Micah Zhang
ML Research Engineer • LLM-Based Apprenticeship • Human-Centered AI • Adaptive Inference
I am an ML research engineer studying how large language models can provide cognitive scaffolding for difficult human work while preserving agency and supporting long-term growth. My current work focuses on LLM-based apprenticeship, adaptive inference, and budgeted refinement methods for research ideation and other high-friction forms of knowledge work.
About
My research interests center on language model systems that help people do difficult cognitive work without replacing human agency. I am especially interested in when AI systems should provide more guidance, refine a candidate more deeply, structure feedback, or support a learner through a difficult step.
I am currently preparing PhD applications to study human-centered LLM systems, support allocation, adaptive inference, and evaluations that distinguish AI as substitution from AI as scaffolding.
Research Vision
I want to build and study LLM-based apprenticeship systems: AI systems that help people bridge the gap between potential and present capability, persist through difficult work, and become more capable over time. I believe support changes outcomes, and I want my research to make that claim technically precise and practically real.
Research Interests
- LLM-based apprenticeship and cognitive scaffolding
- Human-centered AI systems that preserve and grow agency
- Research ideation, support allocation, and budgeted refinement
- Adaptive inference, test-time compute, and quality-cost tradeoffs
- Evaluation of whether AI support improves human capability over time
Selected Work
This project studies a conditional support-allocation problem: given a noisy pool of LLM-generated research ideas, how should limited refinement effort be allocated to produce a stronger, more diverse, more execution-ready portfolio? The work compares raw generation, reranking, uniform refinement, random subset refinement, diversity-aware refinement, and two-stage micro-triage plus heavy refinement.
HALO studies adaptive refinement for improving language-model quality-compute tradeoffs. The project investigates selective latent refinement and controller-based allocation of additional computation for frozen language models.
This paper introduces Fine-tuned Generated Chain-of-Thought, a method combining a fine-tuned BERT encoder, zero-shot chain-of-thought generation, and a fine-tuned LLM for solving BRAINTEASER sentence puzzles.