[{"data":1,"prerenderedAt":48},["Reactive",2],{"news-story-news-1792366853834-do5vn":3},{"id":4,"title_zh":5,"title_en":6,"summary_zh":7,"summary_en":8,"takeaways_zh":9,"takeaways_en":14,"content_zh":19,"content_en":20,"external_links":21,"source":25,"category":32,"impact_score":33,"views_count":34,"likes_count":35,"related_entities":36,"tags":40,"image_url":47},"news-1792366853834-do5vn","新加坡国立与阿里通义千问提出智能体实时纠偏框架 Steering Agent：双路径强化学习突破目标漂移，多跳 QA 提升 7.4 分且执行轮次缩减 26%","NUS, USTC, and Alibaba Qwen Introduce Steering Agent: Dual-Path RL Framework Solves Goal Drift, Boosting Multi-Hop QA by +7.4 Points with 26% Fewer Turns","针对长链路多轮工具调用与多智能体协作中频发的“目标漂移（Goal Drift）”难题，新加坡国立大学、中科大与阿里通义千问团队联合提出 Steering Agent（arXiv:2610.22316）。该框架通过成对排序与组相对策略优化（GRPO）联合训练低成本快速判断路径（When to intervene）与高成本慢速纠偏生成路径（How to intervene），在无需重训基座模型的前提下将多跳 QA 平均基准分从 42.70 提升至 50.10，同时将平均执行轮次从 8.8 轮大幅压缩至 6.5 轮（-26.1%）。","To combat goal drift in long-horizon agent tool-use and multi-agent systems, NUS, USTC, and Alibaba Qwen introduced Steering Agent (arXiv:2610.22316). Jointly trained via pairwise ranking and Group-Relative Policy Optimization (GRPO), the agent operates a low-cost fast path to detect drift and a slow path to generate grounded interventions. It raises Multi-Hop QA benchmark performance from 42.70 to 50.10 while cutting executor turns from 8.8 to 6.5 (-26.1%) without modifying the underlying executor models.",[10,11,12,13],"新加坡国立大学与阿里通义千问团队提出 Steering Agent（arXiv:2610.22316），专攻多轮工具调用与多智能体系统中的目标漂移（Goal Drift）难题","创新设计“快速路径判断何时干预 + 慢速路径生成如何纠偏”的双路径架构，通过成对排序与 GRPO 强化学习联合训练","在多跳问答基准上将平均准确率提升至 50.10（超越无干预基线 7.4 分），并将规划执行轮次从 8.8 轮压减至 6.5 轮（-26.1%）","在挂载真实网络搜索的 GAIA（Pass@4 达 39.81%）与 Frames（Pass@4 达 59.71%）基准上展现出优异的跨模型与跨任务泛化能力",[15,16,17,18],"NUS, USTC, and Alibaba Qwen introduced Steering Agent (arXiv:2610.22316) to tackle goal drift in long-horizon LLM agent workflows","Features a dual-path architecture combining a low-cost fast path (when to intervene) with a slow path (how to intervene), trained via pairwise ranking and GRPO","Raises average Multi-Hop QA accuracy to 50.10 (+7.4 pts over unsteered baseline) while reducing executor turns from 8.8 to 6.5 (-26.1%)","Demonstrates robust out-of-domain transfer on live search benchmarks, reaching 59.71% Pass@4 on Frames and 39.81% Pass@4 on GAIA","### 核心背景与行业痛点\n随着 LLM 智能体与多智能体系统（MAS）在复杂软件工程与深度网络搜索中执行越来越长的工作流，系统在多轮工具调用与代理间交互中极易发生**“目标漂移（Goal Drift）”**——随着上下文膨胀，智能体逐渐偏离用户原始意图，陷入无效工具试错或发散性子任务中，造成严重的 Token 算力浪费与任务失败。现有的静态规划（Planning）或被动验证（Verification）方案通常开销巨大，且无法在执行过程中自适应调节干预时机。为此，新加坡国立大学（NUS）、中国科学技术大学与阿里巴巴通义千问团队联合发表了 **`Steering Agent`（`arXiv:2610.22316`）**，构建了一套在智能体运行期实现动态非侵入式纠偏的监督控制新范式。\n\n### 架构亮点与底层机制\n1. **快慢双路径协同决策（Fast-Slow Dual-Path Architecture）**：\n   - **快速路径（Fast Path, Low-Cost）**：轻量级监控当前执行上下文，仅输出二元决策以判定系统是否偏离全局目标，决定“何时介入（When to intervene）”；\n   - **慢速路径（Slow Path, High-Cost）**：仅在快速路径发出漂移警报时被激活，基于完整上下文进行深度反思与指令重整，生成具体的动作修正建议，决定“如何纠偏（How to intervene）”。\n2. **成对排序与组相对策略优化（GRPO）联合强化学习**：\n   - 团队利用成对排序目标训练快速路径对漂移状态的判别敏感度，并结合 **Group-Relative Policy Optimization (GRPO)** 联合优化慢速路径的干预指令生成策略，使 Steering Agent 具备极强的跨域泛化能力与可控干预频率，**完全无需重训底层的执行器智能体**。\n\n### 权威 Benchmark 与实测跑分对比\n- **域内多跳问答基准（In-Domain Benchmarks, Table 1）**：\n  - 在无干预基线（No Intervention）仅获得 **42.70 分**（执行耗时 6.4 轮）的对照下，静态规划经强化学习后达 45.91 分（8.8 轮）；\n  - **`Steering w/ RL` 斩获 50.10 分（+7.40 分提升）**，且平均执行轮次从规划方案的 8.8 轮大幅缩减至 **6.5 轮（轮次开销降低 26.1%）**；在 `2Wiki`（**50.1** vs 36.9，**+13.2 pts**）与 `HotpotQA`（**53.7** vs 46.9，**+6.8 pts**）上均实现大幅领先。\n- **域外在线搜索与真实 Agent 评测（Out-of-Domain Benchmarks, Table 2）**：\n  - 在挂载真实搜索 API 的复杂评测中，在 `Frames` 基准上 `Avg@4` 从 26.46% 跃升至 **36.32%（+9.86 pts）**，`Pass@4` 从 44.17% 升至 **59.71%（+15.54 pts）**；\n  - 在通用智能体基准 `GAIA` 上，`Avg@4` 从 16.75% 提升至 **23.79%（+7.04 pts）**，`Pass@4` 达到 **39.81%**；在 `xBench-DS` 上 `Pass@4` 达到 **72.00%**（相比无干预 65.00%）。\n\n### 开发者实战落地与开箱指南\n在企业级 Coding Agent 或复杂多智能体流程（如 Multi-Agent Code Review & Debugging）中，开发者可通过非侵入式 Sidecar 模式部署 Steering Agent：\n1. 将快速路径作为轻量级中间件插入智能体 Event Loop，每轮工具调用后评估漂移概率；\n2. 仅当漂移置信度越过预设阈值时，触发慢速路径向执行智能体注入系统级重新聚焦指令（Refocus Prompt），避免无效的死循环与上下文溢出。","### 核心背景与行业痛点\nAs LLM-based autonomous agents and multi-agent systems (MAS) execute increasingly complex, long-horizon software engineering and web workflows, they frequently suffer from **goal drift**—gradually straying from the user's primary objective through recursive tool invocations and subtask proliferation. Existing remedies like static pre-planning or post-hoc verification introduce heavy latency overheads and cannot dynamically modulate intervention frequency. To resolve this, researchers from the National University of Singapore (NUS), USTC, and Alibaba's Qwen Business Unit introduced **`Steering Agent` (`arXiv:2610.22316`)**, an online runtime supervisory framework that detects drift and steers agents back on track.\n\n### 架构亮点与底层机制\n1. **Dual-Path Fast/Slow Steering Architecture**:\n   - **Fast Path (Low-Cost)**: Continuously observes the interaction context to make lightweight binary decisions on whether the system has drifted, deciding *when to intervene*;\n   - **Slow Path (High-Cost)**: Triggered selectively upon drift alerts to diagnose failure points and generate grounded redirect instructions, determining *how to intervene*.\n2. **Joint Reinforcement Learning via Pairwise Ranking & GRPO**:\n   - The fast path is trained using pairwise ranking objectives to maximize drift sensitivity, while the slow path is optimized via Group-Relative Policy Optimization (GRPO) to synthesize concise, high-efficacy corrective prompts. This decouples supervisory intelligence from the executor, enabling seamless steering of black-box LLMs without executor retraining.\n\n### 权威 Benchmark 与实测跑分对比\n- **In-Domain Multi-Hop QA Benchmarks (Table 1)**:\n  - Compared to the unsteered baseline (**42.70** avg accuracy, 6.4 turns) and static RL planning (45.91, 8.8 turns), **`Steering w/ RL` achieves 50.10 avg accuracy (+7.40 points gain)** while cutting executor turns from 8.8 to **6.5 turns (-26.1% reduction)**;\n  - Delivers marked gains on `2Wiki` (**50.1** vs 36.9, **+13.2 pts**) and `HotpotQA` (**53.7** vs 46.9, **+6.8 pts**).\n- **Out-of-Domain Online Search Benchmarks (Table 2)**:\n  - On the `Frames` benchmark with live search APIs, lifts `Avg@4` from 26.46% to **36.32% (+9.86 pts)** and `Pass@4` from 44.17% to **59.71% (+15.54 pts)**;\n  - On `GAIA`, boosts `Avg@4` from 16.75% to **23.79% (+7.04 pts)** and `Pass@4` to **39.81%**; reaches **72.00% Pass@4** on `xBench-DS` (vs 65.00%).\n\n### 开发者实战落地与开箱指南\nIn production multi-agent architectures (such as autonomous repository refactoring or deep research loops):\n1. Deploy the fast-path steering classifier as an event-driven sidecar filter between agent tool turns to monitor trajectory drift scores;\n2. Configure dynamic intervention thresholds so that slow-path redirection prompts are only generated when drift probability exceeds tolerance, pruning unproductive execution branches without inflating token costs.",[22],{"title":23,"url":23,"type":24},"https://arxiv.org/abs/2610.22316","arxiv",{"platform":26,"author_name":27,"author_handle":28,"author_avatar":29,"tweet_url":30,"published_at":31},"twitter","NUS & Alibaba Qwen / USTC","@Alibaba_Qwen","https://unavatar.io/x/Alibaba_Qwen","","2026-10-18T23:40:53.833Z","Agent",93,38,0,{"models":37,"agents":38,"vendors":39},[],[],[],[41,42,43,44,45,46],"Steering Agent","Multi-Agent","Alibaba Qwen","GRPO","Goal Drift","Reinforcement Learning","https://arxiv.org/html/2610.22316v1/Steering-Method-v2.svg",1792905128647]