars that respond more sensitively to users’ emotions.
Torilab is responsible for the entire AI system development lifecycle — from design and model integration to operational infrastructure and continuous product improvement. Our technology platform delivers real-time, emotion-aware avatar interactions optimized for performance, personalization, and safety. Each technology below is marked with its current status and is applied according to the requirements of each product.
Model integration, dialogue control, memory management, personalization, and AI orchestration.
An orchestration layer developed in-house by Torilab that manages prompt engineering, session memory, retrieval and grounding, and real-time context adaptation. This helps maintain coherent dialogue and consistent personas while using models efficiently.
A modular multi-agent architecture that enables avatars to handle multiple tasks, such as booking appointments, capturing and managing business cards, long-term memory, context switching, and real-time information retrieval. The avatar thus acts not just as a conversational partner but as a personal assistant.
Uses user profiles, historical context, and preference signals to personalize conversations. Long-term memory and adaptive behavior continue to be enhanced across multiple interaction sessions.
Image analysis and generation, voice processing, and synchronization of avatar lip movements and expressions.
A multimodal emotion recognition pipeline that processes inputs from text messages, voice, and facial expressions (via camera). The models are optimized for latency and accuracy in real-time inference.
A TTS engine developed in-house by Torilab, focusing on Japanese phonetics and prosody. The engine supports adjustable levels of familiarity and a rich range of emotional expression, generating natural, context-appropriate speech with consistent audio quality.
Upon detecting user emotions, the system orchestrates the avatar’s audio and visual outputs to express the corresponding emotion. Torilab has defined a technical roadmap for smooth transitions between complex emotions such as fear, surprise, and anger.
Data management, account authentication, payments, notifications, and connections to external services.
Building AI inference and distribution infrastructure, system monitoring, performance optimization, and incident response.
Developing lightweight emotion classifiers designed to run directly on edge devices, reducing dependence on cloud GPU resources. This approach lowers latency and costs, supports global scaling, and maintains a consistent experience in mobile environments.
Developing mobile applications, web applications, user interfaces, and administrative tools.
System testing, AI output evaluation, release quality verification, and continuous improvement.
Technologies developed by Torilab are combined with third-party models and services, and are integrated into products only when they meet the technical evaluation requirements of each system.
View All Products →Research publications, product launches, and Torilab’s collaboration milestones.
Today, we have some truly exciting news to share.
A method for steering large language model behavior by rotating vectors in activation space.
As AI becomes increasingly capable of recognizing and steering behavior, research and technology development must go hand in hand with responsibility toward people, data, and real-world applications.
We prioritize user data protection and access management throughout the development and operation of AI systems.
We test and evaluate AI responses both before deployment and during operation.
We detect, evaluate, and address inappropriate responses according to the control mechanisms applied to each system.
Torilab connects academic knowledge with engineering capabilities to conduct in-depth research with partners and co-develop AI products and services under a revenue-sharing model.
Collaborating with researchers and academic institutions to explore foundational AI challenges, exchange knowledge, and develop new research directions.
Combining research and engineering capabilities to evaluate, integrate, and develop AI solutions tailored to the requirements of each system.
Researching, building, and developing AI products or services with partners based on the strengths, resources, and collaboration goals of each party, while sharing revenue generated by the services.