English | 中文
AutoSR (paper) is a Claude Code plugin for symbolic ansatz search. AutoSR performs symbolic regression by searching through Research States, following the process of human scientific discovery.
Log in to autosr.app to get started!
Problems and generated artifacts live outside the plugin, usually in a problem workspace such as artifacts/pile_efficiency/.
The artifacts/ directory is where problem workspaces go; only its README is tracked, so data, runs, and generated outputs stay local.
From a problem workspace:
cd /path/to/problem-workspace
claude --plugin-dir /path/to/AutoSR/agonsr --dangerously-skip-permissions
Then run /llm-mcts 10 problem.md.
Resume an existing run:
/llm-mcts 10 problem.md --resume runs/llm-mcts_YYMMDD_HHMM
problem.md should contain the actual problem definition: objective, variables, data/docs/scripts to read, constraints, evaluation method, scoring rubric, and expected ansatz.md contents. Use paths relative to the problem workspace.
Optional. Put special per-role notes here when they should not live in the general problem statement. Format:
## Notes to ansatz-proposer
## Notes to ansatz-reviewer
## Notes to dispatcher
The dispatcher reads this file and passes each section only to the corresponding subagent.
llm-mcts is a dispatcher. It initializes or resumes a run, gets the next candidate from mcts.py next, sends fixed minimal prompts to ansatz-proposer and ansatz-reviewer, reads the <review score="X"> block from ansatz.md, calls mcts.py update, and finally shows the best candidates.
Run files are written under runs/. See agonsr/references/project_manual.md for the exact workspace and ansatz-file conventions.
@misc{zhang2026autosrautomaticsymbolicregression,
title={AutoSR: Automatic Symbolic Regression by Searching Research States},
author={Youran Sun and Kejia Zhang and Xinyu Ren and Chugang Yi and Haizhao Yang},
year={2026},
eprint={2608.16876},
archivePrefix={arXiv},
primaryClass={cs.SC},
url={https://arxiv.org/abs/2608.16876},
}