A*-Thought: Efficient Reasoning for Low-Resource Settings
xxang/AStar-Thought-QwQ-32B is a 32.8 billion parameter model that leverages the novel A*-Thought framework to enhance reasoning efficiency and performance, especially in environments with limited computational resources. This framework employs a bidirectional compression mechanism to distill complex reasoning chains into compact, effective paths.
Key Capabilities
- Bidirectional Importance Estimation: Quantifies the significance of each thinking step based on its relevance to both the question and the potential solution.
- A Search for Path Optimization:* Efficiently navigates the search space using cost functions that evaluate path quality and conditional self-information of the solution.
- Improved Accuracy and Efficiency: Demonstrates up to 2.39x accuracy and 2.49x ACU (Accuracy-Cost-Utility) improvements in low-budget scenarios (e.g., 512-token inference budget).
- Significant Length Reduction: Achieves up to 33.59% response length reduction without substantial accuracy loss in higher budget settings (e.g., 4096-token budget).
- Generalizability: The A*-Thought framework is compatible with various backbone models, consistently achieving high ACU scores across different budget conditions.
Good For
- Applications requiring efficient reasoning in low-resource or budget-constrained environments.
- Tasks where compact and effective reasoning paths are critical.
- Reducing inference costs and response lengths while maintaining high accuracy.