研究人员推出了一种新颖的分阶段程序TREK（Teacher-Routed Exploration via Forward KL），旨在增强语言模型的能力，特别是在复杂的推理任务中。TREK利用蒸馏并非为了直接模仿，而是为了扩展模型的探索支持，使其能够处理当前策略可能 falter 的提示。该方法在应用于Qwen3等模型时，在AIME 2024和AIME 2025等数学推理基准上显示出显著的改进，并且还提高了ALFWorld和ScienceWorld等agentic任务的成功率。 AI
影响 通过改进探索和精炼策略，增强LLM在复杂任务上的推理能力。
排序理由 该集群包含一篇研究论文，详细介绍了一种提高语言模型在推理任务上性能的新方法。
AI 生成摘要 · Google Gemini · 来自 3 个来源。 我们如何撰写摘要 →
TREK expands exploration support for policy optimization by using distillation for exploration rather than imitation, improving performance on challenging mathematical reasoning and agentic tasks.
arXiv:2607.05339v1 Announce Type: cross Abstract: Group Relative Policy Optimization (GRPO) is effective when the current policy already samples useful reasoning trajectories, but it stalls on hard prompts whose correct solution modes lie outside the student's on-policy support. …
Group Relative Policy Optimization (GRPO) is effective when the current policy already samples useful reasoning trajectories, but it stalls on hard prompts whose correct solution modes lie outside the student's on-policy support. We propose TREK (Teacher-Routed Exploration via Fo…