ChiAdapt
Inference-Time
Personalization
Personalize any frozen language model at inference time — no fine-tuning, no retraining, no user data in the weights. 2.8 KB per user. Deployed on a 1.1B parameter model with measurable gains across all domains.
Results by Domain
Tested on TinyLlama 1.1B. All five domains improved. H₀ (no effect) rejected.
| Domain | Baseline | ChiAdapt | Change |
|---|---|---|---|
| Cooking | 44.2% | 48.8% | +4.6% |
| Science | 41.7% | 44.9% | +3.2% |
| Creative Writing | 39.1% | 43.5% | +4.4% |
| Technical | 46.3% | 52.1% | +5.8% |
| Conversational | 42.8% | 50.5% | +7.7% |
What Makes It Different
No fine-tuning infrastructure. No gradient computation at inference. No user data stored in model weights.
Zero Retraining
Base model stays frozen. Per-user adapters are computed from interaction history without any gradient descent.
2.8 KB Per User
Each user profile is a tiny set of attention-head modulation weights. Stores 1 million users in 2.8 GB.
Privacy by Architecture
User preferences never enter the base model weights. Adapters can be deleted instantly with no residual.
Scales to Any LLM
Validated on GPT-2 (124M) and TinyLlama (1.1B). Architecture-agnostic — works on any transformer.
Real-Time Adaptation
Adapters update on every interaction. No batch reprocessing needed.
Degradation-Safe
Maximum observed degradation is +1.46% on worst-case domain. Mean improvement +5.13%.
Use Cases
AI Assistants
Personalize responses to user style without storing conversation history in the model.
Enterprise Chatbots
Each department gets a tailored experience from the same base model deployment.
Education Platforms
Adapt difficulty, tone, and explanation style per student automatically.
Content Recommendations
Modulate generation style based on user engagement patterns.
Interested in ChiAdapt?
Patent pending. Available for enterprise licensing and consulting engagements.