Research showcase of my work on the LLM-inference subteam of Guided Evolution — a framework that fuses Large Language Models with genetic-algorithm Neural Architecture Search — at Georgia Tech's VIP: Automated Algorithm Design lab.
Attribution. This is a personal write-up of collaborative, lab research. The full system is a team codebase and is not redistributed here. Source (team org): llm-inference-aad/llm-inference. The Guided Evolution framework builds on prior published work; credit to the lab and the original authors.
Guided Evolution (GE) combines the human-like reasoning of LLMs with genetic-algorithm-based Neural Architecture Search. Instead of blind mutation, an LLM proposes code/architecture changes and receives result-driven feedback (the "Evolution of Thought" loop), so the search is intelligently guided while still preserving the genetic diversity evolutionary algorithms rely on. Candidate models are evaluated on accuracy, parameter count, and training time, and the best are evolved forward.
Our subteam applies genetic-programming techniques to reduce LLM inference cost — making the evolutionary loop (which calls LLMs heavily) faster and cheaper without sacrificing search quality. This is a distributed-systems and ML-efficiency problem as much as a research one: many model candidates are trained and evaluated across an HPC cluster, with LLM inference in the critical path.
- LLM quantization experiments — evaluated quantized inference to cut latency/memory in the evolutionary loop while monitoring impact on search quality.
- RAG retrieval-quality research — studied retrieval quality for the framework's retrieval-augmented prompting, to improve the LLM's guidance signal.
- Distributed runs on GT PACE (SLURM) — ran and debugged the LLM-GE pipeline as SLURM-managed, multi-node jobs on Georgia Tech's PACE supercomputing cluster, including diagnosing multi-node inference failures.
This section reflects my individual contributions within the larger team effort.
LLM inference & quantization · RAG / retrieval evaluation · evolutionary algorithms / NAS · PyTorch · distributed computing on SLURM / HPC · Python · experiment debugging at scale
- Team codebase: github.com/llm-inference-aad/llm-inference
- Program: Georgia Tech VIP — Automated Algorithm Design
Maintained by Sri Rithish Palani. This repository documents research contributions and intentionally does not host the lab's source code.