跳转至

SkillReason：针对隐式用户请求的推理增强型智能体技能检索

文章背景与核心概要

在大语言模型（LLM）智能体的实际应用中，如何从庞大的跨领域代码或技能库中检索出合适的工具是一个核心瓶颈。现实中的用户请求往往非常简练且信息不足（隐式请求），用户通常只陈述高级目标，而将必要的执行步骤和所需能力隐去。传统的检索技术在弥合这种高层意图与底层技术能力之间的语义鸿沟时显得力不从心。

为了解决这一痛点，本文提出了 SkillReason-Bench（包含3,729个查询和61,228个技能的跨领域大规模基准测试）以及名为 SkillReason 的全新两阶段训练框架。该框架利用思维链（CoT）推理，通过教师模型引导的监督以及群组相对策略优化（GRPO），将能力推理过程直接内化到查询表示中。最重要的是，它在推理阶段无需进行缓慢的自回归生成，从而在保持极高检索效率的同时，实现了最先进的检索性能。

摘要 (Abstract)

Large language model agents increasingly rely on reusable skills to extend their capabilities beyond parametric knowledge. However, retrieving the appropriate skill from a large-scale library remains challenging because realistic user requests are often concise and underspecified, stating only the task goal while leaving the required capabilities and execution steps implicit. Existing benchmarks provide limited coverage of such requests.

大语言模型智能体越来越依赖可重用的技能来将其能力扩展到参数化知识之外。然而，从大规模库中检索出合适的技能仍然充满挑战，因为现实中的用户请求往往简练且信息不足，通常只说明任务目标，而将所需的权能和执行步骤留作隐式表达。现有的基准测试对这类请求的覆盖范围非常有限。

To address this gap, the authors introduce SkillReason-Bench, a large-scale cross-domain benchmark containing 3,729 queries and a retrieval corpus of 61,228 skills spanning nine domains. They further propose SkillReason, a two-stage framework that uses chain-of-thought reasoning as training-time supervision for skill retrieval: * Stage I: Capability reasoning traces generated by a stronger teacher provide explicit supervision through contrastive learning, retrieval distribution alignment, and language modeling, encouraging the retriever to internalize capability reasoning in its query representation. * Stage II: A retrieval-guided GRPO objective encourages the model to explore reasoning trajectories better suited to its own capabilities and more effective for retrieval.

为了弥补这一空白，作者推出了 SkillReason-Bench，这是一个包含 3,729 个查询和跨越 9 个领域的 61,228 个技能检索语料库的大规模跨领域基准测试。他们进一步提出了 SkillReason，这是一个两阶段框架，利用思维链推理作为技能检索在训练时的监督信号： * 第一阶段： 由更强的教师模型生成的能动推理轨迹，通过对比学习、检索分布对齐和语言建模提供显式监督，鼓励检索器在其查询表示中内化能力推理。 * 第二阶段： 检索引导的 GRPO 目标函数鼓励模型探索更适合自身能力、对检索更有效的推理轨迹。

At inference time, SkillReason directly encodes the original query without autoregressive CoT generation, preserving efficient query-only retrieval. Extensive experiments on SkillReason-Bench, SkillRet, and SRA-Bench show that SkillReason achieves state-of-the-art performance across all three benchmarks, demonstrating that reasoning-enhanced training better bridges the semantic gap between high-level task goals and skill capabilities.

在推理时，SkillReason 直接对原始查询进行编码，无需进行自回归的 CoT 生成，从而保持了高效的“仅查询（query-only）”检索。在 SkillReason-Bench、SkillRet 和 SRA-Bench 上进行的广泛实验表明，SkillReason 在这三个基准测试中均取得了最先进的性能，证明了推理增强训练能够更好地弥合高级任务目标与技能能力之间的语义鸿沟。

框架概览：SkillReason (Framework Overview: SkillReason)

1. SkillReason-Bench 数据集 (SkillReason-Bench Dataset)

  • Scale: 3,729 queries paired with a retrieval corpus of 61,228 skills.
  • Coverage: Spans nine distinct application domains to robustly evaluate cross-domain generalizability.
  • Purpose: Targets implicit user requests where task goals are stated, but necessary capabilities and sequential steps remain unsaid.
  • 规模： 3,729 个查询，配有包含 61,228 个技能的检索语料库。
  • 覆盖范围： 跨越 9 个不同的应用领域，以稳健评估跨领域的泛化能力。
  • 目的： 针对隐式用户请求，即陈述了任务目标，但未说明必要的能力和顺序步骤。

2. 两阶段训练方法 (Two-Stage Training Methodology)

  • Stage I (Teacher-Supervised Internalization): Leverages a stronger teacher model to generate capability reasoning traces. Employs contrastive learning, retrieval distribution alignment, and language modeling to bake reasoning steps directly into the query embedding representation.
  • Stage II (Reinforcement Learning via GRPO): Utilizes a retrieval-guided Group Relative Policy Optimization (GRPO) objective to help the model discover and optimize reasoning trajectories that align with its internal retriever capabilities.
  • 第一阶段（教师监督内化）： 利用更强的教师模型生成能力推理轨迹。采用对比学习、检索分布对齐和语言建模，将推理步骤直接“烘焙”到查询嵌入表示中。
  • 第二阶段（通过 GRPO 进行强化学习）： 利用检索引导的群组相对策略优化（GRPO）目标，帮助模型发现并优化与其内部检索器能力相匹配的推理轨迹。

3. 高效推理 (Efficient Inference)

  • Unlike standard reasoning models that require slow, step-by-step autoregressive generation at test time, SkillReason directly encodes the raw input query. This preserves fast, lightweight query-only retrieval while retaining the semantic benefits of deep reasoning.
  • 与需要在测试时进行缓慢、逐步自回归生成的标准推理模型不同，SkillReason 直接编码原始输入查询。这在保留深度推理语义优势的同时，维持了快速、轻量级的仅查询检索。

实验结果 (Experimental Results)

Evaluated across three rigorous benchmarks—SkillReason-Bench, SkillRet, and SRA-Bench—SkillReason achieves state-of-the-art performance, outperforming existing retrieval models by effectively closing the semantic gap between high-level human intents and technical skill capabilities.

通过在 SkillReason-Bench、SkillRet 和 SRA-Bench 这三个严苛的基准测试中进行评估，SkillReason 实现了最先进的性能，通过有效缩小高级人类意图与技术技能能力之间的语义鸿沟，其表现超越了现有的检索模型。

访问与全文链接 (Access & Full-Text Links)