NanoProteinLM: Towards Hill-Climbable Protein Language Model Research

Agentic autoresearch could extend recursive self-improvement to scientific discovery. Given an objective, an agent that proposes a change, tests it, keeps what works and repeats could explore a design space far larger than any group can cover by hand. Handing a real scientific question to such a loop, however, seldom returns anything useful, usually because the setting around the question was never configured for a closed loop. We ask what makes a research task hill-climbable, propose the ingredients an environment needs, and apply them to protein language-model training. The result is NanoProteinLM, an open and fully automatable environment with a public decontaminated corpus, a readable trainer, frozen evaluations and an explicit autoresearch protocol. Inside it, a simple sequential hill-climber proposes training changes that we combine with owner-selected components and test at larger budgets. The assembled recipe improves long-range contact precision by 8.4 percentage points over our ESMC-style baseline at matched data and token budget. We interpret the four accepted changes retained after validation and explain why the others were dropped. We also share the lessons we learned in making the environment hill-climbable. Protein researchers gain an open recipe they can rerun and improve, and agent researchers gain a benchmark in which their progress returns directly to that recipe. The code and data are open-sourced on GitHub (https://github.com/Lumin-Science/Nano-ProteinLM) and Hugging Face (https://huggingface.co/datasets/LuminScience/LuminBench-Nano-ESMC).