📄 中文摘要
大型语言模型（LLMs）在复杂推理方面取得了显著进展，但思维树（ToT）框架面临探索深度与计算效率之间的关键权衡。现有的ToT实现通常依赖于重型的基于LLM的自我评估或严格的启发式方法进行分支剪枝，这使得其在广泛应用中显得成本高昂且灵活性不足。为了解决这一问题，提出了一种可适应的即插即用预测器DST，作为轻量级的监督启发式方法来指导ToT搜索过程。该预测器实现了动态的、上下文感知的剪枝，使得在简单推理步骤上搜索能够接近贪婪效率，同时在复杂情境下自适应扩展搜索范围。
领域专用的思维树通过即插即用预测器
出处: Domain-Specialized Tree of Thought through Plug-and-Play Predictors
发布: 2026年3月24日
📄 English Summary
Domain-Specialized Tree of Thought through Plug-and-Play Predictors
Large Language Models (LLMs) have made significant strides in complex reasoning; however, the Tree of Thoughts (ToT) framework encounters a critical trade-off between exploration depth and computational efficiency. Existing ToT implementations often depend on heavyweight LLM-based self-evaluation or rigid heuristics for branch pruning, rendering them prohibitively expensive and inflexible for broader applications. To address this, an adaptable, plug-and-play predictor, DST, is introduced as a lightweight, supervised heuristic to guide the ToT search process. This predictor enables dynamic, context-aware pruning, allowing the search to proceed with near-greedy efficiency on simpler reasoning steps while adaptively expanding the search beam in more complex scenarios.