[{"data":1,"prerenderedAt":51},["Reactive",2],{"news-story-news-1792366853834-s8z1w":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":28,"category":35,"impact_score":36,"views_count":37,"likes_count":38,"related_entities":39,"tags":43,"image_url":50},"news-1792366853834-s8z1w","浙江大学与腾讯开源深搜智能体对齐算法 KnowGPO：原子知识状态分组强化学习，8B 模型跨级超越 32B 开源系统并在 GAIA 斩获 40.2 分","Zhejiang University & Tencent Open-Source KnowGPO for Deep Search Agents: Knowledge-Grouped RL Outperforms 32B Models, Scoring 40.2 on GAIA","深度搜索智能体在长链路多轮检索中面临严重的信贷分配（Credit Assignment）难题。浙江大学与腾讯联合开源 KnowGPO（arXiv:2610.22331，GitHub: ZJU-REAL/KnowGPO）。该算法将复杂问答拆解为原子知识条件集合，以知识状态对齐代替杂乱的表面文本对齐，并在同状态组内计算增量知识过程奖励。微调后的 Qwen3-8B 在 BrowseComp-ZH（38.8%）与 GAIA（40.2%）上全面超越 GRPO 与 GiGPO，越级战胜 32B 级开源深搜基线并逼近闭源模型。","Deep search agents struggle with credit assignment across long, noisy retrieved contexts. Zhejiang University and Tencent open-sourced KnowGPO (arXiv:2610.22331, GitHub: ZJU-REAL/KnowGPO), an RL framework that groups search trajectories by atomic knowledge conditions rather than surface observations. Qwen3-8B fine-tuned with KnowGPO achieves 38.8% on BrowseComp-ZH and 40.2% on GAIA, outperforming GRPO/GiGPO baselines and rivaling 32B open-source search systems.",[10,11,12,13],"浙江大学与腾讯联合开源 KnowGPO（arXiv:2610.22331），解决长程深搜智能体中网页噪声大、过程信贷分配失效的核心痛点","以原子知识条件替代表面文本状态，通过同知识状态组内相对优势与增量知识过程奖励实现鲁棒对齐","Qwen3-8B 接入 KnowGPO 后在 BrowseComp-ZH 达到 38.8%、GAIA 达到 40.2%，较标准 GRPO 绝对提升 5.9 / 5.2 个百分点","8B 模型性能跨级超越 WebSailor-32B、ASearcher-Web-32B 等开源 32B 系统，逼近 DeepMiner-32B 与商业闭源模型",[15,16,17,18],"Zhejiang University and Tencent open-sourced KnowGPO (arXiv:2610.22331) for credit assignment in deep search agents","Replaces surface observation matching with atomic knowledge condition states and within-group incremental knowledge rewards","Qwen3-8B + KnowGPO scores 38.8% on BrowseComp-ZH and 40.2% on GAIA, outperforming standard GRPO by +5.9 and +5.2 points","Allows an 8B model to rival and outperform 32B open-source search systems including WebSailor-32B and ASearcher-Web-32B","### 核心背景与行业痛点\n在多轮网络深度搜索与知识密集型推理任务中，基于终局结果奖励的强化学习（如 `GRPO`）只能提供单一的终端正负反馈，无法对长链路搜索过程中的单步决策（如查询改写、页面点击、无用结果跳过）进行准确的**信贷分配（Credit Assignment）**。现有的过程监督方法（如 `GiGPO`、`IGPO`）试图通过轨迹表层文本状态或模型内部隐层状态分配过程奖励，但在多轮搜索场景中，检索到的网页内容极度异构且充满噪声，导致表层匹配彻底失效。为此，浙江大学（ZJU）与腾讯联合提出了 **`KnowGPO`（Knowledge-Grouped Policy Optimization，`arXiv:2610.22331`）**。\n\n### 架构亮点与底层机制\n1. **原子知识条件状态化拆解（Atomic Knowledge Decomposition）**：\n   - `KnowGPO` 摒弃表层文本比对，对于每个复杂查询，算法将回答所需的逻辑链条离散分解为一组原子知识条件（Atomic Knowledge Conditions）。在每一步搜索中，智能体轨迹被抽象为其当前已满足的知识子集向量；\n2. **组内知识增量环境过程奖励（Knowledge-Grouped Reinforcement Learning）**：\n   - 处于相同知识状态的搜索轨迹被自动聚合为同一个对比组，在组内进行相对优势计算；将连续两步之间新增解锁的原子知识增量作为环境接地的客观过程奖励（Environment-Grounded Process Reward），彻底规避了奖励黑客与噪声文档干扰。\n\n### 权威 Benchmark 与实测跑分对比\n- **`BrowseComp-ZH` 与 `GAIA` 深度评测（Table 1）**：\n  - 基于 `Qwen3-8B-SFT`（初始得分 30.1% / 31.1%）基座：\n  - `+ GRPO` 提升至 32.9% / 35.0%；`+ GiGPO` 达到 37.3% / 35.0%；`+ IGPO` 达到 32.2% / 37.9%；\n  - **`+ KnowGPO` 分数飙升至 38.8%（BrowseComp-ZH）与 40.2%（GAIA）**，在强化学习方法中斩获最佳，较终局奖励 GRPO 绝对提升 **+5.9 / +5.2 个百分点**；\n  - 8B 规模的 `KnowGPO` 全面击败了 32B 规模的开源系统 `WebSailor-32B`（25.5% / 51.5%）与 `ASearcher-Web-32B`（15.6% / 52.8%），在 `BrowseComp-ZH` 上逼近 32B `DeepMiner-32B-RL`（40.1%）与闭源商业模型 `GLM-4.5`（37.5%）。\n- **`Xbench` 挑战基准（Table 2）**：\n  - 在 `Xbench-2505` 与 `Xbench-2510` 上分别达到 **48.0%** 与 **36.0%**，持续领跑所有同参数开源 RL 对齐模型。\n\n### 开发者实战落地与开箱指南\n- **官方开源仓库**：项目代码已在 GitHub 开源（`https://github.com/ZJU-REAL/KnowGPO`）。\n- **深搜 Agent 训练落地**：开发者在训练搜索类 Agent 时，可引入原子知识分解器作为轻量级离线标注模块，将多轮检索日志转化为知识状态转移图；在 PPO/GRPO 训练循环中将默认的终端 Outcome Reward 替换为 KnowGPO 知识组内优势函数，即可在 8B 级小模型上复现顶尖的检索决策密度与抗噪能力。","### 核心背景与行业痛点\nDeep search agents solve knowledge-intensive tasks through long-horizon, multi-turn interactions where assigning credit to individual search decisions is notoriously difficult. Outcome-based reinforcement learning (such as `GRPO`) provides only a single scalar terminal reward for an entire trajectory. Recent process-supervision methods (`GiGPO`, `IGPO`) attempt fine-grained credit assignment, but anchor rewards to surface-level text or model activations—both of which fail in deep search where retrieved documents are noisy and lengthy. Researchers from Zhejiang University and Tencent introduced **`KnowGPO` (`arXiv:2610.22331`, GitHub: `ZJU-REAL/KnowGPO`)** to group trajectories by objective knowledge acquisition.\n\n### 架构亮点与底层机制\n1. **Atomic Knowledge State Decomposition**:\n   - For each query, `KnowGPO` decomposes the target answer derivation into a set of atomic knowledge conditions. At each step, an agent's state is represented by the subset of satisfied conditions rather than raw textual observations;\n2. **Group-Relative RL on Incremental Knowledge Gains**:\n   - Trajectories sharing the same knowledge state are clustered into groups for within-group advantage computation. The incremental knowledge gained between consecutive steps acts as an environment-grounded process reward, effectively eliminating reward hacking caused by noisy search snippets.\n\n### 权威 Benchmark 与实测跑分对比\n- **`BrowseComp-ZH` and `GAIA` Benchmarks (Table 1)**:\n  - Starting from `Qwen3-8B-SFT` (30.1% / 31.1%):\n  - `+ GRPO` reaches 32.9% / 35.0%; `+ GiGPO` achieves 37.3% / 35.0%; `+ IGPO` reaches 32.2% / 37.9%;\n  - **`+ KnowGPO` reaches 38.8% on `BrowseComp-ZH` and 40.2% on `GAIA`**, outperforming all RL baselines and gaining **+5.9 / +5.2 percentage points** over standard GRPO;\n  - The 8B `KnowGPO` agent surpasses 32B systems like `WebSailor-32B` (25.5%) and `ASearcher-Web-32B` (15.6%), closely rivaling `DeepMiner-32B-RL` (40.1%) and proprietary `GLM-4.5` (37.5%).\n- **`Xbench` Leaderboard (Table 2)**:\n  - Achieves **48.0%** on `Xbench-2505` and **36.0%** on `Xbench-2510`, setting new state-of-the-art results among open-source 8B search agents.\n\n### 开发者实战落地与开箱指南\n- **Open-Source Repository**: Code and implementation details are available at `https://github.com/ZJU-REAL/KnowGPO`.\n- **Engineering Takeaways**: When post-training search and research agents, developers can integrate atomic knowledge condition evaluators into the RL loop, substituting outcome-only signals with KnowGPO's state-grouped advantage estimates to dramatically improve tool-search efficiency on compact 8B models.",[22,25],{"title":23,"url":23,"type":24},"https://github.com/ZJU-REAL/KnowGPO","github",{"title":26,"url":26,"type":27},"https://arxiv.org/abs/2610.22331","arxiv",{"platform":29,"author_name":30,"author_handle":31,"author_avatar":32,"tweet_url":33,"published_at":34},"twitter","Zhejiang University & Tencent","@huggingface","https://unavatar.io/x/huggingface","","2026-10-18T23:40:53.834Z","OpenSource",92,59,0,{"models":40,"agents":41,"vendors":42},[],[],[],[44,45,46,47,48,49],"KnowGPO","Search Agents","Tencent","Zhejiang University","RL","GRPO","https://arxiv.org/html/2610.22331v1/intro_v0.png",1792905127970]