数学研究代理
AI Skill Hub 强烈推荐：数学研究代理 是一款优质的Agent工作流。AI 综合评分 8.0 分，在同类工具中表现稳健。如果你正在寻找可靠的Agent工作流解决方案，这是一个值得深入了解的选择。
📚 深度解析
数学研究代理 工作流的设计遵循"最小配置，最大复用"原则：核心逻辑已经封装好，用户只需配置自己的 API Key 和业务参数即可快速上手。工作流内置错误处理和重试机制，在网络波动或 API 限速等情况下仍能稳定运行，适合作为生产环境的自动化基础设施。
在实际部署时，建议先在测试环境中运行 3-5 次，验证各个环节的输出结果符合预期，再部署到生产环境。AI Skill Hub 评分 8.0 分，是同类 Agent 工作流中的精选推荐。
📋 工具概览
数学研究代理 是一套完整的 AI Agent 自动化工作流方案。通过可视化的节点编排，将复杂的多步骤任务拆解为清晰的自动化流程，实现全程无人值守的智能处理。支持与数百种外部服务和 API 无缝集成，适合构建数据处理管线、业务自动化和 AI 辅助决策系统。
📖 中文文档
数学研究代理 是一套完整的 AI Agent 自动化工作流方案。通过可视化的节点编排，将复杂的多步骤任务拆解为清晰的自动化流程，实现全程无人值守的智能处理。支持与数百种外部服务和 API 无缝集成，适合构建数据处理管线、业务自动化和 AI 辅助决策系统。
- 可视化 Agent 工作流编排，无需编写复杂代码
- 支持多步骤自动化任务链，实现全流程无人值守
- 与外部 API、数据库和第三方服务无缝集成
- 内置错误处理与自动重试机制，保障稳定运行
- 提供可复用的自动化模板，快速在同类场景部署
- 自动化日常重复性工作，将精力集中于创造性任务
- 构建数据采集 → 处理 → 输出的完整自动化管线
- 实现跨平台、跨系统的数据流转和业务协同
# 克隆仓库 git clone https://github.com/sjtuytc/ResearchMathAgent cd ResearchMathAgent # 查看安装说明 cat README.md # 按 README 完成环境依赖安装后即可使用
- 访问 GitHub 仓库获取工作流文件
- 在对应平台（Dify / Flowise / Make 等）中找到「导入工作流」功能
- 上传工作流文件
- 按照提示配置必要的环境变量和 API Key
- 运行测试确认流程正常后投入使用
# 查看帮助 researchmathagent --help # 基本运行 researchmathagent [options] <input> # 详细使用说明请查阅文档 # https://github.com/sjtuytc/ResearchMathAgent
# researchmathagent 配置说明 # 查看配置选项 researchmathagent --config-example > config.yml # 常见配置项 # output_dir: ./output # log_level: info # workers: 4 # 环境变量（覆盖配置文件） export RESEARCHMATHAGENT_CONFIG="/path/to/config.yml"
RMA: an Agentic System for Research-Level Mathematical Problems
COLM 2026 · The 2nd Workshop on Lifelong Agents: Learning, Aligning, and Evolving
Language: English | 中文
</div>
---
Abstract
<details> <summary>Read full abstract</summary>
We present Research Math Agents (RMA), an agentic framework for automated reasoning on research-level mathematical problems. Unlike prior studies centered on competition mathematics or formal theorem proving, RMA targets research-level mathematical problems that require long-horizon reasoning, literature grounding, and iterative proof refinement. RMA decomposes research-level proof solving into specialized modules for problem analysis, literature search and understanding, fair comparison, knowledge-bank construction, and proof verification, all coordinated by initializer, proposer, and verifier agents through a shared structured memory. Within this unified framework, these agents operate in a multi-role, multi-round workflow, collaboratively generating, refining, and verifying candidate proofs through iterative feedback. We evaluate RMA on the First Proof benchmark, which consists of ten research-level problems contributed by expert mathematicians across diverse domains. Through comprehensive expert evaluation, RMA outperforms strong baselines on the First Proof benchmark, including GPT-5.2R and Aletheia, solving eight out of ten research problems and producing more logically sound and readable proofs. Our comprehensive ablation studies further show that performance gains arise from the interaction of structured reasoning modules, iterative refinement, and verifier-based feedback, rather than any single component.
</details>
---
Overview
RMA targets research-level mathematics (not competition math or formal theorem proving) by combining specialized modules for problem analysis, literature search and understanding, fair comparison, knowledge-bank construction, and proof verification.
Within a multi-role, multi-round workflow, initializer/proposer/verifier agents share structured memory to iteratively generate, refine, and validate candidate proofs. On the First Proof benchmark, RMA reports stronger results than strong baselines through structured modules, iterative refinement, and verifier feedback.
---
Highlights
RMA is the first agentic framework that targets research-level mathematical proof — not competition problems, not formal theorem proving — by combining specialized agents, structured memory, and iterative verifier feedback.
| Feature | Detail |
|---|---|
| **7 research-level datasets** | First Proof Rounds 1 & 2, Erdős Problems, Formal Conjectures, ResearchMath-14k, Unsolved Math, AIM Problem Lists — totalling 22,000+ problems |
| **Multi-agent pipeline** | Initializer → Proposer → Verifier → Refiner, coordinated through shared structured memory |
| **State-of-the-art results** | Solves **8 / 10** First Proof Round 1 problems; outperforms GPT-5.2R and Aletheia |
| **Two Claude backends** | Anthropic Messages API (pay-per-token) *or* Claude Code local CLI (Pro/Max subscription) |
| **Live web UI** | Streaming step-by-step viewer, live PDF preview, per-question issue tracker, token cost display with provider attribution and pie charts |
| **Autonomous daily worker** | Runs the solver overnight with no human in the loop, writes dated reports to documents/ |
| **Benchmark-fair sandbox** | Contamination boundary enforced in code — prior solutions are never read by the solver |
| **Agentic GitHub Issues API** | REST API (/api/gh/issues) so multiple agents can coordinate on real GitHub Issues |
---
Paper Build
latexmk -pdf -interaction=nonstopmode -halt-on-error main.tex---
⚡ Quickstart
Solve a research-level math problem on your own Claude subscription in three steps — no API key, no Google Cloud / Vertex, no per-token billing:
git clone https://github.com/sjtuytc/ResearchMathAgent
cd ResearchMathAgent
./scripts/quick_install.sh # sets up an isolated env + the `rma` CLI + Claude Code
source .venv/bin/activate # activate it (the installer prints this line)
claude login # log in with YOUR Claude Pro/Max subscription
rma solve q6 # solve a problem — billed to your subscriptionrma solve <q> uses the Claude Code backend by default, so every run is billed to your claude login subscription and never to a developer's API account or Vertex AI. Pick any First Proof problem q1–q10 (or a dataset problem with --dataset <slug>). Want a no-LLM dry run? rma solve q6 --model-name rma-skeleton.
---
Quick Start
Fastest path — your Claude subscription, no API key (see ⚡ Quickstart above):
./scripts/quick_install.sh # isolated env + `rma` CLI + Claude Code backend
source .venv/bin/activate # activate it (the installer prints this)
claude login # your Claude Pro/Max subscription
rma solve q6 # solve on your subscription (default backend)Other options:
```bash
Pay-per-token Anthropic API instead of the subscription
export ANTHROPIC_API_KEY="<your key>" rma solve q6 --model-name claude-opus-4-8 ```
---
`rma/` — CLI entry points
| File | Role |
|---|---|
cli.py | Top-level rma argument parser; dispatches to solve, push, memory, doctor subcommands. |
push.py | rma push — runs push-forward (issues + meetings + documents), refreshes concepts/insights/proof-eval, then builds the master context-report PDF. |
solve.py | rma solve — runs a full solver agent on one problem: parse → propose → verify → refine → consolidate. |
models.py | Model name constants and aliases used across CLI flags and API calls. |
memory.py | rma memory — prints or clears the push-forward state file. |
doctor.py | rma doctor — environment health check: Python version, tectonic, Claude CLI / API keys. |
__main__.py | python -m rma entry point; delegates to cli.py. |
`webapp/` — FastAPI server
| File | Role |
|---|---|
server.py | All API endpoints for the research web app (proof CRUD, PDF compile, issues, meetings, insights, context reports, literature). |
agent.py | Base agent class and prompt-execution loop shared by all agent types (critic, solver, meeting, document). |
claude_code.py | Claude Code CLI driver — runs the claude binary for subscription-based (Pro/Max) LLM calls, plus the one-shot complete_via_cli() helper. |
llm.py | One-shot completion helper (complete()) routed through the Claude subscription CLI. |
context_report.py | Builds book-style LaTeX context reports per problem (Problem → Evaluation → Best Proof → Meetings → Issues → Insights) and compiles them to PDF via tectonic; also builds the combined master PDF for rma push. |
proof_eval.py | LLM rubric evaluation of the best proof: answer accuracy, logical correctness, proof completeness, proof clarity — stored in documents/questions/<pid>/proof_eval.json. |
insight_agents.py | LLM agents that generate system-level, dataset-level, and per-question insight summaries from current project state. |
insight_loop.py | Background polling loop that periodically regenerates insight summaries. |
insights.py | Load/save insight JSON files from webapp/insights/<level>/. |
issue_agents.py | Critic agent (discovers proof gaps → opens issues) and solver agent (resolves open issues via LLM proof-writing). |
issue_loop.py | Background loop that auto-runs the issue solver on open issues while the server is running. |
issue_pdf.py | Compiles a single issue thread or all issues for a problem into a PDF. |
issues.py | CRUD for issue JSON files under webapp/issues/<dataset>/<pid>/. |
meet_agents.py | Mathematician-persona discussion agents that run multi-round research meetings and produce action plans. |
meet.py | CRUD for meeting rooms stored under documents/questions/<pid>/meets/. |
meet_pdf.py | Compiles a meeting room's notes (plan + discussion transcript) into a PDF. |
push_forward.py | Orchestrator: for each problem runs issue-discovery → solver → meeting → document update; called by rma push and the nightly cron. |
push_forward_cli.py | CLI wrapper so push_forward can run outside uvicorn's auto-reload process. |
concepts.py | Load/save/generate per-problem concept lists (core + background) from problem LaTeX. |
concepts_pdf.py | Compiles a problem's concept list into a standalone PDF. |
proofs.py | Load/store best-proof records; get_best_proof, consolidate_best, compile_best_pdf. |
proof_history.py | Load and summarize the version history of proof attempts for a problem. |
problem_pdf.py | Compiles a raw problem .tex file (with shared preamble) into a standalone PDF. |
problem_export.py | Export problem statements in various formats (JSON, plain text, LaTeX). |
latex.py | tectonic / pdflatex wrapper: compile_tex, compile_problem_pdf, safe_pdf_name, PDF directory helpers. |
dataset_store.py | Read/query problem metadata from data/datasets/<slug>/ (titles, statements, solution status). |
documents.py | List and read document files under documents/questions/<pid>/ (overview, strategies, timeline, etc.). |
rich_documents.py | Regenerate question overview/progress documents with AI-written content after each push. |
doc_bundle.py | Build the combined "bundle.pdf" from all documents for a question or dataset. |
literature.py | Search, download, and seed the global paper library for a problem's area. |
hero.py | Generates the overview/strategy document for a problem (the "hero" document shown in the Documents tab). |
runs.py | Track experiment run metadata: start time, model name, experiment name, completion status. |
smoke_pipeline.py | External eval pipeline: POST /api/solve → async proof generation + LLM rubric evaluation. |
solvability_eval.py | Load/save per-problem solvability scores produced by the filter app. |
solve_finalize.py | Post-solve cleanup: consolidate proof files, update best-proof record, write summary. |
todos.py | Per-problem TODO list CRUD (stored in documents/questions/<pid>/todos.json). |
token_log.py | Track and display LLM token usage and cost per session, with provider attribution. |
tools.py | Tool definitions (file read/write/search/run) available to claude_code agents during solving. |
github_issues.py | GitHub Issues REST API wrapper for agent-coordinated issue tracking on real GitHub repositories. |
devlog.py | Append timestamped entries to documents/devlog.jsonl for session and event history. |
daily.py | Scheduled daily tasks — currently wraps push-forward as a cron-style target. |
seed_fp2.py | One-shot importer: seeds the first_proof_2 dataset from the GitHub 1stproof/batch-2 repository. |
prefix.py | API_PREFIX constant (/rmac/solve) shared by modules that generate absolute API URLs. |
__main__.py | python -m webapp entry point: starts uvicorn with HOST/PORT env vars. |
Named experiment + skeleton model (pipeline test)
rma solve --all --exp-name proofs_test_all_june13 --model-name rma-skeleton
Web UI — streaming solver, live PDF preview, per-question issue tracker
pip install -e ".[webapp]" python -m webapp # → http://127.0.0.1:8000
高质量的AI工作流，适用于数学研究
- 构建多智能体协作系统的 Agent 开发者
- Agent 任务先做 dry-run 验证工具调用链，再开启自主执行
- API key 直接提交到 git 仓库（请用 .env 并加入 .gitignore）
- CLI：直接 npm install -g / pip install，命令行调用
- 云端托管：可放在 Vercel / Railway / Fly.io 等 PaaS 平台
⚡ 核心功能
- 可视化 Agent 工作流编排，无需编写复杂代码
- 支持多步骤自动化任务链，实现全流程无人值守
- 与外部 API、数据库和第三方服务无缝集成
- 内置错误处理与自动重试机制，保障稳定运行
- 提供可复用的自动化模板，快速在同类场景部署
- 构建多智能体协作系统的 Agent 开发者
- Agent 任务先做 dry-run 验证工具调用链，再开启自主执行
- API key 直接提交到 git 仓库（请用 .env 并加入 .gitignore）
👥 适合人群
🎯 使用场景
- 自动化日常重复性工作，将精力集中于创造性任务
- 构建数据采集 → 处理 → 输出的完整自动化管线
- 实现跨平台、跨系统的数据流转和业务协同
⚖️ 优点与不足
- +大幅减少重复性人工操作
- +可视化流程，清晰直观
- +可扩展性强，支持复杂场景
- −未明确开源协议，商用场景需谨慎评估
- −初始配置和调试需投入一定时间
- −强依赖外部服务的稳定性
- −复杂场景需具备一定技术基础
该工具未明确声明开源协议，商业使用前请联系原作者确认授权范围，避免侵权风险。
AI Skill Hub 为第三方内容聚合平台，本页面信息基于公开数据整理，不对工具功能和质量作任何法律背书。
建议在沙箱或测试环境中充分验证后，再部署至生产环境，并做好必要的安全评估。
🔗 相关工具推荐
❓ 常见问题 FAQ
总体来看，数学研究代理 是一款质量优秀的Agent工作流，在同类工具中具备一定竞争力。AI Skill Hub 将持续追踪其更新动态，建议收藏备用，结合自身场景选择合适时机引入使用。
| 原始名称 | ResearchMathAgent |
| 原始描述 | 开源AI工作流：Research Math Agents, Official code release for our paper RMA。⭐17 · TeX |
| Topics | aiai-agentsmathscience |
| GitHub | https://github.com/sjtuytc/ResearchMathAgent |
| 语言 | TeX |
收录时间：2026-06-16 · 更新时间：2026-06-16 · License：未公布 · AI Skill Hub 不对第三方内容的准确性作法律背书。
🤖 交给 Agent 安装 · 数学研究代理
选择 Agent 类型，复制安装指令后粘贴到对应客户端