Ryan Kim

Built bioinformatics agent RL environments (Chimera) end to end that hold GPT-6 Astra, Claude Opus 5.5, and Fable 5.1 to <0.5/1.0 reward: a synthetic generator of verified-solvable tasks, Docker/Harbor execution, and reward grading. Sped up task generation 6× by adding caching and parallelism to the generator pipeline. Analyzed agent trajectories with an LLM-as-a-judge that checks each run against expected milestones.

Led research extending Energy-Based Transformers to multi-token prediction through Explorative Modeling (XM). Implemented the full training stack and compared candidate energy formulations against block-diffusion LLMs.

Drove 37% of the gain over baseline on NanoPath v1 (26% on v2) by building an LLM autoresearch agent with dual-branch search and anti-overfitting cycles (3rd-author paper, accepted at a NeurIPS 2026 workshop). Cut config errors 80% and token use 15× by designing agent infrastructure (knowledge graphs, SLURM, worktree isolation). Raised mean PathoROB robustness 0.726→0.891 and improved downstream AUC across 7 pathology foundation models by introducing SPECTRA, label-free contrastive LoRA fine-tuning (first-author paper, accepted at a NeurIPS 2026 workshop).

Found that error doubles out of distribution (MAE ~0.10→~0.20) by benchmarking 12 vision foundation models for tumor-purity regression on TCGA and internal cohorts in PyTorch.

Engineered a CPU-only LLM pipeline that classified 370 clinically annotated lung-cancer CT reports, optimizing prompts for Llama 3.1-8B and Qwen3-4B with DSPy MIPROv2 (3rd-author abstract, accepted at ACRO 2026).

Predicted survival of 185 glioblastoma patients with a DNABERT-based Cox model on non-coding DNA mutations.