Zhigang Li

Jiaxiang Chen, Song Wang, Zhucong Li et al.

Currently, prompting techniques can be mainly divided into two categories:1)shot method implicitly inspires the model to answer the question by mimicing the steps in the given example, e.g., the few-shot CoT. 2) Guideline method explicitly instructs the model to reason by following guidelines, which contains succinct and concise task-specific knowledge. Shot method is prone to difficulties in terms of selection of shots type, the number of shots, and the design of the reasoning steps, so a question arises: can we only use guideline instead of shot in the prompt? To this end, we propose the FGT framework to automatically learn task-specific guidelines from dataset consisting of Feedback, Guideline, and Tree-gather agents. First, the feedback agent is designed to evaluate the outcomes, both right and wrong, of each Q&A to gather insights guiding more effective optimization strategies. Next, the guideline agent is tasked with deriving guidelines from each piece of feedback and storing them in local memory. Lastly, the tree-gather agent aggregates all guidelines hierarchically through a tree structure, ultimately obtaining all unduplicated guidelines from a global perspective. In addition, we induce the model to generate intermediate processes to ensure the reasoning consistent with the guidelines. Experimental results demonstrate that our approach achieves superior performance across multiple tasks, thereby highlighting the effectiveness of using the guidelines in prompt.

Jiaxiang Chen, Zhuo Wang, Mingxi Zou et al. · microsoft-research

Large language models (LLMs) have advanced general-purpose reasoning, showing strong performance across diverse tasks. However, existing methods often rely on implicit exploration, where the model follows stochastic and unguided reasoning paths-like walking without a map. This leads to unstable reasoning paths, lack of error correction, and limited learning from past experience. To address these issues, we propose a framework that shifts from implicit exploration to structured reasoning through guideline and refinement. First, we extract structured reasoning patterns from successful trajectories and reflective signals from failures. During inference, the model follows these guidelines step-by-step, with refinement applied after each step to correct errors and stabilize the reasoning process. Experiments on BBH and four additional benchmarks (GSM8K, MATH-500, MBPP, HumanEval) show that our method consistently outperforms strong baselines across diverse reasoning tasks. Structured reasoning with stepwise execution and refinement improves stability and generalization, while guidelines transfer well across domains and flexibly support cross-model collaboration, matching or surpassing supervised fine-tuning in effectiveness and scalability.