We conduct focused research in Applied Meta-Learning.

New scaling laws for train-time knowledge injection: hypernetworks beat LoRA and full fine-tuning on out-of-distribution generalization, and the gap grows with target model size.

Adapting model behavior at inference time without retraining, large context windows, or touching the primary model's weights.

NEMA: meta-learning architectures that generate specialized, task-specific agents automatically.

Memory augmentation and model editing for handling outdated and domain-specific knowledge.

Benchmarking Nace's Safety Model against AWS Bedrock, Azure, NeMo Guardrails, and prompt-based approaches.

A theoretical framework connecting the geometry of data representations to training dynamics.

Efficient fine-tuning of LLMs through task-specific parameter generation.

Improving task adaptability in large language models.

A modular RAG pipeline tailored to enterprise needs: over 90% F1 on verification.

How structured, verifiable knowledge complements LLMs where they fall short. No ML background required.

Tailored, meta-learned Small Language Models for rapid, efficient adaptation.

Most post-training performance gains are reranking: aligning the proposal distribution with the task.