Public repository · Research prototype
TRIAD
Teacher-Routed Interaction-Adaptive Distillation. A continual personalization research prototype using a frozen teacher and per-user LoRA to learn from interactions.
YUKIUSA LAB / PROJECTS
Projects
Building tools, AI systems, and experimental ideas in code.
Purpose / problem
Explore how a model changes when interaction history accumulates in user-specific trainable weights, beyond prompts or conversation memory. Change is not assumed to mean improvement.
What I built
A local research CLI connecting a frozen Teacher, a shared frozen base and per-user LoRA adapters. Online updates, replay, checkpoints, restoration and before/after evaluation form a single experimental workflow.
Teacher → Student
A frozen teacher interprets interactions as learning targets. The student performs real backpropagation updates to per-user LoRA while the shared base model stays frozen.
Learning across experiences
Per-user adapters and replay memory persist across interactions. Checkpoints and rollback help manage learning state.
Inspecting change
The repository includes workflows to compare personalization and drift. It uses Python, PyTorch, Transformers with PEFT, and Ollama for teacher inference.
Based on the repository README. Personalization may improve, distort, or degrade responses. The small offline demo exercises update, storage, and evaluation workflows; it does not demonstrate or guarantee natural-language quality.
FUTURE RESEARCH
Self-Improving Humanoid Robot
Long-term vision · Coming Soon
A long-term research vision for humanoid systems that learn robot states and experiences through neural networks, and gradually improve through experience.
This is a future direction, not an implemented project or a set of validated capabilities. It connects Embodied AI, Vision-Language-Action, Robot Foundation Models, and Continual Learning.