ARise: Towards Knowledge-Augmented Reasoning via Risk-Adaptive Search

ARise is an advanced reasoning framework that implements Monte Carlo Tree Search (MCTS) for complex question answering tasks. The system decomposes complex questions into manageable sub-questions, retrieves relevant information, and synthesizes a final answer through iterative search and risk-assessment.

🎉🎉 News: Our paper has been accepted to ACL 2025 main conference!

Project homepage: https://opencausalab.github.io/ARise/

Features

• Iterative Refinement through Decomposition: Breaks down complex reasoning tasks into manageable steps, enabling more precise and controlled reasoning processes.
• Retrieval-then-Reasoning Approach: Augments LLMs with fine-grained knowledge retrieval to enhance reasoning capabilities with external information.
• Monte Carlo Tree Search (MCTS): Expands linear reasoning into tree-based exploration, mitigating error propagation by allowing backtracking when necessary.
• Risk-Adaptive Search: Employs Bayesian risk minimization to dynamically evaluate reasoning states and optimize search strategies.
• Dynamic Path Exploration: Enables exploration of multiple reasoning paths simultaneously, focusing computational resources on the most promising directions.

Project Structure

ARise/
├── asset/                  # Project assets
│   ├── pipe.png            # Pipeline diagram
│   ├── com.png             # Comparison diagram
│   └── res.png             # Results visualization
├── src/                    # Core implementation
│   ├── base.py             # Base reasoning functions
│   ├── mcts.py             # Monte Carlo Tree Search implementation
│   ├── node.py             # Tree node definition
│   └── task.py             # Task definition and execution
├── utils/                  # Utility functions
│   ├── extract.py          # Data extraction utilities
│   ├── inference_model.py  # Model inference wrapper
│   ├── prompts.py          # Prompt templates
│   ├── rag.py              # Retrieval-augmented generation
│   ├── value_function.py   # Value functions for MCTS
│   ├── value_model.py      # Value model implementation
│   ├── verify.py           # Verification utilities
│   └── wrap.py             # Prompt wrapping utilities
├── run.py                  # Main execution script
└── nltk_data.zip           # NLTK data package

Configuration

Key parameters can be configured in the MCTSTask class:

• time_limit: Time limit for search in milliseconds
• iteration_limit: Maximum number of iterations
• exploration_constant: UCT exploration constant
• multihops: Number of sub-queries
• total_depth: Total depth of the search tree
• temperature: Sampling temperature
• run_mode: Reasoning strategy (MCTS, zero-shot, etc.)
• value_mode: Value function mode (risk, similarity, etc.)

Evaluation

The system evaluates performance using:

• Exact match accuracy
• F1 score for supporting facts

Citation

If you use ARise in your research, please cite:

@inproceedings{zhang2025arise,
  title   = {ARise: Towards Knowledge-Augmented Reasoning via Risk-Adaptive Search},
  author  = {Yize Zhang and Tianshu Wang and Sirui Chen and Kun Wang and Xingyu Zeng and Hongyu Lin and Xianpei Han and Le Sun and Chaochao Lu},
  booktitle={Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (ACL)},
  year    = {2025}
}