Research

Nonlinear query projections in transformer attention. Parameter-matched residual MLP queries beat the linear baseline and a model with 12.5% more non-embedding parameters. Headed to an ICLR 2026 workshop.

• PyTorch
• Transformers
• arXiv 2603.13381

A multimodal framework reframing image-text understanding as bidirectional predictive coding over a shared latent memory, co-authored with A. Yeh.

• TensorFlow
• Predictive coding
• Multimodal

A training-free system for the ARC-AGI-3 interactive reasoning benchmark that uses zero LLM calls at inference time, pairing LLM-proposed world-model rules with exact replay verification and forward-search planning.

• World models
• Search
• Python

Minimal-pair test suites at Colby's Linguistics & NLP Lab that isolate implicature and presupposition, separating real pragmatic inference from surface pattern matching.

• LLM evaluation
• NLP
• Diagnostics

Interpretability of learned world representations and JEPA-style predictive models, with an eye toward physical and agent-based applications such as autonomous driving.

• JEPA
• Interpretability

An MCP server that generates complete, schema-validated Power BI projects from live SQL schema introspection, plus a multi-tenant row-level-security analytics platform.

• Python
• MCP
• Power BI
• SQL Server

An LLM governance module with tiered human-in-the-loop review and usage protocols adopted across finance, legal, and HR teams.

• LLM orchestration
• Governance
• RAG