Claude Desktop용 추론 구현으로, 빔 검색과 몬테카를로 트리 검색(MCTS) 모두를 사용할 수 있게 해줍니다. 솔직히 말해서, 이 방법은 Claude를 복잡한 문제 해결에 더 잘할 수 있는지 알아보기 위한 시도였습니다… 알고 보니 확실히 가능하다는 것이 밝혀졌습니다.

현재 버전:

v2.0.0

새로운 소식:

두 가지 실험적 추론 알고리즘을 추가했습니다:

• mcts-002-alpha

• A* 검색 기법과 초기 알파 정책 시뮬레이션 계층 구현을 함께 사용합니다

• 적응형 탐사 시뮬레이터와 결과 기반 추론 시뮬레이터의 초기 알파 구현도 포함

참고 이 알파 시뮬레이터의 구현은 아직 완성되지 않았으며 변경될 수 있습니다

• mcts-002alt-alpha

• 양방향 검색 기법과 초기 알파 정책 시뮬레이션 계층 구현을 함께 사용합니다

• 적응형 탐사 시뮬레이터와 결과 기반 추론 시뮬레이터의 초기 알파 구현도 포함

참고 이 알파 시뮬레이터의 구현은 아직 완성되지 않았으며 변경될 수 있습니다

mcts-001-alpha와 mcts-001alt-alpha는 어떻게 된 거야?

간단히 말해: 쓸모가 없었고 기본 mcts 방식과 거의 비슷했습니다.

MCP Reasoner

A reasoning implementation for Claude Desktop that lets you use both Beam Search and Monte Carlo Tree Search (MCTS). tbh this started as a way to see if we could make Claude even better at complex problem-solving… turns out we definitely can.

Current Version:

v2.0.0

What’s New:

Added 2 Experimental Reasoning Algorithms:

- `mcts-002-alpha`

    - Uses the A* Search Method along with an early *alpha* implementation of a Policy Simulation Layer

    - Also includes an early *alpha* implementation of Adaptive Exploration Simulator & Outcome Based Reasoning Simulator

*NOTE* the implementation of these alpha simulators is not complete and is subject to change

- `mcts-002alt-alpha`

    - Uses the Bidirectional Search Method along with an early *alpha* implementation of a Policy Simulation Layer

    - Also includes an early *alpha* implementation of Adaptive Exploration Simulator & Outcome Based Reasoning Simulator

*NOTE* the implementation of these alpha simulators is not complete and is subject to change

What happened to mcts-001-alpha and mcts-001alt-alpha?

Quite simply: It was useless and near similar to the base mcts method. After initial testing the results yielded in basic thought processes was near similar showing that simply adding policy simulation may not have an effect.

So why add Polciy Simulation Layer now?

Well i think its important to incorporate Policy AND Search in tandem as that is how most of the algorithms implement them.

Previous Versions:

v1.1.0

Added model control over search parameters:

beamWidth - lets Claude adjust how many paths to track (1-10)

numSimulations - fine-tune MCTS simulation count (1-150)

Features

• Two search strategies that you can switch between:
  • Beam search (good for straightforward stuff)
  • MCTS (when stuff gets complex) with alpha variations (see above)
• Tracks how good different reasoning paths are
• Maps out all the different ways Claude thinks through problems
• Analyzes how the reasoning process went
• Follows the MCP protocol (obviously)

Installation

git clone https://github.com/frgmt0/mcp-reasoner.git 

OR clone the original:

git clone https://github.com/Jacck/mcp-reasoner.git

cd mcp-reasoner
npm install
npm run build

Configuration

Add to Claude Desktop config:

{
  "mcpServers": {
    "mcp-reasoner": {
      "command": "node",
      "args": ["path/to/mcp-reasoner/dist/index.js"],
    }
  }
}

Testing

[More Testing Coming Soon]

Benchmarks

[Benchmarking will be added soon]

Key Benchmarks to test against:

• MATH500

• GPQA-Diamond

• GMSK8

• Maybe Polyglot &/or SWE-Bench

License

This project is licensed under the MIT License - see the LICENSE file for details.