Microsoft’s Learning Zone leverages Phi Silica with LoRA finetuning to create high-quality Kahoot! quizzes on-device. This approach reduces rejection rates by 75% and boosts quiz quality 4.6X, enabling educators to generate engaging, personalized lessons efficiently using AI-powered local models.
Phi Silica & LoRA: Revolutionizing On-Device AI for Education
At Build 2025, Microsoft unveiled a breakthrough in AI fine-tuning: LoRA (low-rank adaptation) integration with Phi Silica. This local Small Language Model (SLM) powers Copilot+ PCs, enabling smarter, faster, and more efficient AI customization.
What’s New: LoRA Fine-Tuning Meets Phi Silica
LoRA fine-tuning updates only a small subset of model parameters using custom data. This means improved task-specific performance without compromising the model’s overall capabilities. Microsoft applied this to generate high-quality Kahoot! quizzes directly on-device, slashing rejection rates by 75% and boosting subjective quality scores by 4.6 times.
“LoRA makes fine-tuning more efficient by updating only a small subset of parameters of the model with custom data.”
Microsoft Learning Zone & Kahoot! Integration
Microsoft Learning Zone, aka Project Spark, is a free AI-powered learning companion app for Copilot+ PCs. It helps educators create personalized lessons and interactive quizzes seamlessly. Partnering with Kahoot!, Microsoft enables engaging classroom games generated entirely by Phi Silica.
Instead of training separate models for each task, LoRA adapters specialize a single Phi Silica base model to handle diverse generation needs with minimal overhead.
Quality: Verifiable vs. Subjective
Quality control splits into two parts:
- Verifiable Quality: Enforces strict output formats like character limits and structure, ensuring smooth UX across devices.
- Subjective Quality: Focuses on engagement, clarity, and educational value, judged via human-like rubrics and scaled with a novel agentic evaluation framework.
Behind the Scenes: Dataset Curation & LoRA Training
Microsoft curated a rich dataset of 13,000 synthetic Kahoot!-style Q&A pairs using GPT-4o as a teacher model. This distillation approach bootstrapped high-quality training data from real-world learning materials.
Using the AI toolkit’s LoRA feature, Microsoft fine-tuned adapters on the quantized Phi Silica model. This customization allowed on-device generation that fits Kahoot!’s strict format requirements.
System Prompt Optimization
Initially, the system prompt was long and complex, specifying JSON formats and question styles. After LoRA training, a shorter prompt sufficed, reducing latency and resource use.
“The combination of the LoRA adapter and this new system prompt gave us the desired output.”
During inference, a reinforced JSON format prompt ensured consistent output, crucial for Kahoot! quiz compatibility.
Hyperparameter Tuning & Evaluation
Microsoft experimented with hyperparameters but found default settings worked best for stability and quality. They used a multi-agentic “agent-as-a-judge” framework to automate subjective quality checks, cutting down costly human reviews.
Results showed a 75% reduction in guardrail rejections and a 4.6X boost in subjective quality, directly enhancing user experience by lowering generation delays and increasing quiz relevance.
Why It Matters
This breakthrough demonstrates how efficient fine-tuning with LoRA can unlock powerful, personalized AI experiences right on your device. For educators, it means faster, smarter tools that create engaging content without cloud dependencies.
In short, Microsoft’s Phi Silica + LoRA combo sets a new standard for on-device AI in education, blending cutting-edge research with practical classroom impact.
From the Windows Blog
