陈定伟
, 宗泽方
, 马志鹏
, 罗利奥
, 李阳
李成明
陈鹏
蒋杰
腾讯公司
深圳北理莫斯科大学
cuso4cdw@gmail.com
licm@smbu.edu.cn
{willzong,thomasyngli}@tencent.com
本工作完成于腾讯公司实习期间。通讯作者。
摘要
基于智能体的强化学习已被证明能有效训练基于大语言模型的智能体,使其具备外部工具使用能力。然而,我们发现智能体强化学习训练会导致冗余工具调用增加,并模糊模型的内在知识边界——模型难以区分何时需要调用工具、何时仅凭参数化知识即可作答。现有基于奖励塑形的解决方案会生成粗粒度的优化目标,往往倾向于不加区分地抑制工具调用,从而引发奖励作弊行为。本文提出AKBE(智能体知识边界增强),一种在训练过程中通过双路径(带工具与不带工具)轨迹展开来动态探测模型内在知识边界的在策略方法。我们将知识边界定义为:针对每个实例判断是否需要工具,以及所需的最少工具调用次数。通过比较不同路径的正确性,AKBE对轨迹进行分类,并构建针对性的监督信号,引导每个问题采用高效的工具使用模式。这些信号无缝集成到智能体强化学习训练循环中。在七个问答基准上的实验表明,与标准智能体强化学习相比,AKBE平均提升了任务准确率,并将工具调用次数减少了18%,在无需权衡准确率与效率的前提下,实现了25%的工具生产力提升。进一步分析表明,该方法在不同强化学习算法间具有即插即用的兼容性,并揭示了各类信号类别的作用机制。我们的代码已开源:https://github.com/CuSO4-Chen/AKBE。
基于在策略内在知识边界增强的高效智能体强化学习
丁伟晨♠◇††致谢:本工作完成于腾讯实习期间,宗泽方♠,马志鹏♠,罗利奥♠,李阳♠,李成明,陈鹏♠,蒋杰♠22脚注标记:2 ♠腾讯 ◇香港中文大学 ♡深圳北理莫斯科大学 cuso4cdw@gmail.com, licm@smbu.edu.cn {willzong,thomasyngli}@tencent.com
1 引言
大语言模型智能体通过将内部推理与外部工具交互相结合(Yao等人,2023;Schick等人,2023;Si等人,2026;Luo等人,2026),在解决复杂任务方面展现出卓越能力。通过使用搜索引擎和代码解释器等工具,这些智能体将其推理能力扩展至参数化知识之外。近年来,强化学习已成为一种强大的后训练范式,用于进一步增强智能体能力,其中GRPO(Shao等人,2024)、DAPO(Yu等人,2025)以及专门的智能体强化学习算法(Feng等人,2025;Dong等人,2025;Zong等人,2026)在工具增强推理基准测试上取得了显著改进。
然而,智能体强化学习训练中一个关键但尚未充分探索的副作用是:当模型被优化以增强借助工具进行推理的能力时,它会越来越多地产生冗余的工具调用——要么在参数化知识足够时仍调用工具,要么在少量调用即可满足需求时过度调用,这被定义为认知卸载(Wang et al., 2025; Xie et al., 2026)。如图1所示,这表现为训练过程中工具调用数量的持续增长。这种对工具调用的过度依赖存在两个问题:(1)它浪费了计算资源并增加了推理延迟;(2)不必要的工具调用可能会引入噪声,用误导性的检索信息覆盖正确的内部推理,从而导致答案质量下降。
现有的高效智能体强化学习方法主要通过奖励塑形来解决这个问题,即将工具调用模式纳入奖励函数中(Wang et al., 2025; Wu et al., 2025b)。然而,直接将工具调用行为与奖励信号耦合会形成一个粗粒度的优化目标。这会激励模型减少整体工具使用量以获取额外奖励,而不考虑具体调用是否必要,从而导致奖励破解和任务准确率下降。更根本的是,这种基于奖励的方法无法捕捉每个实例中必要调用与冗余调用之间的区别,也无法适应训练过程中模型知识边界的动态演变。
在本文中,我们提出了 AKBE(智能体知识边界增强),这是一种在策略方法,通过在训练过程中显式探测模型的内在知识边界来解决这一局限性。我们将知识边界定义为:针对每个实例,判断是否需要外部工具,以及在需要时,为得出正确答案所需的最少工具调用次数,这代表了每个问题最有效的工具调用模式。关键思路在于,对于训练批次中的每个问题,我们分别执行带外部工具和不带外部工具的双路径 rollout。通过比较这两条路径的正确性,我们能够判断该问题是处于模型的参数化知识范围内,还是确实需要外部工具调用,并进一步确定在后者情况下所需的最少工具使用量。基于此识别结果,AKBE 对每个问题进行分类并构建针对性的监督信号:工具依赖型选择最少工具调用的正确轨迹以强化高效工具使用,效率型选择无工具的正确轨迹以消除冗余调用,幻觉型选择无工具的正确轨迹以减轻有害的工具依赖,而两者皆错型则不提供信号,仅依赖 RL 目标。这些由知识边界引导的信号,作为辅助的在策略训练损失,与标准 RL 目标无缝集成到训练循环中,在不修改 RL 奖励或优化过程的前提下,提供细粒度的实例级指导。我们的贡献总结如下:
-
我们提出了 AKBE,一种用于高效智能体 RL 的在策略知识边界增强方法。该方法通过双路径 rollout 动态探测模型的内在知识边界,并构建由边界引导的监督信号,以消除冗余工具调用并强化高效的工具使用模式。
-
我们在两个骨干模型上,针对七个问答基准进行了大量实验,结果表明,与标准智能体强化学习相比,AKBE 平均提升了任务准确率,并将工具调用次数减少了 18%,从而使工具生产率提高了 25%。在大多数情况下,它的表现优于基线方法,且无需在准确率与效率之间进行任何权衡。
-
我们进一步证明,AKBE 作为一个即插即用模块,能够兼容多种不同的智能体强化学习算法,并揭示了模型的知识边界在训练过程中会动态演化,其中每个信号类别会自然地适应,以解决工具使用行为中一种特定的失败模式。
2 相关工作
近期研究将强化学习应用于训练具备外部工具使用能力的大语言模型智能体(Shao 等人,2024;Yu 等人,2025;Zheng 等人,2025)。此外,一系列工作设计了针对智能体场景的专用算法,例如基于熵的展开和信用分配(Jin 等人,2025;Dong 等人,2025;Ji 等人,2025;Zong 等人,2026;Chen 等人,2026)。然而,这些方法在训练过程中均表现出工具调用冗余不断增加的问题(Xie 等人,2026)。为缓解这一问题,OTC-PO(Wang 等人,2025)引入了工具生产力奖励项,-GRPO(Wu 等人,2025b)加入了置信度阈值,HiPRAG(Wu 等人,2025a)则应用分层过程奖励来评估每一步的工具调用。然而,这些基于奖励的方法要么对整体工具调用行为施加粗粒度的惩罚,导致智能体总是学会不加区分地减少工具调用以获取额外奖励,从而引发奖励破解;要么单独评估每一步的工具调用,但依赖外部模型或 API(Wu 等人,2025a),引入了额外的开销和依赖。SMART(Qian 等人,2025)则离线构建元认知 SFT 数据,但静态数据集无法追踪强化学习训练过程中不断变化的知识边界。与这些方法不同,我们提出的 AKBE 在强化学习训练循环内部运行,通过在线策略的双路径(带工具和不带工具)展开动态探测模型的内在知识边界,从而构建边界引导的监督信号,该信号可作为即插即用模块无缝集成到任何智能体强化学习算法中。
3 预备知识
3.1 任务定义
我们考虑一个智能体场景,其中语言模型策略迭代地与外部工具环境交互,以回答给定的问题。遵循 ReAct 范式(Yao 等人,2023),智能体生成一系列推理与行动交替的步骤。在每一步,智能体根据当前上下文生成一个思考和一个行动。该行动要么是调用外部工具(工具会返回一个观察结果并附加到上下文中),要么是一个结束行动,用于终止本轮交互并返回最终答案。一次完整的交互因此形成一条轨迹,其中 表示最终步骤。根据最终答案是否与真实答案匹配,会分配一个结果奖励。学习目标是最大化训练分布上的期望奖励:
| (1) |
3.2 智能体强化学习
虽然 PPO(Schulman 等人,2017)提供了一个通用的策略优化框架,但它对独立价值评估器的依赖引入了大量的内存和训练开销。GRPO(Shao 等人,2024)通过引入组相对优势解决了这一问题,并已成为近期智能体强化学习研究中的主流算法(Jin 等人,2025;Dong 等人,2025;Ji 等人,2025)。
具体来说,对于每个问题,GRPO 从当前策略中采样一组轨迹,并计算组相对优势:
| (2) |
策略通过最大化带有 KL 正则化项的裁剪策略目标函数来更新:
| (3) |
其中 是重要性采样比率, 是裁剪阈值, 控制相对于参考策略 的 KL 正则化强度。请注意,来自工具观察结果的 token 在训练期间会被屏蔽。
4 方法
在本节中,我们提出 AKBE,该方法通过从双路径展开中获取的、以知识边界为指导的训练信号来增强智能体强化学习目标。通过探测模型对每个问题是否需要外部工具以及最少需要多少次调用,AKBE 选择高效的轨迹作为有针对性的在线策略优化信号,从而消除冗余的工具调用,同时在确实需要外部工具的场景下强化高效的工具使用。我们在图 2 中展示了该框架,并在算法 1 中详细说明了训练流程。
4.1 双路径轨迹展开
对于训练批次中的每个问题,AKBE 并行执行双路径轨迹展开(带工具和不带工具):
带工具轨迹展开:我们对策略可以访问外部工具的智能体展开进行采样。其轨迹包含一次或多次工具调用。设 表示是否至少有一条带工具轨迹得出了正确答案。
不带工具轨迹展开:我们对禁用工具访问的展开进行采样,迫使策略仅依赖其参数化知识。设 表示是否至少有一条不带工具轨迹得出了正确答案。
我们将策略在问题 上的知识边界定义为:
| (4) |
其中 表示 位于模型的内在知识范围内(即无需工具调用),而 表示需要外部工具。由于不带工具的展开不涉及任何工具交互或环境延迟,与带工具展开相比,其时间消耗显著降低,使得这一探测步骤在计算上非常高效。
4.2 边界引导的信号构建
基于双路径结果 ,我们将每个问题的轨迹分为四类,并构建相应的训练信号:
工具依赖(=✓, =✗)。模型仅在需要工具调用时,才能通过工具调用正确作答。我们选择正确使用工具且工具调用次数最少的轨迹作为目标,在保留必要工具调用的同时,强化高效的工具使用模式。当多条正确轨迹具有相同的最小工具调用次数时,我们随机采样其中一条以避免偏差。在更细粒度上,每次工具调用都反映了动态的步骤级知识边界决策:当模型的参数化知识不足以支撑某个具体过程推理步骤时,模型便会调用工具。因此,选择最小工具调用轨迹,能够针对特定问题,在每一步强化最广泛可达的知识边界。
高效(=✓, =✓)。模型无需工具即可正确作答,此时工具调用是多余的。我们随机选择一条正确的无工具轨迹作为目标,教导模型在知识边界范围内的问题上,绕过不必要的工具调用。
模型幻觉(=✗, =✓)。模型无需工具即可正确作答,但使用工具后反而答错,这表明工具调用引入了有害噪声,或将模型引向了错误的推理路径。我们选择一条正确的无工具轨迹作为目标,引导模型在特定问题上摆脱有害的工具依赖。
双错(=✗, =✗)。两种路径均未得出正确答案。无法构建可靠的监督信号;对于这些实例,我们仅依赖原始的强化学习目标。
4.3 联合训练目标
整体训练目标将原始强化学习损失与知识边界引导的训练目标相结合:
| (5) |
其中 可替换为任何经典的智能体强化学习损失(例如 DAPO、GSPO),而 是基于所选目标轨迹的在线策略交叉熵训练目标:
| (6) |
其中 分别表示来自工具依赖、高效和模型幻觉类别中具有构建信号的题目集合, 是第 4.2 节中所述的针对题目 所选的目标轨迹。系数 控制边界引导目标相对于强化学习损失的强度。
关键在于,由于两者都是基于当前模型的在策略(on-policy) rollout 计算得出的,因此知识边界在每个训练步骤都会被动态重新评估。随着模型通过强化学习训练不断改进,特定问题的知识边界可能会发生变化,而边界引导信号也会相应调整。这种在策略特性使得 AKBE 与那些使用静态离线数据、无法追踪此类动态演变的方法区分开来。此外,AKBE 被设计为一个即插即用模块:它可以通过在训练期间简单地添加相关项,无缝集成到任何智能体强化学习算法中,无论该算法的具体形式如何。
5 实验
| 方法 | 多跳 | 单跳 | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Hotpot | 2Wiki | MuSiQue | Bamb. | 平均 | TC | TP | NQ | TriviaQA | PopQA | 平均 | TC | TP | |
| 骨干模型:Qwen3-4B | |||||||||||||
| ReAct | 30.42 | 32.92 | 12.83 | 44.80 | 30.01 | 3.43 | 8.75 | 26.75 | 53.53 | 35.34 | 41.31 | 2.27 | 18.19 |
| Search-o1 | 35.18 | 38.45 | 14.21 | 48.00 | 34.83 | 3.31 | 10.52 | 33.85 | 57.42 | 40.27 | 46.11 | 1.72 | 26.81 |
| R1-Searcher | 45.62 | 49.83 | 19.51 | 50.40 | 45.20 | 3.28 | 13.78 | 46.85 | 64.52 | 49.43 | 54.95 | 1.48 | 37.13 |
| Search-R1 | 44.80 | 50.66 | 18.94 | 52.00 | 45.40 | 3.16 | 14.33 | 45.98 | 65.17 | 49.18 | 54.97 | 1.41 | 38.99 |
| OTC-PO | 38.74 | 47.71 | 15.16 | 48.00 | 41.27 | 2.06 | 20.03 | 45.31 | 64.35 | 49.86 | 54.90 | 1.02 | 53.83 |
| -GRPO | 45.83 | 51.49 | 21.06 | 56.00 | 46.39 | 3.01 | 15.41 | 47.00 | 64.73 | 49.95 | 55.30 | 1.53 | 36.15 |
| 离线 AKBE | 45.21 | 51.12 | 19.84 | 54.40 | 45.84 | 2.45 | 18.71 | 46.52 | 64.55 | 49.72 | 55.06 | 1.22 | 45.13 |
| AKBE(我们的方法) | 46.53 | 51.89 | 21.17 | 56.80 | 46.82 | 2.60 | 18.01 | 47.33 | 65.18 | 50.80 | 55.90 | 1.16 | 48.19 |
| 骨干模型:Qwen2.5-7B | |||||||||||||
| ReAct | 2.85 | 1.94 | 0.58 | 4.00 | 2.10 | 4.04 | 0.52 | 4.34 | 10.67 | 9.32 | 9.23 | 1.36 | 6.79 |
| Search-o1 | 18.52 | 16.73 | 5.42 | 16.00 | 16.10 | 3.92 | 4.11 | 15.87 | 32.45 | 22.18 | 25.37 | 1.28 | 19.82 |
| R1-Searcher | 47.21 | 44.85 | 20.63 | 44.00 | 43.02 | 3.91 | 11.00 | 45.12 | 63.28 | 49.05 | 54.07 | 1.08 | 50.06 |
| Search-R1 | 46.64 | 45.60 | 19.97 | 45.90 | 43.19 | 3.77 | 11.45 | 44.35 | 63.65 | 48.70 | 53.96 | 1.05 | 51.39 |
| OTC-PO | 42.16 | 43.72 | 20.48 | 44.80 | 40.72 | 1.56 | 26.10 | 48.72 | 63.46 | 47.64 | 53.89 | 1.00 | 53.89 |
| -GRPO | 48.02 | 47.92 | 21.60 | 48.00 | 45.13 | 3.34 | 13.51 | 47.89 | 64.21 | 48.18 | 54.35 | 1.03 | 52.76 |
| 离线 AKBE | 47.38 | 46.81 | 20.95 | 46.40 | 44.22 | 2.68 | 16.50 | 47.15 | 63.92 | 48.05 | 54.08 | 1.01 | 53.54 |
| AKBE(我们的方法) | 48.69 | 47.59 | 22.23 | 48.80 | 45.24 | 2.94 | 15.39 | 49.64 | 64.63 | 47.91 | 54.59 | 1.00 | 54.59 |
5.1 实验设置
数据集。我们在工具增强的搜索场景下,使用七个问答评测基准来评估 AKBE。遵循 Search-R1(Jin 等人,2025)的设置,我们部署了一个基于维基百科的轻量级搜索引擎作为外部工具环境。这些评测基准分为两类:多跳问答,包括 HotpotQA(Yang 等人,2018)、2WikiMultihopQA(Ho 等人,2020)、MuSiQue(Trivedi 等人,2022)和 Bamboogle(Press 等人,2023),这些需要多步检索和推理;以及单跳问答,包括 Natural Questions(NQ)(Kwiatkowski 等人,2019)、TriviaQA(Joshi 等人,2017)和 PopQA(Mallen 等人,2022),这些通常只需要一次检索。所有评测基准均使用精确匹配(EM)作为主要评估指标。我们还额外报告了工具调用次数(TC),定义为每个问题平均调用的工具次数,以及工具生产力(TP),用于衡量每单位工具使用所对应的准确率。
基线方法。我们将 AKBE 与以下方法进行对比:(1)ReAct(Yao 等人,2023):一种基于提示词的方法,作为没有经过强化学习训练的参考基准;(2)Search-o1(Li 等人,2025):一个将智能体搜索流程整合到推理过程中的框架;(3)R1-Searcher(Song 等人,2025)和(4)Search-R1(Jin 等人,2025):两种经典的智能体强化学习框架,它们部署 GRPO 来增强搜索能力;(5)OTC-PO(Wang 等人,2025):一种奖励塑形方法,通过引入工具生产力项来惩罚冗余的工具调用;(6)-GRPO(Wu 等人,2025b):一种奖励塑形方法,它引入了一个基于置信度的阈值来减少不确定性;以及(7)离线 AKBE:AKBE 的一种离线变体,它使用相同的知识边界引导信号构建策略,但从一个固定的、经过 GRPO 训练的检查点生成信号数据,作为直接对比来验证在线动态信号构建的必要性。请注意,基线方法和 AKBE 的更多实现细节在附录 A 中提供。
5.2 AKBE 的主要结果
我们在表 1 中展示了基于两种骨干模型和七个基准测试的主要结果。AKBE 在多跳和单跳基准测试上均获得了最高的平均 EM 分数,同时大幅降低了 TC,并在大多数情况下带来了持续的 TP 提升。在 Qwen3-4B 上,AKBE 在所有七个基准测试中的平均 EM 相比其基础方法提升了 +1.85,同时将 TC 降低了 18%,工具生产力提升约 25%。同样的效果在 Qwen2.5-7B 上也成立,证实了其在不同模型架构和规模上的通用性。相比之下,OTC-PO 在所有设置中实现了最低的 TC(在表 1 中用下划线标出),但严重牺牲了准确性,这证实了粗粒度的奖励塑造会激励不加区分地抑制工具调用,从而导致奖励作弊。-GRPO 通过其置信度阈值避免了 EM 崩溃,但提供的 TC 降低有限。AKBE 实现了严格意义上更好的平衡:在比 -GRPO 更大程度降低 TC 的同时,还提升了 EM。
将 AKBE 与其离线变体(Offline AKBE)进行比较,揭示了在线策略信号构建的重要性。Offline AKBE 在 EM 分数上始终低于 AKBE,尽管其 TC 更低,这反映了从冻结的已训练策略中生成的过于激进的“减少工具调用”信号。离线数据所捕获的知识边界反映了模型在训练后期阶段的能力,这对于早期训练中较弱的策略来说过于乐观。由此产生的静态边界信号无法与模型在整个训练过程中不断演变的知识状态保持一致,导致过早的工具抑制和精度下降。这验证了我们的核心主张:动态的在线策略知识边界追踪对于实现 EMTC 平衡至关重要。
5.3 分析
5.3.1 即插即用泛化
由于 AKBE 通过辅助监督信号增强模型的知识边界感知,而非修改 RL 奖励或优化流程,因此它天然地与基础智能体 RL 算法的选择正交,可作为即插即用模块使用。为验证这一点,我们将 AKBE 与四种智能体 RL 算法集成:GRPO (Shao et al., 2024)、DAPO (Yu et al., 2025)、GSPO (Zheng et al., 2025) 和 AEPO (Dong et al., 2025),每种算法代表不同的优化策略,例如动态采样、序列级优化和熵驱动探索。
如表 2 所示,AKBE 在所有四种基础算法上均一致地提升了平均 EM 并降低了 TC。值得注意的是,无论基础方法本身的特性如何,这种提升都是一致的:DAPO 由于其针对多样化轨迹的动态采样策略,已经实现了较低的 TC,但 AKBE 仍能进一步将其降低,同时提升 EM。对于基础 TC 较高(3.23 和 3.08)的 GSPO 和 AEPO,AKBE 带来了更大的 TC 降低,同时 EM 也持续提升。TP 指标在所有四种组合中均一致提升,增益范围从 到 。这些结果证实 AKBE 是一个高效的正交模块。边界引导的训练目标提供了互补的学习信号,提升了工具调用效率,同时不干扰基础 RL 算法的优化动态。
| 方法 | Hotpot | 2Wiki | MuSi. | Bamb. | 平均 | TC | TP |
|---|---|---|---|---|---|---|---|
| GRPO | 44.80 | 50.66 | 18.94 | 52.00 | 45.40 | 3.16 | 14.33 |
| + AKBE | 46.53 | 51.89 | 21.17 | 56.80 | 46.82 | 2.60 | 18.01 |
| DAPO | 45.95 | 51.81 | 21.68 | 51.20 | 46.65 | 2.61 | 17.85 |
| + AKBE | 45.83 | 52.35 | 22.51 | 52.80 | 47.01 | 2.38 | 19.75 |
| GSPO | 47.07 | 49.25 | 22.68 | 49.60 | 45.69 | 3.23 | 14.16 |
| + AKBE | 47.62 | 49.78 | 23.50 | 48.00 | 46.24 | 2.84 | 16.28 |
| AEPO | 46.36 | 51.78 | 23.47 | 52.00 | 46.96 | 3.08 | 15.25 |
| + AKBE | 46.89 | 52.26 | 24.24 | 54.40 | 47.50 | 2.73 | 17.40 |
| 配置 | 平均 EM | TC | TP |
|---|---|---|---|
| GRPO | 45.40 | 3.16 | 14.33 |
| AKBE(我们的方法) | 46.86 | 2.60 | 18.02 |
| 无工具依赖 | 43.56 | 2.15 | 20.26 |
| 无效率 | 46.50 | 2.92 | 15.93 |
| 无幻觉 | 46.55 | 2.58 | 18.04 |
| 仅依赖工具 | 46.02 | 2.85 | 16.15 |
5.3.2 轨迹类别消融研究
为理解每个信号类别的贡献,我们通过从知识边界引导的训练目标中逐一移除单个类别来进行消融实验。
在表 3 中,我们发现移除依赖工具信号会导致 EM 显著低于 GRPO,尽管其 TC 达到最低。剩余的效率与幻觉类别仅监督无工具轨迹,导致对必要工具调用的过度抑制,从而降低了任务准确率。这证实了依赖工具信号作为一种关键保护机制,防止了效率导向信号过度抑制必要的工具调用。移除效率信号导致 TC 上升,表明它是消除冗余工具调用的主要力量。移除幻觉信号导致 EM 小幅下降,而 TC 保持可比,验证了幻觉类别纠正了工具调用覆盖正确内部推理的有害工具调用路径,从而促进了 EM 的提升。值得注意的是,仅依赖工具信号一项就已优于 GRPO,表明即使在无工具展开大多失败的高难度问题上,AKBE 仍然有效。完整的 AKBE 优于所有子集,证实了这三个类别是互补的:依赖工具信号教导何时以及如何高效使用工具,效率信号教导何时不需要工具,幻觉信号教导何时工具是有害的。
5.3.3 系数的选择策略
我们研究了平衡强化学习损失与边界引导目标之间系数的选择策略。图4报告了在Qwen3-4B多跳任务上,当系数在某个范围内变化时,平均精确匹配率、工具调用次数和工具调用通过率的实验结果。AKBE在所有系数取值下的精确匹配率均持续优于GRPO,并在某个系数值上取得了最佳平衡。当系数超过0.2时,精确匹配率急剧下降,这表明过强的边界引导目标主导了强化学习损失,导致对工具调用的过度抑制。值得注意的是,在所有系数取值下,工具调用次数和工具调用通过率均持续优于GRPO,这表明无论信号强度如何,AKBE都能可靠地提升工具使用效率。最优系数自然地在两个目标之间平衡了梯度贡献,因为边界引导目标每个问题最多作用于一条目标轨迹,而强化学习损失则基于多条采样轨迹计算。关于这一点,我们在附录B中提供了详细的理论分析。
5.3.4 训练过程中的轨迹分布
为了考察知识边界在训练过程中如何演变,我们在图3中比较了Qwen2.5-7B多跳任务上训练早期(第1-40步)和训练后期(第201-240步)的轨迹类别分布。
最显著的变化是“两者皆错”比例大幅下降,这表明智能体强化学习训练逐步使模型能够解决先前无法解答的问题。关键在于,“效率”类别的提升幅度最大,证明 AKBE 成功促进了知识内化——模型越来越倾向于利用其参数化知识来回答问题。与此同时,“模型幻觉”显著减少,证实了幻觉信号在训练过程中有效纠正了有害的工具调用路径。这些变化验证了我们设计的两个关键方面:(1)知识边界在训练过程中并非静态,这证明了采用在线策略信号构建方法优于静态离线方法;(2)AKBE 的边界引导目标与强化学习目标协同作用——强化学习增强了模型的工具增强推理能力,而 AKBE 则提供知识边界引导的效率信号,指导模型最大化利用其知识边界,从而实现高效推理路径,并最大程度减少冗余的工具调用。
5.3.5 计算开销
一个自然的担忧是,AKBE 中额外的不使用工具的 rollout 是否会引入过高的计算开销。图 5 比较了在 Qwen3-4B 多跳任务上,GRPO 与 AKBE 每个训练步骤的时间消耗。令人惊讶的是,尽管每批次执行了额外的不使用工具的 rollout,AKBE 平均仍比 GRPO 快 15%。这一结果源于两个因素:(1)不使用工具的 rollout 完成速度远快于使用工具的 rollout,因为它们不涉及工具交互或环境延迟;(2)随着 AKBE 在训练过程中逐步减少工具调用,使用工具的 rollout 本身也因工具调用次数减少而变得更短,导致训练后期步骤时间加速。这表明 AKBE 引入的计算开销很小,并且在大多数情况下,其带来的效率提升足以弥补这部分开销。我们进一步在附录 E 中提供了工具调用次数和响应长度的详细逐步骤对比。
6 结论
在本文中,我们提出了 AKBE,一种简单而有效的方法,通过在智能体强化学习训练期间进行同策略双路径推演,动态探测模型的内在知识边界。通过构建知识边界引导的监督信号,AKBE 消除了冗余的工具调用,同时保留了必要的工具调用,并引导模型形成高效的工具调用模式。与容易遭受奖励破解问题的奖励塑形方法不同,AKBE 在不修改强化学习目标的情况下,在实例层面提供更细粒度的引导,从而同时提升任务准确率和工具调用效率。在七个问答基准测试和两个骨干模型上的实验验证了其有效性,表明显式的同策略知识边界建模是一种有前景且通用的高效智能体强化学习策略。
7 局限性
尽管由于在训练后期减少了工具调用,AKBE 的平均训练时间比 GRPO 更快,但在训练早期阶段,当工具调用尚未减少时,额外的无工具推演确实引入了额外的计算成本。未来的工作可以探索更高效的推演策略,例如自适应采样,仅对可能处于知识边界内的问题选择性执行无工具推演。此外,系数在整个训练过程中是固定的,而 和 之间的最优平衡可能随训练阶段和任务难度而变化。一种能够根据当前轨迹分布或任务复杂度,在每一步自适应调整 的方法,可能进一步提升性能。
参考文献
- D. Chen, Z. Zong, Z. Ma, L. Luo, Y. Li, C. Li, P. Chen, and J. Jiang (2026) A 2 tgpo: agentic turn-group policy optimization with adaptive turn-level clipping. arXiv preprint arXiv:2605.06200. 引用自:§2.
- G. Dong, L. Bao, Z. Wang, K. Zhao, X. Li, J. Jin, J. Yang, H. Mao, F. Zhang, K. Gai, G. Zhou, Y. Zhu, J. Wen, and Z. Dou (2025) Agentic entropy-balanced policy optimization. External Links: 2510.14545, Link 引用自:§A.2, §1, §2, §3.2, §5.3.1.
- L. Feng, Z. Xue, T. Liu 和 B. An (2025) 《组内组策略优化用于大语言模型智能体训练》。arXiv 预印本 arXiv:2505.10978。引用自:§1。
- X. Ho, A. Duong Nguyen, S. Sugawara 和 A. Aizawa (2020) 《构建用于全面评估推理步骤的多跳问答数据集》。载于《第 28 届国际计算语言学会议论文集》,D. Scott, N. Bel 和 C. Zong 主编,西班牙巴塞罗那(线上),第 6609–6625 页。外部链接:链接,文献标识码。引用自:§A.3, §5.1。
- Y. Ji, Z. Ma, Y. Wang, G. Chen, X. Chu 和 L. Wu (2025) 《用于大语言模型智能体强化学习的树搜索》。外部链接:2509.21240,链接。引用自:§2, §3.2。
- B. Jin, H. Zeng, Z. Yue, J. Yoon, S. Arik, D. Wang, H. Zamani 和 J. Han (2025) 《Search-R1:训练大语言模型进行推理并利用强化学习调用搜索引擎》。arXiv 预印本 arXiv:2503.09516。引用自:§A.1, §A.2, §A.6, §2, §3.2, §5.1, §5.1。
- M. Joshi, E. Choi, D. Weld 和 L. Zettlemoyer (2017) 《TriviaQA:一个用于阅读理解的大规模远程监督挑战数据集》。载于《第 55 届计算语言学协会年会论文集(第一卷:长论文)》,R. Barzilay 和 M. Kan 主编,加拿大温哥华,第 1601–1611 页。外部链接:链接,文献标识码。引用自:§A.3, §5.1。
- T. Kwiatkowski, J. Palomaki, O. Redfield, M. Collins, A. Parikh, C. Alberti, D. Epstein, I. Polosukhin, J. Devlin, K. Lee, K. Toutanova, L. Jones, M. Kelcey, M. Chang, A. M. Dai, J. Uszkoreit, Q. Le 和 S. Petrov (2019) 《Natural Questions:一个用于问答研究的基准》。计算语言学协会汇刊 7,第 452–466 页。外部链接:链接,文献标识码。引用自:§A.3, §5.1。
- X. Li, G. Dong, J. Jin, Y. Zhang, Y. Zhou, Y. Zhu, P. Zhang 和 Z. Dou (2025) 《Search-o1:智能体搜索增强的大型推理模型》。载于《2025 年自然语言处理经验方法会议论文集》,第 5420–5438 页。引用自:§5.1。
- Z. Luo, Z. Luo, M. Zhang 和 R. Mao (2026) 《TabTracer:用于大语言模型复杂表格推理的蒙特卡洛树搜索》。arXiv 预印本 arXiv:2602.14089。引用自:§1。
- A. Mallen、Asai、Akari、V. Zhong、R. Das、H. Hajishirzi 和 D. Khashabi(2022)《何时不应信任语言模型:探究参数化与非参数化记忆的有效性与局限性》。arXiv 预印本。引用自:§A.3、§5.1。
- O. Press、M. Zhang、S. Min、L. Schmidt、N. A. Smith 和 M. Lewis(2023)《衡量并缩小语言模型中的组合性差距》。外部链接:2210.03350,链接。引用自:§A.3、§5.1。
- C. Qian、E. C. Acikgoz、H. Wang、X. Chen、A. Sil、D. Hakkani-Tur、G. Tur 和 H. Ji(2025)《SMART:用于缓解工具过度使用的自我感知智能体》。收录于《计算语言学协会发现:ACL 2025》,第 4604–4621 页。引用自:§2。
- Qwen、:、A. Yang、B. Yang、B. Zhang、B. Hui、B. Zheng、B. Yu、C. Li、D. Liu、F. Huang、H. Wei、H. Lin、J. Yang、J. Tu、J. Zhang、J. Yang、J. Yang、J. Zhou、J. Lin、K. Dang、K. Lu、K. Bao、K. Yang、L. Yu、M. Li、M. Xue、P. Zhang、Q. Zhu、R. Men、R. Lin、T. Li、T. Tang、T. Xia、X. Ren、X. Ren、Y. Fan、Y. Su、Y. Zhang、Y. Wan、Y. Liu、Z. Cui、Z. Zhang 和 Z. Qiu(2025)《Qwen2.5 技术报告》。外部链接:2412.15115,链接。引用自:§A.7。
- T. Schick、J. Dwivedi-Yu、R. Dessi、R. Raileanu、M. Lomeli、E. Hambro、L. Zettlemoyer、N. Cancedda 和 T. Scialom(2023)《Toolformer:语言模型可以自学使用工具》。收录于《第三十七届神经信息处理系统会议》,外部链接:链接。引用自:§1。
- J. Schulman、F. Wolski、P. Dhariwal、A. Radford 和 O. Klimov(2017)《近端策略优化算法》。外部链接:1707.06347,链接。引用自:§3.2。
- Z. Shao、P. Wang、Q. Zhu、R. Xu、J. Song、X. Bi、H. Zhang、M. Zhang、Y. K. Li、Y. Wu 和 D. Guo(2024)《DeepSeekMath:突破开放语言模型数学推理的极限》。外部链接:2402.03300,链接。引用自:§1、§2、§3.2、§5.3.1。
- G. Sheng、C. Zhang、Z. Ye、X. Wu、W. Zhang、R. Zhang、Y. Peng、H. Lin 和 C. Wu(2024)《HybridFlow:一种灵活高效的 RLHF 框架》。arXiv 预印本:arXiv:2409.19256。引用自:§A.7。
- S. Si, H. Zhao, Y. Lei, Q. Wang, D. Chen, Z. Wang, Z. Wang, K. Luo, Z. Wang, G. Chen 等人 (2026) 《从上下文到技能:语言模型能否从上下文中熟练学习?》. arXiv 预印本 arXiv:2604.27660. 引用于:§1.
- H. Song, J. Jiang, Y. Min, J. Chen, Z. Chen, W. X. Zhao, L. Fang, 和 J. Wen (2025) 《R1-searcher:通过强化学习激励大语言模型的搜索能力》. arXiv 预印本 arXiv:2503.05592. 引用于:§5.1.
- H. Trivedi, N. Balasubramanian, T. Khot, 和 A. Sabharwal (2022) 《MuSiQue:通过单跳问题组合实现多跳问题》. 《计算语言学协会汇刊》第10卷,第539–554页. 外部链接:链接, 文献 引用于:§A.3, §5.1.
- H. Wang, C. Qian, W. Zhong, X. Chen, J. Qiu, S. Huang, B. Jin, M. Wang, K. Wong, 和 H. Ji (2025) 《少行动即多推理!教会模型高效行动》. arXiv 预印本 arXiv:2504.14870. 引用于:§1, §1, §2, §5.1.
- L. Wang, N. Yang, X. Huang, B. Jiao, L. Yang, D. Jiang, R. Majumder, 和 F. Wei (2022) 《通过弱监督对比预训练的文本嵌入向量》. arXiv 预印本 arXiv:2212.03533. 引用于:§A.6.
- P. Wu, M. Zhang, K. Wan, W. Zhao, K. He, X. Du, 和 Z. Chen (2025a) 《Hiprag:用于高效智能体检索增强生成的分层过程奖励》. arXiv 预印本 arXiv:2510.07794. 引用于:§2.
- P. Wu, M. Zhang, X. Zhang, X. Du, 和 Z. Chen (2025b) 《明智搜索:通过降低不确定性来缓解次优的智能体搜索》. 载于《2025年自然语言处理经验方法会议论文集》,第19734–19745页. 引用于:§1, §2, §5.1.
- R. Xie, D. Gopinath, D. Qiu, D. Lin, H. Sun, S. Potdar, 和 B. Dhingra (2026) 《搜索增强型大语言模型中的过度搜索》. arXiv 预印本 arXiv:2601.05503. 引用于:§1, §2.
- A. Yang, A. Li, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Gao, C. Huang, C. Lv, C. Zheng, D. Liu, F. Zhou, F. Huang, F. Hu, H. Ge, H. Wei, H. Lin, J. Tang, J. Yang, J. Tu, J. Zhang, J. Yang, J. Yang, J. Zhou, J. Zhou, J. Lin, K. Dang, K. Bao, K. Yang, L. Yu, L. Deng, M. Li, M. Xue, M. Li, P. Zhang, P. Wang, Q. Zhu, R. Men, R. Gao, S. Liu, S. Luo, T. Li, T. Tang, W. Yin, X. Ren, X. Wang, X. Zhang, X. Ren, Y. Fan, Y. Su, Y. Zhang, Y. Zhang, Y. Wan, Y. Liu, Z. Wang, Z. Cui, Z. Zhang, Z. Zhou, 和 Z. Qiu (2025) Qwen3 技术报告。外部链接: 2505.09388, 链接 引用自: §A.7。
- Z. Yang, P. Qi, S. Zhang, Y. Bengio, W. Cohen, R. Salakhutdinov, 和 C. D. Manning (2018) HotpotQA:一个用于多样化、可解释的多跳问答的数据集。载于《2018年自然语言处理经验方法会议论文集》,E. Riloff, D. Chiang, J. Hockenmaier, 和 J. Tsujii 编,比利时布鲁塞尔,第2369–2380页。外部链接: 链接, 文献标识码 引用自: §A.3, §5.1。
- S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. R. Narasimhan, 和 Y. Cao (2023) ReAct:在语言模型中协同推理与行动。载于《第十一届国际学习表征会议》,外部链接: 链接 引用自: §1, §3.1, §5.1。
- Q. Yu, Z. Zhang, R. Zhu, Y. Yuan, X. Zuo, YuYue, W. Dai, T. Fan, G. Liu, J. Liu, L. Liu, X. Liu, H. Lin, Z. Lin, B. Ma, G. Sheng, Y. Tong, C. Zhang, M. Zhang, R. Zhang, W. Zhang, H. Zhu, J. Zhu, J. Chen, J. Chen, C. Wang, H. Yu, Y. Song, X. Wei, H. Zhou, J. Liu, W. Ma, Y. Zhang, L. Yan, Y. Wu, 和 M. Wang (2025) DAPO:一个大规模开源大语言模型强化学习系统。载于《第三十九届神经信息处理系统年度会议》,外部链接: 链接 引用自: §A.4, §1, §2, §5.3.1。
- C. Zheng, S. Liu, M. Li, X. Chen, B. Yu, C. Gao, K. Dang, Y. Liu, R. Men, A. Yang, J. Zhou, 和 J. Lin (2025) 组序列策略优化。外部链接: 2507.18071, 链接 引用自: §A.4, §2, §5.3.1。
- Z. Zong, D. Chen, Y. Li, Q. Yi, B. Zhou, C. Li, B. Qian, P. Chen, 和 J. Jiang (2026) AT²po:基于树搜索的智能体回合制策略优化。arXiv 预印本 arXiv:2601.04767。引用自: §A.2, §1, §2。
附录 A 实现细节
A.1 奖励设计
我们的训练流程采用一种二元结果奖励,该奖励将答案正确性与结构格式要求相结合。正确性信号遵循 Search-R1(Jin 等人,2025)的奖励公式,使用精确匹配作为主要评估标准。
精确匹配奖励。
给定从智能体轨迹中提取的最终答案和真实答案,EM 奖励定义如下:
| (7) |
这种严格的二元公式消除了部分评分带来的模糊性,推动策略向完全正确的答案收敛,为智能体强化学习提供了清晰的优化信号。
格式约束。
除了正确性之外,每条轨迹还必须满足结构有效性要求。响应必须包含由 `<think>...</think>` 标签包裹的推理过程,以及由 `<answer>...</answer>` 标签包裹的最终答案,且答案需进一步用 `\boxed{}` 括起来。格式指示器为:
| (8) |
违反此格式的响应无论答案正确与否均不得分,从而确保可靠的工具调用解析和最终答案提取。
最终奖励。
整体奖励结合了上述两个部分:
| (9) |
只有当轨迹满足格式要求并提供完全正确的答案时,才能获得最高奖励 1;格式违规将受到明确惩罚。
A.2 提示词模板
AKBE 需要两个提示词模板用于其双路径展开,如图 6 和图 7 所示。
带工具提示词。
带工具模板(图 6)遵循先前智能体强化学习工作中采用的基于标签的格式(Jin 等人,2025;Dong 等人,2025;Zong 等人,2026)。每次展开被结构化为由专用标签对分隔的语义不同区域:推理步骤在 `<think></think>` 内表述,检索查询通过 `<search></search>` 发出,环境观察结果注入到 `<result></result>` 中,最终预测在 `<answer></answer>` 内输出,且标准答案用 `\boxed{}` 括起来以便进行精确匹配提取。
无工具提示词。
无工具模板(图 7)移除了所有与工具相关的指令和标签(`<search>` 和 `<result>`),仅保留推理(`<think></think>`)和回答(`<answer></answer>`)部分。这迫使模型仅依靠其参数化知识生成答案,从而使 AKBE 能够通过比较两条路径的正确性来探测知识边界。
A.3 数据集
我们在两类广泛使用的问答基准上进行了实验,以评估我们提出的 AKBE 的有效性。
多跳问答。
该类别评估多轮工具使用和组合推理能力,正确答案无法从单次检索的段落中获得。HotpotQA(Yang 等人,2018)是一个基于维基百科的大规模基准,包含支持事实标注,是广泛使用的多跳问答测试平台。2WikiMultiHopQA(Ho 等人,2020)将维基百科段落与 Wikidata 三元组结合,生成需要显式多跳实体推理的问题。MuSiQue(Trivedi 等人,2022)包含约 2.5 万个问题,涵盖 2 到 4 跳推理,通过单跳原语的受控组合合成,以探测细粒度的推理深度。Bamboogle(Press 等人,2023)提供了一组规模虽小但具有对抗性的组合查询,作为智能体 RL 策略的鲁棒性探针。
单跳问答。
该类别验证单步检索任务的性能。Natural Questions(NQ)(Kwiatkowski 等人,2019)汇总了从维基百科回答的真实用户查询,是检索增强生成的标准基准。TriviaQA(Joshi 等人,2017)的问题与支持证据之间存在显著的词汇和句法差异,用于测试对表面形式变化的鲁棒性。PopQA(Mallen 等人,2022)是一个以实体为中心的基准,旨在区分外部检索与参数化记忆的贡献,因此它自然适用于诊断策略是否真正利用了搜索工具,还是依赖于记忆的事实。
A.4 AKBE 设置
关于我们AKBE的实现细节,我们使用的训练批次大小为 ,小批次大小为 ,最大响应长度为 。在策略执行阶段,对于带工具的策略执行,我们使用的策略执行大小为 ,对于无工具的策略执行,策略执行大小为 ,最大工具使用次数设置为 。AKBE目标的裁剪阈值设置为 (与GRPO相同)。遵循先前的研究工作(Yu等人,2025;Zheng等人,2025),我们移除了KL正则化项(),以允许策略探索多样化的策略执行策略。在多跳和单跳设置下,AKBE系数均设置为 ,这从经验上平衡了 和 之间的梯度贡献(详见§B分析)。
A.5 基线设置
| 配置 | 数值 |
| 优化器 | AdamW |
| 学习率 | 1e-6 |
| 裁剪比率 | 0.2 |
| 训练批次大小 | 64 |
| PPO小批次大小 | 8 |
| 策略执行次数 | 16 |
| 最大提示词长度 | 2000 |
| 最大响应长度 | 6192 |
| 最大工具调用轮次 | 6 |
| 奖励指标 | EM |
| 检索器 | 本地维基百科 |
| 前K个检索段落 | 3 |
表4总结了所有基于强化学习的基线所使用的共享超参数。对于特定方法的配置,我们遵循各自原始论文中报告的设置。所有基线均在没有任何额外SFT阶段的情况下进行训练。我们选择并报告在所有评估基准上达到最高平均EM值的检查点结果。
A.6 搜索工具环境
我们的搜索工具环境遵循Search-R1(Jin等人,2025)的设置。我们使用维基百科快照作为检索语料库,并使用e5-base-v2(Wang等人,2022)作为密集检索器。该知识库包含约2100万个维基百科条目,为单跳和多跳查询提供了广泛的事实覆盖。在策略发出检索动作的每一轮中,搜索引擎会根据查询对候选段落进行评分,并返回最相关的3个条目,这些条目作为工具观察结果注入到上下文中,供模型进行推理。
A.7 硬件与工件
所有训练和评估实验均在配备 8 块 NVIDIA H20 GPU 的单节点上进行。我们采用两个公开可用的检查点作为骨干策略:Qwen3-4B(Yang 等人,2025)和 Qwen2.5-7B(Qwen 等人,2025),选择它们是因为其强大的推理能力以及与智能体后训练的良好兼容性。我们的训练基础设施基于 VeRL 框架(Sheng 等人,2024),这是一个混合控制器强化学习系统,其模块化的 rollout 接口支持 AKBE 双路径设计所需的多轮、工具交互式 rollout 调度。
附录 B 系数的理论分析
我们从 与 之间的梯度贡献平衡角度,对 的最优值进行了理论分析。
设置。
考虑一个包含 个问题的训练批次。对于每个问题 ,GRPO 目标函数是在 条带工具 rollout 轨迹上计算的,而 AKBE 目标函数最多选择一条目标轨迹 。总训练损失为:
| (10) |
梯度分析。
来自 GRPO 损失的每个问题梯度贡献涉及 条轨迹:
| (11) |
其中 是组内相对优势。假设每条轨迹贡献近似相等的梯度幅度 ,则每个问题的预期梯度范数为:
| (12) |
其中 是组内优势估计值的标准差。
对于 AKBE 损失,每个问题最多贡献一条目标轨迹:
| (13) |
其梯度范数近似为 。然而,并非所有问题都能产生信号。设 为可构造信号的问题比例()。AKBE 对每个问题的有效梯度贡献(在批次上取平均)为:
| (14) |
平衡条件。
为了使两个目标函数对整体参数更新的贡献相当,我们需要:
| (15) |
代入每个问题的估计值,并注意到 GRPO 损失对轨迹进行了归一化,而 AKBE 则作用于单条轨迹:
| (16) |
实际估计。
在我们的设定中,采用二元奖励()并结合组相对归一化时,优势标准差在构造上是固定的。信号比例通常较高(约 70–80% 的问题在两条路径中至少有一条正确轨迹)。然而,关键缩放因子是轨迹数量的比值:对每个问题从条轨迹中聚合梯度(每条轨迹按权重加权),而则只使用一条完整权重的轨迹。为防止单条 AKBE 轨迹主导 RL 轨迹,应按下式缩放:
| (17) |
在我们的实验中,取时,得到,与我们的经验最优值高度吻合。考虑到任务难度和信号比例变化等动态因素,我们发现,在我们的实验设定中,构成一个合理范围。
附录 C 交叉熵与 DPO 在信号整合中的对比
一个自然的问题是,边界引导信号能否通过偏好优化(例如 DPO)而非交叉熵来整合。由于 AKBE 的信号构造为每个问题识别出偏好轨迹,因此还可以额外选择被拒绝轨迹,并应用 DPO 风格的目标函数:
| (18) |
其中和分别表示偏好轨迹和被拒绝轨迹。
信号构造差异。
基于交叉熵的 AKBE 仅保留正信号,引导模型朝向选定的目标轨迹,而不显式惩罚其他备选轨迹。相比之下,基于 DPO 的变体还会从错误或低效的 rollout 中额外选择被拒绝轨迹,具体选择工具调用次数最多(即与偏好模式偏差最大)的轨迹作为负信号。这就在高效与低效的工具使用行为之间创建了显式的偏好对。
实验对比。
我们在 Qwen3-4B 多跳任务上比较了三种配置:标准 GRPO、带交叉熵的 AKBE(我们的方法)以及带 DPO 整合的 AKBE。图 8 展示了 EM 评估结果和训练奖励动态。两个实验均采用。
结果与分析。
如图 8 所示,基于 DPO 的变体最初表现良好,甚至在训练中期阶段(第 80–140 步)超越了 GRPO。然而,随后它突然崩溃,在第 150 步之后,EM 和训练奖励均急剧下降。相比之下,AKBE(基于交叉熵)在整个训练过程中保持了稳定且单调的提升。
我们将这种不稳定性归因于偏好轨迹与拒绝轨迹之间固有的相似性。正面和负面信号都是涉及推理和工具调用的轨迹;它们仅在工具调用策略或最终正确性上有所不同。由于 DPO 明确降低了拒绝轨迹的概率,模型逐渐学会了惩罚偏好样本和拒绝样本中共同存在的工具调用模式,而不是学习高效与低效工具使用之间的细微差别。这导致工具调用行为被逐步过度抑制,最终引发训练崩溃。交叉熵通过仅提供正向监督来避免这种失败模式:它教会模型高效模式是什么样的,而不明确惩罚其他替代方案,从而为知识边界增强提供了更稳定、更具针对性的优化信号。
附录 D 知识边界估计的可靠性
我们讨论一个潜在问题:基于无工具 rollout 中“至少一个正确”(在我们的实验设置中)的知识边界估计是否可靠,或者它是否被模型仅凭运气猜对的问题所主导。
具体而言,我们分析了训练过程中所有被归类为 NT=1 的问题的正确展开(rollout)次数分布。如图 9 所示,NT=1 的问题在训练早期平均有 5.0/8 个正确展开,在训练后期平均有 5.2/8 个正确展开。仅有 18.1%(早期)和 16.7%(后期)的 NT=1 问题恰好有 1/8 正确,而大多数问题(早期 63.4%,后期 67.1%)实现了 4/8 的正确展开。值得注意的是,有 30.9%(早期)至 36.7%(后期)的问题达到了完美的 8/8,这表明模型能够充分且可靠地利用其参数化知识。
这些结果表明,知识边界估计具有高置信度,而非由噪声驱动。此外,从训练早期到后期的改进证实,AKBE 的在线策略(on-policy)设计随着训练进行,逐步增强了边界估计的可靠性。对于少数可能存在噪声的情况(1/8 正确),较小的系数确保这些弱信号无法覆盖主导的强化学习目标,而在线策略机制则提供了额外的保障。
附录 E 补充开销分析
为了补充正文(§5.3.5)中的每步开销对比,我们对驱动 AKBE 计算效率的两个底层因素进行了补充分析:工具调用频率和响应长度。
工具调用次数。
图 10(左)展示了每个训练批次中搜索调用总数随训练步数的变化。两种方法在初始阶段处于相似水平(每批次 320 次调用),但随着训练推进,其变化轨迹出现分化。GRPO 的工具调用次数稳步增长,从约 320 次上升至第 300 步时的 370 次以上,这反映了强化学习训练后的智能体在仅以正确性作为奖励时,倾向于增加工具使用的普遍现象。相比之下,AKBE 在整个训练过程中将工具调用次数稳定维持在 290–300 次左右,在训练中期(第 50–150 步)甚至略有下降。平均而言,AKBE 在整个训练过程中相比 GRPO 减少了 8.4% 的工具调用。
响应长度。
图 10(右)揭示了平均响应长度的显著差异。GRPO 呈现强劲上升趋势,平均响应长度从 2,300 个模型 token 增长至第 300 步时的 2,500 个模型 token 以上。这一增长与更多工具调用产生更长的多轮交互序列相关。相比之下,AKBE 在训练早中期阶段(第 1–100 步)呈现下降趋势,此后稳定在 1,600–1,700 个模型 token 左右。整体响应长度减少了 23.0%,这直接转化为更低的推理延迟和计算成本。
这些趋势证实了 AKBE 基于知识边界的引导信号产生了复合效率效应:通过在训练早期教会模型避免不必要的工具调用,后续的生成过程自然变得更短,从而同时加速了训练和推理。
附录 F 案例研究
我们展示了训练过程中(Qwen3-4B 多跳任务)来自 AKBE 三种信号类别的代表性案例。这些案例说明了双路径对比如何揭示模型的知识边界并指导信号构建。对于每个案例,我们展示了来自两条路径的完整推理轨迹、信号类别分类,以及哪条轨迹被选为监督目标。
案例 #1 分析(效率)。
此案例展示了效率信号类别,即模型拥有足够的参数化知识,无需外部检索即可正确作答。在带工具路径中,模型明确表示“我对《Donker Mag》并不熟悉”并启动了搜索,尽管无工具路径表明它能够从其内部知识中轻松回忆起该专辑与Die Antwoord的关联。这种差异暴露了第一级知识边界违规:模型在非必要时仍默认使用工具。通过将无工具轨迹选为监督目标,AKBE教会模型信任其对已知事实的参数化记忆,从而减少冗余工具调用并提升推理效率。
案例#2分析(工具依赖型)。
此案例展示了工具依赖类别,即外部检索确实必要。无工具路径暴露了明显的知识缺口:模型产生了看似合理但错误的答案(“沃伦·巴菲特”),在多个猜测间循环却未能得出正确答案(塞思·卡拉曼)。相比之下,带工具路径制定了针对性查询(“Baupost Group创始人”),立即检索到相关段落,从而实现了正确推理。此处,工具调用次数最少(TC=1)的带工具轨迹被选为目标。这一信号强化了第二级边界意识:对于超出参数化知识范围的问题,模型应学习高效的检索模式,通过单次工具调用解决查询。
案例#3分析(模型幻觉)。
这个案例揭示了一种微妙的失败模式:检索引发的模型幻觉。模型搜索“Arline Burks Gant 死亡日期”,但检索器返回了关于“Barbara Stoddard Burks”的不相关段落。使用工具路径没有识别出检索失败,反而基于相同的姓氏错误地将两人混为一谈,通过有缺陷的推理(“也许 Arline Burks Gant 和 Barbara Stoddard Burks 是同一个人?”)得出了错误答案。与此同时,不使用工具路径正确地回忆起 Maurice Pialat 死于 2003 年,而 Arline Burks Gant 一直活到 2011 年。这个案例表明,当检索结果嘈杂或偏离目标时,使用工具可能产生负面影响。通过选择不使用工具的轨迹作为目标,AKBE 引导模型避免在参数知识比嘈杂的搜索结果更可靠的查询上过度依赖检索。
这三个案例共同说明了 AKBE 的双路径比较机制如何在知识边界自适应地识别适当行为。信号构建并非统一偏向某一路径;相反,它会针对每个特定查询选择能反映最恰当工具使用决策的轨迹。这种细粒度、逐查询的监督使模型能够在不同场景下发展出细微的知识边界意识。
| 问题:Donker Mag 是哪个非洲嘻哈乐团的专辑?正确答案:Die Antwoord |
|---|
| 使用工具路径(TC=1, EM=1.0) |
| <think> 好的,用户询问的是专辑“Donker Mag”,想知道它属于哪个非洲嘻哈乐团。首先,我需要弄清楚这张专辑与哪些艺术家相关。我对“Donker Mag”并不熟悉,所以应该查一下。 </think> |
| <search> Donker Mag 专辑 </search> |
| <result> 助手回复 Donker Mag 是南非嘻哈乐团 Die Antwoord 的第三张录音室专辑,于 2014 年 6 月 3 日由 Zef Records 发行。 </result> |
| 第 1 页:“Die Antwoord” 然而,这张专辑遭到了许多评论家的平庸乃至负面评价。继《Ten$Ion》之后,他们发行了一首非专辑单曲《XP N$IV $H1T》。在《XP N$IV $H1T》发布后,Die Antwoord 于 6 月至 7 月在欧洲进行了一次短暂的巡演。巡演开始后不久,一首名为《Cookie Thumper!》的新单曲片段被发布,该单曲还将配有音乐视频。在新单曲发布前后,Die Antwoord 还公布了他们的第三张专辑《Donker Mag》的名称,该专辑于 2014 年 6 月 3 日发行。《Cookie Thumper!》的音乐视频于 2014 年 5 月 27 日作为该专辑的首支单曲发布。 |
| 第 2 页:《Magnet(杂志)》 《Magnet》是一本音乐杂志,主要关注另类、独立或非主流的乐队。该杂志每年出版四期,由 Eric T. Miller 独立拥有并编辑。20 世纪 90 年代,具有类似关注点的音乐杂志包括《Option》、《Raygun》和《Alternative Press》。《Magnet》的第一期于 1993 年中期出版。 |
| 第 3 页:“Green Monkey Records” (CDR GM1028) The OF - Escape Goat (CD GM1029) Jim McIver - Sunlight Reaches (CD GM1030) Tom Dyer’s New Pagan Gods - History Of Northwest Rock Vol. 1, 1959-1968 (CD GM1031) Fur For Fairies (CD GM1032) The Fuzz - Best Kept Secret… |
| </result> |
| <思考> 嗯,第一个搜索结果提到了“Die Antwoord”与专辑《Donker Mag》相关。Die Antwoord 是一个南非嘻哈组合。其他结果似乎不相关。所以,答案应该是 Die Antwoord。 </思考> |
| <答案> 最终答案是 \boxed{Die Antwoord}。 </答案> |
| 无工具路径(TC=0, EM=1.0) 已选为目标 |
| 《Donker Mag》这张专辑属于哪个非洲嘻哈团体?答案是:Die Antwoord。 |
| 最终答案是 \boxed{Die Antwoord}。 |
| 问题:Baupost Group 曾邀请哪位美国投资者、对冲基金经理兼慈善家来管理该对冲基金?真实答案:Seth Klarman |
|---|
| 带工具路径(TC=1,EM=1.0)被选为目标(最小 TC) |
| 好的,让我想想。问题是关于 Baupost Group 曾邀请哪位美国投资者、对冲基金经理兼慈善家来管理该对冲基金。我需要找出答案。首先,我应该回想一下 Baupost Group 是什么。Baupost 是一家对冲基金公司,对吧?他们可能有一位知名的创始人或与之相关的著名人物。问题提到他们邀请了一个特定的人来管理该基金。所以,我需要找到那个人的名字。也许我可以搜索“Baupost Group 创始人”来找到创始人的名字。让我查一下。 |
| <search> Baupost Group 创始人</search> |
| <result> |
| 第 1 页:“Baupost Group”……截至 2013 年 12 月 31 日,管理资产规模达 299 亿美元。Baupost Group 是一家只做多头的对冲基金,由哈佛大学教授 William Poorvu 与合伙人 Howard Stevenson、Jordan Baruch 及 Isaac Auerbach 于 1982 年共同创立。Poorvu 邀请 Seth Klarman 协助管理该基金,至今 Klarman 仍担任其负责人。Baupost Group 的投资理念强调风险管理。这家全球规模最大的对冲基金之一,是一家价值投资管理机构。据彭博社报道,Baupost 自成立以来的净收益排名第四。 |
| 第 2 页:“Baupost Group”Baupost Group 是一家只做多头的对冲基金,由哈佛大学教授 William Poorvu 于 1982 年创立……Poorvu 邀请 Seth Klarman 协助管理该基金,至今 Klarman 仍担任其负责人。 |
| 第 3 页:“Baupost Group”……Baupost 首席执行官 Seth Klarman 在向麻省理工学院学生发表演讲时解释说,由情绪驱动的投资研究风险很高,可能导致糟糕的投资决策。自成立以来,该公司的三个私人合伙基金实现了平均年化 19% 的回报率。 |
| </result> |
| <think> 因此,根据搜索结果,Baupost Group 由 William Poorvu 等人创立。Poorvu 邀请 Seth Klarman 协助管理该基金,至今 Klarman 仍担任其负责人。问题问的是 Baupost Group 邀请了谁来管理该对冲基金。答案是 Seth Klarman。我来确认一下:搜索结果一致提到 Seth Klarman 是现任负责人,并且他是受 Poorvu 之邀来协助管理该基金的。所以答案应该是 Seth Klarman。 </think> |
| <answer> 最终答案是 \boxed{Seth Klarman}。 </answer> |
| 无工具路径(TC=0, EM=0.0) |
| <think> 好的,让我想想。问题是关于 Baupost Group 询问哪位美国投资者、对冲基金经理和慈善家来管理该对冲基金。嗯。我记得 Baupost Group 是一家对冲基金,对吧?所以他们可能任命了一位董事总经理或其他什么人。等等,我想答案可能是沃伦·巴菲特这样的人。不,等等,巴菲特是伯克希尔·哈撒韦的 CEO。也许是另一个人。让我想想。有个人叫比尔·阿克曼?不,他更像是一位激进投资者。或者是罗杰·L·伊韦尔特?不,那不对。等等,我想正确答案是迪克·富尔德。不,他是雷曼兄弟的 CEO。也不是。也许是别人。Baupost Group 由罗伯特·F·德雷珀和杰里·罗兹创立。但问题是他们请谁来管理该基金。也许答案终究是沃伦·巴菲特。 </think> |
| <answer> 最终答案是 \boxed{沃伦·巴菲特}。 </answer> |
| 问题:莫里斯·皮亚拉和阿琳·伯克斯·甘特,谁先去世?真实答案:莫里斯·皮亚拉 |
|---|
| 使用工具路径(TC=4, EM=0.0) |
| <think> 好的,问题是问莫里斯·皮亚拉和阿琳·伯克斯·甘特谁先去世。我需要找到他们的去世日期并进行比较。首先,我应该查一下莫里斯·皮亚拉是什么时候去世的。让我搜索一下。 </think> |
| <search> 莫里斯·皮亚拉 去世日期 </search> |
| <result> |
| 第 1 页:“莫里斯·皮亚拉” 莫里斯·皮亚拉(;1925 年 8 月 31 日 – 2003 年 1 月 11 日)是一位法国电影导演、编剧和演员,以其电影严谨而不煽情的风格而闻名。他的作品常被描述为“现实主义”。皮亚拉出生于法国多姆山省的坎拉。 |
| 第 2 页:“莫里斯·皮亚拉” ……他总是愿意根据经验来调整他的叙事…… 莫里斯·皮亚拉(;1925 年 8 月 31 日 – 2003 年 1 月 11 日)是一位法国电影导演…… |
| 第 3 页:“莫里斯·罗内” ……他因癌症在巴黎一家医院去世…… 莫里斯·罗内(1927 年 4 月 13 日 – 1983 年 3 月 14 日)是一位法国电影演员、导演和作家。 |
| </result> |
| <think> 莫里斯·皮亚拉于2003年1月11日去世。现在,我需要查找阿琳·伯克斯·甘特的去世时间。让我查一下。 </think> |
| <search> 阿琳·伯克斯·甘特去世日期 </search> |
| <result> |
| 第1页:“芭芭拉·斯托达德·伯克斯”……多年后。伯克斯于1927年与斯坦福大学国家化学研究奖学金获得者赫尔曼·拉姆斯珀格结婚;两人一直维持婚姻关系直至1932年他去世。1943年,她与长期好友罗伯特·库克订婚,但她在两人结婚前便去世了。伯克斯于1943年5月25日去世,当时她“要么是失足,要么是跳下纽约市乔治·华盛顿大桥”身亡。 |
| 第2页:“理查德·芬纳·伯吉斯”……他与埃塞尔·佩特里·谢尔顿结婚……他于1945年1月13日在埃尔帕索去世。 |
| 第3页:“阿比·G·罗杰斯”……于1894年5月21日在纽约市意外去世,享年53岁。她当时正在接受肿瘤切除手术。 |
| </result> |
| <think> 嗯,第一个结果提到了芭芭拉·斯托达德·伯克斯,但那是另一个人,不是阿琳·伯克斯·甘特。第二个结果是关于理查德·芬纳·伯吉斯的。第三个是关于阿比·G·罗杰斯的。这些都不是阿琳·伯克斯·甘特。等等,也许阿琳·伯克斯·甘特就是芭芭拉·斯托达德·伯克斯?两者都出现了“伯克斯”这个姓氏。或者,可能是搞混了。假设阿琳·伯克斯·甘特与1943年去世的那位“伯克斯”有关,那么她是在莫里斯·皮亚拉(2003年)之前去世的。因此,阿琳·伯克斯·甘特先去世。 </think> |
| <answer> 最终答案是 \boxed{阿琳·伯克斯·甘特}。 </answer> |
| 无工具路径(TC=0,EM=1.0)被选为目标 |
| <think> 问题是问莫里斯·皮亚拉和阿琳·伯克斯·甘特谁先去世。我记得莫里斯·皮亚拉是著名的法国电影导演,于2003年1月11日去世。至于阿琳·伯克斯·甘特,她不太出名,但我相信她活到了2011年。因此,莫里斯·皮亚拉在2003年先于阿琳·伯克斯·甘特去世。 </think> |
| <answer> 最终答案是 \boxed{莫里斯·皮亚拉}。 </answer> |
Dingwei Chen
, Zefang Zong
, Zhipeng Ma
, Leo Luo
, Yang Li
Chengming Li
Peng Chen
Jie Jiang
Tencent Inc
Shenzhen MSU-BIT University
cuso4cdw@gmail.com
licm@smbu.edu.cn
{willzong,thomasyngli}@tencent.com
Work was done during the internship at Tencent Inc. Corresponding author.
Abstract
Agentic reinforcement learning (RL) has proven effective for training LLM-based agents with external tool-use capabilities. However, we identify that agentic RL training induces increasing redundant tool calls and blurs the model’s intrinsic knowledge boundary, where the model fails to distinguish when tools are needed versus when parametric knowledge suffices. Existing solutions based on reward shaping create coarse-grained optimization targets that tend to incentivize indiscriminate tool-call suppression, leading to reward hacking. In this paper, we propose AKBE (Agentic Knowledge Boundary Enhancement), an on-policy method that dynamically probes the model’s intrinsic knowledge boundary through dual-path (with-tool and no-tool) rollouts during training. We define the knowledge boundary as the per-instance determination of whether tools are required and the minimum tool calls necessary. By comparing correctness across paths, AKBE categorizes trajectories and constructs targeted supervisory signals that guide efficient tool-use patterns for each question. These signals are integrated seamlessly into the agentic RL training loop. Experiments on seven QA benchmarks demonstrate that AKBE improves task accuracy by on average and reduces tool calls by 18% over standard agentic RL, yielding 25% higher tool productivity without any accuracy-efficiency trade-off. Further analysis suggests its plug-and-play compatibility across different RL algorithms and the mechanism of each signal category. Our code is available at https://github.com/CuSO4-Chen/AKBE.
Efficient Agentic Reinforcement Learning with On-Policy Intrinsic Knowledge Boundary Enhancement
Dingwei Chen♠◇††thanks: Work was done during the internship at Tencent Inc., Zefang Zong♠, Zhipeng Ma♠, Leo Luo♠, Yang Li♠ Chengming Li, Peng Chen♠, Jie Jiang♠22footnotemark: 2 ♠Tencent Inc ◇The Chinese University of Hong Kong ♡Shenzhen MSU-BIT University cuso4cdw@gmail.com, licm@smbu.edu.cn {willzong,thomasyngli}@tencent.com
1 Introduction
Large language model (LLM) agents have demonstrated remarkable capabilities in solving complex tasks by integrating internal reasoning with external tool interactions (Yao et al., 2023; Schick et al., 2023; Si et al., 2026; Luo et al., 2026). Using tools such as search engines and code interpreters, these agents extend their reasoning beyond parametric knowledge. Recently, reinforcement learning has emerged as a powerful post-training paradigm for further enhancing agentic capabilities, with methods such as GRPO (Shao et al., 2024), DAPO (Yu et al., 2025), and specialized agentic RL algorithms (Feng et al., 2025; Dong et al., 2025; Zong et al., 2026) achieving promising improvements on tool-augmented reasoning benchmarks.
However, a critical yet underexplored side effect of agentic RL training is that: as the model is optimized to enhance reasoning capability with tool access, it increasingly produces redundant tool calls, either invoking tools when parametric knowledge suffices or making excessive calls when fewer would suffice, which is defined as cognitive offloading (Wang et al., 2025; Xie et al., 2026). This manifests itself as a steady growth in tool calls during training, as illustrated in Figure 1. Such an over-reliance on tool calls is problematic in two ways: (1) it wastes computational resources and increases inference latency; and (2) unnecessary tool calls may introduce noise that overrides correct internal reasoning with misleading retrieved information, leading to degradation of answer quality.
Existing approaches to efficient agentic RL address this issue primarily through reward shaping, incorporating tool-call patterns into the reward function (Wang et al., 2025; Wu et al., 2025b). However, directly coupling tool-call behavior with reward signals creates a coarse-grained optimization target. This incentivizes the model to reduce overall tool usage to gain extra reward regardless of whether specific calls are necessary, leading to reward hacking and degraded task accuracy. More fundamentally, such reward-level approaches cannot capture the per-instance distinction between necessary and redundant tool calls, nor adapt to the dynamic evolution of the model’s knowledge boundary throughout training.
In this paper, we propose AKBE (Agentic Knowledge Boundary Enhancement), an on-policy method that addresses this limitation by explicitly probing the model’s intrinsic knowledge boundary during training. We define the knowledge boundary as the per-instance determination of whether external tools are required and, when required, the minimum tool invocations necessary to reach the correct answer, representing the most efficient tool-call pattern for each question. The key insight is that for each question in a training batch, we perform dual-path rollouts with and without external tools. By comparing the correctness of these two paths, we identify whether a question lies within the model’s parametric knowledge or genuinely requires external tool calls, and further determine the minimum tool usage required in the latter case. Based on this identification, AKBE categorizes each question and constructs targeted supervisory signals: Tool-dependent selects minimum tool-call correct trajectories to reinforce efficient tool use, Efficiency selects no-tool correct trajectories to eliminate redundant calls, Hallucination selects no-tool correct trajectories to alleviate harmful tool reliance, and Both-wrong provides no signal, relying solely on the RL objective. These knowledge boundary-guided signals are integrated seamlessly into the training loop with the standard RL objective as an auxiliary on-policy training loss, providing fine-grained instance-level guidance without modifying the RL reward or optimization process. Our contributions are summarized as follows:
-
We propose AKBE, an on-policy knowledge boundary enhancement method for efficient agentic RL that dynamically probes the model’s intrinsic knowledge boundary through dual-path rollouts and constructs boundary-guided supervisory signals to eliminate redundant tool calls and reinforce efficient tool-use patterns.
-
We conduct extensive experiments on seven QA benchmarks across two backbone models, demonstrating that AKBE improves task accuracy by on average and reduces tool calls by 18% over standard agentic RL, yielding 25% higher tool productivity. It outperforms baseline methods in most cases without any accuracy-efficiency trade-off.
-
We further demonstrate that AKBE serves as a plug-and-play module compatible across diverse agentic RL algorithms, and reveal that the model’s knowledge boundary evolves dynamically during training, where each signal category naturally adapts to address a distinct failure mode of tool-use behavior.
2 Related Work
Recent work applies reinforcement learning to train LLM-based agents with external tool-use capabilities (Shao et al., 2024; Yu et al., 2025; Zheng et al., 2025). Furthermore, a series of work designs specialized algorithms tailored to agentic settings such as entropy-driven rollout and credit assignment (Jin et al., 2025; Dong et al., 2025; Ji et al., 2025; Zong et al., 2026; Chen et al., 2026). However, these methods all exhibit increasing redundant tool calls during training (Xie et al., 2026). To mitigate this, OTC-PO (Wang et al., 2025) introduces a tool-productivity reward term, -GRPO (Wu et al., 2025b) incorporates confidence thresholds, and HiPRAG (Wu et al., 2025a) applies hierarchical process rewards to evaluate the tool-call of each step. However, these reward-based methods either apply coarse-grained penalties on overall tool-call behavior where agents always learn to reduce tool calls indiscriminately to gain extra reward, leading to reward hacking, or evaluate each tool-call step individually but rely on external models or APIs (Wu et al., 2025a), introducing additional overhead and dependencies. SMART (Qian et al., 2025) instead constructs metacognitive SFT data offline, but static datasets cannot track the evolving knowledge boundary during RL training. Unlike these approaches, our proposed AKBE operates within the RL training loop, dynamically probing the model’s intrinsic knowledge boundary via on-policy dual-path (with-tool and no-tool) rollouts to construct boundary-guided supervisory signals that seamlessly integrate with any agentic RL algorithm as a plug-and-play module.
3 Preliminary
3.1 Task Definition
We consider an agentic setting where a language model policy iteratively interacts with an external tool environment to answer a given question . Following the ReAct paradigm (Yao et al., 2023), the agent generates a sequence of interleaved reasoning-and-action turns. At each turn , the agent produces a thought and an action conditioned on the current context . The action is either an invocation of an external tool, which returns an observation appended to the context, or a finish action that terminates the episode and returns the final answer. A complete interaction thus forms a trajectory , where denotes the final step. An outcome reward is assigned based on whether the final answer matches the ground truth. The learning objective is to maximize the expected reward over the training distribution :
| (1) |
3.2 Agentic Reinforcement Learning
While PPO (Schulman et al., 2017) provides a general policy optimization framework, its reliance on a separate value evaluator introduces substantial memory and training overhead. GRPO (Shao et al., 2024) addresses this by introducing the group-relative advantages, and has become the predominant algorithm in recent agentic RL research (Jin et al., 2025; Dong et al., 2025; Ji et al., 2025).
Specifically, for each question , GRPO samples a group of trajectories from the current policy and computes group-relative advantages:
| (2) |
The policy is updated by maximizing the clipped policy objective with a KL regularization term:
| (3) |
where is the importance sampling ratio, is the clipping threshold, and controls the strength of KL regularization against a reference policy . Note that tokens from tool observation are masked out during training.
4 Method
In this section, we present AKBE, which augments the agentic RL objective with knowledge boundary-guided training signals derived from dual-path rollouts. By probing whether the model needs external tools for each question and how many calls are minimally required, AKBE selects efficient trajectories as targeted on-policy optimization signals that eliminate redundant tool calls while reinforcing efficient tool use where external tools are genuinely needed. We illustrate the framework in Figure 2 and detail the training procedure in Algorithm 1.
4.1 Dual-Path Trajectory Rollout
For each question in a training batch, AKBE performs a dual-path trajectory rollout (with-tool and no-tool) in parallel:
With-tool trajectory rollout: We sample agentic rollouts where policy has access to external tools. Their trajectories consist of one or more tool calls. Let denote whether at least one with-tool trajectory yields a correct answer.
No-tool trajectory rollout: We sample rollouts in which tool access is disabled, forcing to rely solely on its parametric knowledge. Let denote whether at least one no-tool trajectory yields a correct answer.
We define the knowledge boundary of on question as:
| (4) |
where indicates that lies within the model’s intrinsic knowledge (i.e., tool calls are unnecessary), and indicates that external tools are required. Since the no-tool rollouts do not involve any tool interaction or environment latency, they incur substantially lower time consumption compared to with-tool rollouts, making this probing step computationally efficient.
4.2 Boundary-Guided Signal Construction
Based on the dual-path outcomes , we classify trajectories for each question into four categories and construct corresponding training signals:
Tool-dependent (=✓, =✗). The model can only answer correctly with tool calls (), where tool calls are necessary. We select the correct with-tool trajectory with the minimum number of tool calls as the target , reinforcing efficient tool-use patterns while preserving necessary tool invocations. When multiple correct trajectories share the same minimum tool-call count, we randomly sample one to avoid bias. At a finer granularity, each tool invocation reflects a dynamic step-level knowledge boundary decision: the model invokes a tool when its parametric knowledge is insufficient for a specific process reasoning step. Selecting the minimum tool-call trajectory thus reinforces the broadest achievable knowledge boundary at each step for a specific question.
Efficiency (=✓, =✓). The model can answer correctly without tools (), making tool calls redundant. We randomly select a correct no-tool trajectory as the target , teaching the model to bypass unnecessary tool invocations for questions within its knowledge boundary.
Hallucination (=✗, =✓). The model answers correctly without tools but incorrectly with tools (), indicating that tool calls introduce harmful noise or lead the model towards erroneous reasoning paths. We select a correct no-tool trajectory as the target , steering the model away from detrimental tool reliance for a specific question.
Both-wrong (=✗, =✗). Neither path yields a correct answer. No reliable supervisory signal can be constructed; we rely solely on the original RL objective for these instances.
4.3 Joint Training Objective
The overall training objective combines the original RL loss with the knowledge boundary-guided training objective:
| (5) |
where can be replaced by any classic agentic RL loss (e.g., DAPO, GSPO), and is the on-policy cross-entropy training objective over the selected target trajectories:
| (6) |
where denotes the set of questions with constructed signals from the Tool-dependent, Efficiency, and Hallucination categories respectively, and is the selected target trajectory for question as described in §4.2. The coefficient controls the strength of the boundary-guided objective relative to the RL loss.
Crucially, since both and are computed from on-policy rollouts of the current , the knowledge boundary is dynamically re-evaluated at every training step. As the model improves through RL training, the knowledge boundary for a specific question may shift, and the boundary-guided signal adapts accordingly. This on-policy nature distinguishes AKBE from approaches with static offline data which cannot track such dynamic evolution. Furthermore, AKBE is designed as a plug-and-play module: it can be seamlessly integrated with any agentic RL algorithm by simply adding the term during training regardless of the specific form of .
5 Experiments
| Method | Multi-Hop | Single-Hop | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Hotpot | 2Wiki | MuSiQue | Bamb. | Avg. | TC | TP | NQ | TriviaQA | PopQA | Avg. | TC | TP | |
| Backbone Model: Qwen3-4B | |||||||||||||
| ReAct | 30.42 | 32.92 | 12.83 | 44.80 | 30.01 | 3.43 | 8.75 | 26.75 | 53.53 | 35.34 | 41.31 | 2.27 | 18.19 |
| Search-o1 | 35.18 | 38.45 | 14.21 | 48.00 | 34.83 | 3.31 | 10.52 | 33.85 | 57.42 | 40.27 | 46.11 | 1.72 | 26.81 |
| R1-Searcher | 45.62 | 49.83 | 19.51 | 50.40 | 45.20 | 3.28 | 13.78 | 46.85 | 64.52 | 49.43 | 54.95 | 1.48 | 37.13 |
| Search-R1 | 44.80 | 50.66 | 18.94 | 52.00 | 45.40 | 3.16 | 14.33 | 45.98 | 65.17 | 49.18 | 54.97 | 1.41 | 38.99 |
| OTC-PO | 38.74 | 47.71 | 15.16 | 48.00 | 41.27 | 2.06 | 20.03 | 45.31 | 64.35 | 49.86 | 54.90 | 1.02 | 53.83 |
| -GRPO | 45.83 | 51.49 | 21.06 | 56.00 | 46.39 | 3.01 | 15.41 | 47.00 | 64.73 | 49.95 | 55.30 | 1.53 | 36.15 |
| Offline AKBE | 45.21 | 51.12 | 19.84 | 54.40 | 45.84 | 2.45 | 18.71 | 46.52 | 64.55 | 49.72 | 55.06 | 1.22 | 45.13 |
| AKBE (Ours) | 46.53 | 51.89 | 21.17 | 56.80 | 46.82 | 2.60 | 18.01 | 47.33 | 65.18 | 50.80 | 55.90 | 1.16 | 48.19 |
| Backbone Model: Qwen2.5-7B | |||||||||||||
| ReAct | 2.85 | 1.94 | 0.58 | 4.00 | 2.10 | 4.04 | 0.52 | 4.34 | 10.67 | 9.32 | 9.23 | 1.36 | 6.79 |
| Search-o1 | 18.52 | 16.73 | 5.42 | 16.00 | 16.10 | 3.92 | 4.11 | 15.87 | 32.45 | 22.18 | 25.37 | 1.28 | 19.82 |
| R1-Searcher | 47.21 | 44.85 | 20.63 | 44.00 | 43.02 | 3.91 | 11.00 | 45.12 | 63.28 | 49.05 | 54.07 | 1.08 | 50.06 |
| Search-R1 | 46.64 | 45.60 | 19.97 | 45.90 | 43.19 | 3.77 | 11.45 | 44.35 | 63.65 | 48.70 | 53.96 | 1.05 | 51.39 |
| OTC-PO | 42.16 | 43.72 | 20.48 | 44.80 | 40.72 | 1.56 | 26.10 | 48.72 | 63.46 | 47.64 | 53.89 | 1.00 | 53.89 |
| -GRPO | 48.02 | 47.92 | 21.60 | 48.00 | 45.13 | 3.34 | 13.51 | 47.89 | 64.21 | 48.18 | 54.35 | 1.03 | 52.76 |
| Offline AKBE | 47.38 | 46.81 | 20.95 | 46.40 | 44.22 | 2.68 | 16.50 | 47.15 | 63.92 | 48.05 | 54.08 | 1.01 | 53.54 |
| AKBE (Ours) | 48.69 | 47.59 | 22.23 | 48.80 | 45.24 | 2.94 | 15.39 | 49.64 | 64.63 | 47.91 | 54.59 | 1.00 | 54.59 |
5.1 Experiment Settings
Datasets. We evaluate AKBE on seven question answering benchmarks in a tool-augmented search setting. Following the setup of Search-R1(Jin et al., 2025), we deploy a lightweight search engine based on Wikipedia as the external tool environment. The benchmarks are organized into two categories: Multi-Hop QA, including HotpotQA (Yang et al., 2018), 2WikiMultihopQA (Ho et al., 2020), MuSiQue (Trivedi et al., 2022), and Bamboogle (Press et al., 2023), which require multi-step retrieval and reasoning; and Single-Hop QA, including Natural Questions (NQ) (Kwiatkowski et al., 2019), TriviaQA (Joshi et al., 2017), and PopQA (Mallen et al., 2022), which typically require a single retrieval. All benchmarks are evaluated using Exact Match (EM) as the primary metric. We additionally report Tool Calls (TC), defined as the average number of tool calls per question, and Tool Productivity (TP). , which measures accuracy per unit of tool usage.
Baselines. We compare AKBE against the following methods: (1) ReAct (Yao et al., 2023): a prompting-based approach, serving as the reference without RL training; (2) Search-o1 (Li et al., 2025): a framework that integrates an agentic search workflow into reasoning process; (3) R1-Searcher (Song et al., 2025) and (4) Search-R1 (Jin et al., 2025): two classic agentic RL frameworks that deploy GRPO for search enhancement; (5) OTC-PO (Wang et al., 2025): a reward shaping method with a tool-productivity term to penalize redundant tool calls; (6) -GRPO (Wu et al., 2025b): a reward shaping method which introduces a confidence-based threshold to reduce uncertainty; and (7) Offline AKBE: an offline variant of AKBE that uses the same strategy of knowledge boundary-guided signal construction but generates the signal data from a fixed GRPO-trained checkpoint, serving as a direct comparison to validate the necessity of on-policy dynamic signal construction. Note that additional implementation specifics of baselines and AKBE are provided in Section A.
5.2 Main Results of AKBE
We present the main results across two backbone models and seven benchmarks in Table 1. AKBE obtains the highest average EM score on both Multi-Hop and Single-Hop benchmarks while substantially reducing TC, yielding consistent TP improvements in most cases. On Qwen3-4B, AKBE improves EM by +1.85 on average across all seven benchmarks over its base method, while reducing TC by 18%, yielding approximately a 25% gain in tool productivity. The same effect holds on Qwen2.5-7B, confirming its generality across different model architectures and scales. In contrast, OTC-PO achieves the lowest TC across all settings (underlined in Table 1), but at a severe cost to accuracy, confirming that coarse-grained reward shaping incentivizes indiscriminate suppression of tool calls, leading to reward hacking. -GRPO avoids EM collapse through its confidence threshold but provides limited TC reduction. AKBE achieves a strictly better balance: larger TC reduction than -GRPO while simultaneously improving EM.
Comparing AKBE with its offline variant (Offline AKBE) reveals the importance of on-policy signal construction. Offline AKBE consistently underperforms AKBE in EM score despite achieving even lower TC, reflecting overly aggressive “reduce tool calls” signals generated from the frozen trained policy. The knowledge boundary captured by offline data reflects the model’s capability at a late training stage, which is overly optimistic for the weaker policy during early training. The resulting static boundary signals cannot align with the model’s evolving knowledge state throughout training, leading to premature tool suppression and degraded accuracy. This validates our core claim that dynamic on-policy knowledge boundary tracking is essential for achieving the EMTC balance.
5.3 Analysis
5.3.1 Plug-and-Play Generalization
Since AKBE enhances the model’s knowledge boundary awareness through auxiliary supervisory signals rather than modifying the RL reward or optimization procedure, it is naturally orthogonal to the choice of base agentic RL algorithm and can serve as a plug-and-play module. To verify this, we integrate AKBE with four agentic RL algorithms: GRPO (Shao et al., 2024), DAPO (Yu et al., 2025), GSPO (Zheng et al., 2025), and AEPO (Dong et al., 2025), each representing a distinct optimization strategy, such as dynamic sampling, sequence-level optimization, and entropy-driven exploration.
As shown in Table 2, AKBE consistently improves average EM and reduces TC across all four base algorithms. Notably, the improvements are consistent regardless of the base method’s inherent nature: DAPO already achieves low TC () due to its dynamic sampling strategy for diverse trajectories, yet AKBE still further reduces it to while improving EM ( Avg.). For GSPO and AEPO, which exhibit higher base TC (3.23 and 3.08), AKBE delivers larger TC reductions ( and ) alongside consistent EM gains ( and ). The TP metric improves uniformly across all four pairings, with gains ranging from to . These results confirm that AKBE acts as an efficiently orthogonal module. The boundary-guided training objective provides complementary learning signals that enhance tool call efficiency without interfering with the optimization dynamics of base RL algorithms.
| Method | Hotpot | 2Wiki | MuSi. | Bamb. | Avg. | TC | TP |
|---|---|---|---|---|---|---|---|
| GRPO | 44.80 | 50.66 | 18.94 | 52.00 | 45.40 | 3.16 | 14.33 |
| + AKBE | 46.53 | 51.89 | 21.17 | 56.80 | 46.82 | 2.60 | 18.01 |
| DAPO | 45.95 | 51.81 | 21.68 | 51.20 | 46.65 | 2.61 | 17.85 |
| + AKBE | 45.83 | 52.35 | 22.51 | 52.80 | 47.01 | 2.38 | 19.75 |
| GSPO | 47.07 | 49.25 | 22.68 | 49.60 | 45.69 | 3.23 | 14.16 |
| + AKBE | 47.62 | 49.78 | 23.50 | 48.00 | 46.24 | 2.84 | 16.28 |
| AEPO | 46.36 | 51.78 | 23.47 | 52.00 | 46.96 | 3.08 | 15.25 |
| + AKBE | 46.89 | 52.26 | 24.24 | 54.40 | 47.50 | 2.73 | 17.40 |
| Configuration | Avg. EM | TC | TP |
|---|---|---|---|
| GRPO | 45.40 | 3.16 | 14.33 |
| AKBE (Ours) | 46.86 | 2.60 | 18.02 |
| w/o Tool-dependent | 43.56 | 2.15 | 20.26 |
| w/o Efficiency | 46.50 | 2.92 | 15.93 |
| w/o Hallucination | 46.55 | 2.58 | 18.04 |
| w/ Tool-dependent only | 46.02 | 2.85 | 16.15 |
5.3.2 Ablation Study on Trajectory Categories
To understand the contribution of each signal category, we conduct ablation experiments by selectively removing individual categories from the knowledge boundary-guided training objective.
In Table 3, we find that removing Tool-dependent signals causes EM to drop below GRPO significantly, despite achieving the lowest TC. The remaining Efficiency and Hallucination categories exclusively supervise toward no-tool trajectories, leading to over-suppression of necessary tool calls and degraded task accuracy. This confirms that Tool-dependent signals serve as a crucial protective mechanism that prevents the efficiency-oriented signals from over-suppressing necessary tool calls. Removing Efficiency signals yields a TC increase, identifying it as the primary force for eliminating redundant tool calls. Removing Hallucination signals results in a modest EM drop while TC remains comparable, validating that the Hallucination category corrects harmful tool-call paths where tool invocations override correct internal reasoning, contributing to the EM improvement. Notably, Tool-dependent alone already improves over GRPO, demonstrating that AKBE remains effective even on complex questions where no-tool rollouts mostly fail. Full AKBE outperforms all subsets, confirming the three categories are complementary: Tool-dependent teaches when and how efficiently tools should be used, Efficiency teaches when tools are unnecessary, and Hallucination teaches when tools are harmful.
5.3.3 Selection Strategy of the Coefficient
We investigate the selection strategy of the coefficient that balances the RL loss and the boundary-guided objective. Figure 4 reports Avg. EM, TC, and TP on Qwen3-4B Multi-Hop as varies in . AKBE consistently outperforms GRPO in EM across , with achieving the best balance. As increases beyond 0.2, EM degrades sharply, indicating that an overly strong boundary-guided objective dominates the RL loss and leads to over-suppression of tool calls. Notably, TC and TP consistently outperform GRPO across all values, indicating that AKBE reliably improves tool-use efficiency regardless of signal strength. The optimal naturally balances the gradient contributions between the two objectives, as operates on at most one target trajectory per question while is computed over rollouts. For this, we provide a detailed theoretical analysis in Appendix B.
5.3.4 Trajectory Distribution During Training
To examine how the knowledge boundary evolves during training, we compare the distribution of the trajectory categories between the early (steps 1–40) and late (steps 201–240) training phases on Qwen2.5-7B Multi-Hop in Figure 3.
The most obvious change is a substantial decrease in Both-wrong proportion, indicating that agentic RL training progressively enables the model to solve previously intractable questions. Crucially, the Efficiency category shows the largest increase, demonstrating that AKBE successfully promotes knowledge internalization, where the model increasingly learns to answer questions using its parametric knowledge. Meanwhile, Hallucination decreases notably, confirming that Hallucination signals effectively correct harmful tool-call paths during training. These shifts validate two key aspects of our design: (1) the knowledge boundary is non-static during training, justifying on-policy signal construction over static offline approaches, and (2) AKBE’s boundary-guided objective and the RL objective work synergistically, where RL strengthens the tool-augmented reasoning capability of model while AKBE delivers knowledge boundary-guided efficiency signals that guide the model to maximize the utilization of its knowledge boundary, achieving efficient reasoning paths with minimal redundant tool calls.
5.3.5 Computational Overhead
A natural concern is whether the additional no-tool rollouts in AKBE introduce prohibitive computational overhead. Figure 5 compares the time consumption per training step between GRPO and AKBE on Qwen3-4B Multi-Hop. Surprisingly, AKBE is on average 15% faster than GRPO, despite performing additional no-tool rollouts per batch. This result arises from two factors: (1) no-tool rollouts complete substantially faster than with-tool rollouts as they involve no tool interaction or environment latency, and (2) as AKBE progressively reduces tool calls during training, the with-tool rollouts themselves become shorter with fewer tool calls, leading to accelerating step times in later training stages. This demonstrates that AKBE introduces little computational overhead, which pays for itself through the efficiency gains it induces in most cases. We further provide a detailed per-step comparison of tool call counts and response lengths in Appendix E.
6 Conclusion
In this paper, we presented AKBE, a simple but effective method that dynamically probes the model’s intrinsic knowledge boundary through on-policy dual-path rollouts during agentic RL training. By constructing knowledge boundary-guided supervisory signals, AKBE eliminates redundant tool calls while preserving necessary ones, and guides the model toward efficient tool-call patterns. Unlike reward shaping approaches that suffer from reward hacking, AKBE provides more fine-grained guidance at the instance level without modifying the RL objective, enabling simultaneous improvement in task accuracy and tool-call efficiency. Experiments across seven QA benchmarks and two backbone models validate its effectiveness, demonstrating that explicit on-policy knowledge boundary modeling is a promising and general strategy for efficient agentic reinforcement learning.
7 Limitations
Although AKBE achieves faster average training time than GRPO due to reduced tool calls in later stages, the additional no-tool rollouts do introduce extra computational cost in the early training phase when tool calls have not yet decreased. Future work could explore more efficient rollout strategies, such as adaptive sampling that selectively performs no-tool rollouts only for questions likely to be within the knowledge boundary. Besides, the coefficient is fixed throughout training, while the optimal balance between and may vary across training stages and task difficulties. An adaptive that adjusts per step based on the current trajectory distribution or task complexity could further improve performance.
References
- D. Chen, Z. Zong, Z. Ma, L. Luo, Y. Li, C. Li, P. Chen, and J. Jiang (2026) A 2 tgpo: agentic turn-group policy optimization with adaptive turn-level clipping. arXiv preprint arXiv:2605.06200. Cited by: §2.
- G. Dong, L. Bao, Z. Wang, K. Zhao, X. Li, J. Jin, J. Yang, H. Mao, F. Zhang, K. Gai, G. Zhou, Y. Zhu, J. Wen, and Z. Dou (2025) Agentic entropy-balanced policy optimization. External Links: 2510.14545, Link Cited by: §A.2, §1, §2, §3.2, §5.3.1.
- L. Feng, Z. Xue, T. Liu, and B. An (2025) Group-in-group policy optimization for llm agent training. arXiv preprint arXiv:2505.10978. Cited by: §1.
- X. Ho, A. Duong Nguyen, S. Sugawara, and A. Aizawa (2020) Constructing a multi-hop QA dataset for comprehensive evaluation of reasoning steps. In Proceedings of the 28th International Conference on Computational Linguistics, D. Scott, N. Bel, and C. Zong (Eds.), Barcelona, Spain (Online), pp. 6609–6625. External Links: Link, Document Cited by: §A.3, §5.1.
- Y. Ji, Z. Ma, Y. Wang, G. Chen, X. Chu, and L. Wu (2025) Tree search for llm agent reinforcement learning. External Links: 2509.21240, Link Cited by: §2, §3.2.
- B. Jin, H. Zeng, Z. Yue, J. Yoon, S. Arik, D. Wang, H. Zamani, and J. Han (2025) Search-r1: training llms to reason and leverage search engines with reinforcement learning. arXiv preprint arXiv:2503.09516. Cited by: §A.1, §A.2, §A.6, §2, §3.2, §5.1, §5.1.
- M. Joshi, E. Choi, D. Weld, and L. Zettlemoyer (2017) TriviaQA: a large scale distantly supervised challenge dataset for reading comprehension. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), R. Barzilay and M. Kan (Eds.), Vancouver, Canada, pp. 1601–1611. External Links: Link, Document Cited by: §A.3, §5.1.
- T. Kwiatkowski, J. Palomaki, O. Redfield, M. Collins, A. Parikh, C. Alberti, D. Epstein, I. Polosukhin, J. Devlin, K. Lee, K. Toutanova, L. Jones, M. Kelcey, M. Chang, A. M. Dai, J. Uszkoreit, Q. Le, and S. Petrov (2019) Natural questions: a benchmark for question answering research. Transactions of the Association for Computational Linguistics 7, pp. 452–466. External Links: Link, Document Cited by: §A.3, §5.1.
- X. Li, G. Dong, J. Jin, Y. Zhang, Y. Zhou, Y. Zhu, P. Zhang, and Z. Dou (2025) Search-o1: agentic search-enhanced large reasoning models. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pp. 5420–5438. Cited by: §5.1.
- Z. Luo, Z. Luo, M. Zhang, and R. Mao (2026) TabTracer: monte carlo tree search for complex table reasoning with large language models. arXiv preprint arXiv:2602.14089. Cited by: §1.
- A. Mallen, Asai,Akari, V. Zhong, R. Das, H. Hajishirzi, and D. Khashabi (2022) When not to trust language models: investigating effectiveness and limitations of parametric and non-parametric memories. arXiv preprint. Cited by: §A.3, §5.1.
- O. Press, M. Zhang, S. Min, L. Schmidt, N. A. Smith, and M. Lewis (2023) Measuring and narrowing the compositionality gap in language models. External Links: 2210.03350, Link Cited by: §A.3, §5.1.
- C. Qian, E. C. Acikgoz, H. Wang, X. Chen, A. Sil, D. Hakkani-Tur, G. Tur, and H. Ji (2025) SMART: self-aware agent for tool overuse mitigation. In Findings of the Association for Computational Linguistics: ACL 2025, pp. 4604–4621. Cited by: §2.
- Qwen, :, A. Yang, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Li, D. Liu, F. Huang, H. Wei, H. Lin, J. Yang, J. Tu, J. Zhang, J. Yang, J. Yang, J. Zhou, J. Lin, K. Dang, K. Lu, K. Bao, K. Yang, L. Yu, M. Li, M. Xue, P. Zhang, Q. Zhu, R. Men, R. Lin, T. Li, T. Tang, T. Xia, X. Ren, X. Ren, Y. Fan, Y. Su, Y. Zhang, Y. Wan, Y. Liu, Z. Cui, Z. Zhang, and Z. Qiu (2025) Qwen2.5 technical report. External Links: 2412.15115, Link Cited by: §A.7.
- T. Schick, J. Dwivedi-Yu, R. Dessi, R. Raileanu, M. Lomeli, E. Hambro, L. Zettlemoyer, N. Cancedda, and T. Scialom (2023) Toolformer: language models can teach themselves to use tools. In Thirty-seventh Conference on Neural Information Processing Systems, External Links: Link Cited by: §1.
- J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov (2017) Proximal policy optimization algorithms. External Links: 1707.06347, Link Cited by: §3.2.
- Z. Shao, P. Wang, Q. Zhu, R. Xu, J. Song, X. Bi, H. Zhang, M. Zhang, Y. K. Li, Y. Wu, and D. Guo (2024) DeepSeekMath: pushing the limits of mathematical reasoning in open language models. External Links: 2402.03300, Link Cited by: §1, §2, §3.2, §5.3.1.
- G. Sheng, C. Zhang, Z. Ye, X. Wu, W. Zhang, R. Zhang, Y. Peng, H. Lin, and C. Wu (2024) HybridFlow: a flexible and efficient rlhf framework. arXiv preprint arXiv: 2409.19256. Cited by: §A.7.
- S. Si, H. Zhao, Y. Lei, Q. Wang, D. Chen, Z. Wang, Z. Wang, K. Luo, Z. Wang, G. Chen, et al. (2026) From context to skills: can language models learn from context skillfully?. arXiv preprint arXiv:2604.27660. Cited by: §1.
- H. Song, J. Jiang, Y. Min, J. Chen, Z. Chen, W. X. Zhao, L. Fang, and J. Wen (2025) R1-searcher: incentivizing the search capability in llms via reinforcement learning. arXiv preprint arXiv:2503.05592. Cited by: §5.1.
- H. Trivedi, N. Balasubramanian, T. Khot, and A. Sabharwal (2022) MuSiQue: multihop questions via single-hop question composition. Transactions of the Association for Computational Linguistics 10, pp. 539–554. External Links: Link, Document Cited by: §A.3, §5.1.
- H. Wang, C. Qian, W. Zhong, X. Chen, J. Qiu, S. Huang, B. Jin, M. Wang, K. Wong, and H. Ji (2025) Acting less is reasoning more! teaching model to act efficiently. arXiv preprint arXiv:2504.14870. Cited by: §1, §1, §2, §5.1.
- L. Wang, N. Yang, X. Huang, B. Jiao, L. Yang, D. Jiang, R. Majumder, and F. Wei (2022) Text embeddings by weakly-supervised contrastive pre-training. arXiv preprint arXiv:2212.03533. Cited by: §A.6.
- P. Wu, M. Zhang, K. Wan, W. Zhao, K. He, X. Du, and Z. Chen (2025a) Hiprag: hierarchical process rewards for efficient agentic retrieval augmented generation. arXiv preprint arXiv:2510.07794. Cited by: §2.
- P. Wu, M. Zhang, X. Zhang, X. Du, and Z. Chen (2025b) Search wisely: mitigating sub-optimal agentic searches by reducing uncertainty. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pp. 19734–19745. Cited by: §1, §2, §5.1.
- R. Xie, D. Gopinath, D. Qiu, D. Lin, H. Sun, S. Potdar, and B. Dhingra (2026) Over-searching in search-augmented large language models. arXiv preprint arXiv:2601.05503. Cited by: §1, §2.
- A. Yang, A. Li, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Gao, C. Huang, C. Lv, C. Zheng, D. Liu, F. Zhou, F. Huang, F. Hu, H. Ge, H. Wei, H. Lin, J. Tang, J. Yang, J. Tu, J. Zhang, J. Yang, J. Yang, J. Zhou, J. Zhou, J. Lin, K. Dang, K. Bao, K. Yang, L. Yu, L. Deng, M. Li, M. Xue, M. Li, P. Zhang, P. Wang, Q. Zhu, R. Men, R. Gao, S. Liu, S. Luo, T. Li, T. Tang, W. Yin, X. Ren, X. Wang, X. Zhang, X. Ren, Y. Fan, Y. Su, Y. Zhang, Y. Zhang, Y. Wan, Y. Liu, Z. Wang, Z. Cui, Z. Zhang, Z. Zhou, and Z. Qiu (2025) Qwen3 technical report. External Links: 2505.09388, Link Cited by: §A.7.
- Z. Yang, P. Qi, S. Zhang, Y. Bengio, W. Cohen, R. Salakhutdinov, and C. D. Manning (2018) HotpotQA: a dataset for diverse, explainable multi-hop question answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, E. Riloff, D. Chiang, J. Hockenmaier, and J. Tsujii (Eds.), Brussels, Belgium, pp. 2369–2380. External Links: Link, Document Cited by: §A.3, §5.1.
- S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. R. Narasimhan, and Y. Cao (2023) ReAct: synergizing reasoning and acting in language models. In The Eleventh International Conference on Learning Representations, External Links: Link Cited by: §1, §3.1, §5.1.
- Q. Yu, Z. Zhang, R. Zhu, Y. Yuan, X. Zuo, YuYue, W. Dai, T. Fan, G. Liu, J. Liu, L. Liu, X. Liu, H. Lin, Z. Lin, B. Ma, G. Sheng, Y. Tong, C. Zhang, M. Zhang, R. Zhang, W. Zhang, H. Zhu, J. Zhu, J. Chen, J. Chen, C. Wang, H. Yu, Y. Song, X. Wei, H. Zhou, J. Liu, W. Ma, Y. Zhang, L. Yan, Y. Wu, and M. Wang (2025) DAPO: an open-source LLM reinforcement learning system at scale. In The Thirty-ninth Annual Conference on Neural Information Processing Systems, External Links: Link Cited by: §A.4, §1, §2, §5.3.1.
- C. Zheng, S. Liu, M. Li, X. Chen, B. Yu, C. Gao, K. Dang, Y. Liu, R. Men, A. Yang, J. Zhou, and J. Lin (2025) Group sequence policy optimization. External Links: 2507.18071, Link Cited by: §A.4, §2, §5.3.1.
- Z. Zong, D. Chen, Y. Li, Q. Yi, B. Zhou, C. Li, B. Qian, P. Chen, and J. Jiang (2026) AT 2 po: agentic turn-based policy optimization via tree search. arXiv preprint arXiv:2601.04767. Cited by: §A.2, §1, §2.
Appendix A Implementation Details
A.1 Reward Design
Our training pipeline employs a binary outcome reward that combines answer correctness with a structural format requirement. The correctness signal follows the reward formulation of Search-R1 (Jin et al., 2025), using Exact Match as the primary evaluation criterion.
Exact Match Reward.
Given the final answer extracted from the agent’s trajectory and the ground-truth answer , the EM reward is defined as:
| (7) |
This strict binary formulation eliminates the ambiguity of partial-credit scoring and drives the policy toward fully correct answers, providing a clear optimization signal for agentic RL.
Format Constraint.
In addition to correctness, each trajectory must satisfy a structural validity requirement. The response must contain both a reasoning trace wrapped by <think>...</think> tags and a final answer wrapped by <answer>...</answer> tags, with the answer further enclosed in \boxed{}. The format indicator is:
| (8) |
Responses violating this schema receive no credit regardless of answer correctness, ensuring reliable tool-call parsing and final-answer extraction.
Final Reward.
The overall reward combines both components:
| (9) |
A trajectory earns the maximal reward of 1 only when it satisfies the format requirement and delivers an exactly correct answer; format violations are explicitly penalized with .
A.2 Prompt Template
AKBE requires two prompt templates for its dual-path rollouts, as shown in Figure 6 and Figure 7.
With-tool Prompt.
The with-tool template (Figure 6) follows the tag-based format adopted in prior agentic RL work (Jin et al., 2025; Dong et al., 2025; Zong et al., 2026). Each rollout is structured into semantically distinct regions delimited by dedicated tag pairs: reasoning steps are verbalized within <think></think>, retrieval queries are issued via <search></search>, environment observations are injected within <result></result>, and the final prediction is emitted within <answer></answer> with the canonical answer enclosed in \boxed{} for Exact Match extraction.
No-tool Prompt.
The no-tool template (Figure 7) removes all tool-related instructions and tags (<search> and <result>), retaining only the reasoning (<think></think>) and answer (<answer></answer>) components. This forces the model to generate answers solely from its parametric knowledge, enabling AKBE to probe the knowledge boundary by comparing correctness across the two paths.
A.3 Datasets
We conduct experiments on two categories of widely-used question answering benchmarks to evaluate the effectiveness of our proposed AKBE.
Multi-Hop QA.
This category evaluates multi-turn tool use and compositional reasoning, where correct answers cannot be obtained from a single retrieved passage. HotpotQA (Yang et al., 2018) is a large-scale Wikipedia-derived benchmark with supporting-fact annotations, serving as a widely used testbed for multi-hop question answering. 2WikiMultiHopQA (Ho et al., 2020) combines Wikipedia passages with Wikidata triples, producing questions that require explicit multi-hop entity reasoning. MuSiQue (Trivedi et al., 2022) contains approximately 25k questions spanning 2–4 reasoning hops, synthesized through controlled composition of single-hop primitives to probe fine-grained reasoning depth. Bamboogle (Press et al., 2023) offers a small but adversarial set of compositional queries, serving as a robustness probe for agentic RL policies.
Single-Hop QA.
This category verifies performance on single-step retrieval tasks. Natural Questions (NQ) (Kwiatkowski et al., 2019) aggregates real user queries answered from Wikipedia and serves as a standard benchmark for retrieval-augmented generation. TriviaQA (Joshi et al., 2017) features substantial lexical and syntactic divergence between questions and supporting evidence, testing robustness to surface variation. PopQA (Mallen et al., 2022) is an entity-centric benchmark designed to separate the contribution of external retrieval from parametric memorization, making it a natural diagnostic for whether the policy genuinely leverages the search tool versus relying on memorized facts.
A.4 AKBE Settings
For implementation details of our AKBE, we use a training batch size of , a mini-batch size of , and a maximum response length of . During rollout, we use a rollout size of for with-tool rollout and a rollout size of for no-tool rollout, with the maximum tool usage set to . The clipping thresholds for the AKBE objective are set to (the same as GRPO). Following prior work (Yu et al., 2025; Zheng et al., 2025), we remove the KL regularization term () to allow the policy to explore diverse rollout strategies. The AKBE coefficient is set to for both Multi-Hop and Single-Hop settings, which empirically balances the gradient contributions between and (see §B for detailed analysis).
A.5 Baseline Settings
| Config | Value |
| optimizer | AdamW |
| learning rate | 1e-6 |
| clip_ratio | 0.2 |
| training batch size | 64 |
| PPO mini batch size | 8 |
| rollout_n | 16 |
| max prompt length | 2000 |
| max response length | 6192 |
| max tool-call turns | 6 |
| reward metrics | EM |
| retriever | local wiki |
| top-K retrieval passages | 3 |
Table 4 summarizes the shared hyperparameters used across all RL-based baselines. For method-specific configurations, we follow the settings reported in the respective original papers. All baselines are trained without any additional SFT phase. We select and report results from the checkpoint achieving the highest average EM across all evaluation benchmarks.
A.6 Search Tool Environment
Our search tool environment follows the setup of Search-R1 (Jin et al., 2025). We use a Wikipedia snapshot as the retrieval corpus with e5-base-v2 (Wang et al., 2022) as the dense retriever. The knowledge base contains approximately 21M Wikipedia entries, providing broad factual coverage for both single-hop and multi-hop queries. At each turn where the policy emits a retrieval action, the search engine scores candidate passages against the query and returns the top-3 most relevant entries, which are injected into the context as tool observations for reasoning.
A.7 Hardware and Artifacts
All training and evaluation experiments are conducted on a single node with 8 NVIDIA H20 GPUs. We adopt two publicly available checkpoints as backbone policies: Qwen3-4B (Yang et al., 2025) and Qwen2.5-7B (Qwen et al., 2025), selected for their strong reasoning capabilities and demonstrated compatibility with agentic post-training. Our training infrastructure is built on the VeRL framework (Sheng et al., 2024), a hybrid-controller RL system whose modular rollout interface supports the multi-turn, tool-interactive rollout schedule required by AKBE’s dual-path design.
Appendix B Theoretical Analysis for Coefficient
We provide a theoretical analysis for the optimal value of from the perspective of gradient contribution balancing between and .
Setup.
Consider a training batch of questions. For each question , the GRPO objective is computed over with-tool rollout trajectories, while the AKBE objective selects at most one target trajectory . The total training loss is:
| (10) |
Gradient Analysis.
The per-question gradient contribution from the GRPO loss involves trajectories:
| (11) |
where is the group-relative advantage. Under the assumption that each trajectory contributes approximately equal gradient magnitude , the expected per-question gradient norm is:
| (12) |
where is the standard deviation of the advantage estimates within the group.
For the AKBE loss, each question contributes at most one target trajectory:
| (13) |
with gradient norm approximately . However, not all questions produce a signal. Let denote the proportion of questions with constructible signals (). The effective per-question gradient contribution from AKBE (averaged over the batch) is:
| (14) |
Balancing Condition.
For the two objectives to contribute comparably to the overall parameter update, we require:
| (15) |
Substituting the per-question estimates and noting that the GRPO loss normalizes over trajectories while AKBE operates on single trajectories:
| (16) |
Practical Estimate.
In our setting, the advantage standard deviation under binary rewards () with group-relative normalization is by construction. The signal proportion is typically high (around 70–80% of questions produce at least one correct trajectory in either path). However, the key scaling factor is the ratio of trajectory counts: aggregates gradients from trajectories per question (each weighted by ), while uses exactly one trajectory at full weight. To prevent the single AKBE trajectory from dominating the RL trajectories, should scale as:
| (17) |
With in our experiments, this yields , closely matching our empirical optimum of . Considering dynamic factors such as task difficulty and signal proportion variability, we find that constitutes a reasonable range in our experimental setting.
Appendix C Cross-Entropy vs. DPO for Signal Integration
A natural concern is whether the boundary-guided signals could be integrated via preference optimization (e.g., DPO) rather than cross-entropy. Since AKBE’s signal construction identifies preferred trajectories for each question, one could additionally select rejected trajectories and apply a DPO-style objective:
| (18) |
where and denote the preferred and rejected trajectories respectively.
Signal Construction Difference.
The AKBE based on cross-entropy formulation retains only positive signals, supervising the model toward the selected target trajectory without explicit penalization of alternatives. In contrast, the DPO-based variant additionally selects rejected trajectories from incorrect or inefficient rollouts, specifically choosing the trajectory with the highest tool-call count (i.e., maximum divergence from the preferred pattern) as the negative signal. This creates explicit preference pairs that contrast efficient and inefficient tool-use behaviors.
Experimental Comparison.
We compare three configurations on Qwen3-4B Multi-Hop: standard GRPO, AKBE with cross-entropy (our method), and AKBE with DPO-based integration. Figure 8 presents the EM evaluation and training reward dynamics. is applied to both experiments.
Results and Analysis.
As shown in Figure 8, the DPO-based variant initially shows promising performance, even surpassing GRPO during the mid-training phase (steps 80–140). However, it subsequently undergoes a sudden collapse, with both EM and training reward dropping sharply after step 150. In contrast, AKBE (CE based) maintains stable and monotonic improvement throughout training.
We attribute this instability to the inherent similarity between preferred and rejected trajectory patterns. Both positive and negative signals are trajectories that involve reasoning and tool calls; they differ only in the tool-call strategies or final correctness. As DPO explicitly reduces the probability of rejected trajectories, the model gradually learns to penalize the shared tool-call patterns present in both preferred and rejected samples, rather than learning the fine-grained distinction between efficient and inefficient tool use. This causes a progressive over-suppression of tool-call behavior that eventually leads to training collapse. Cross-entropy avoids this failure mode by providing positive-only supervision: it teaches the model what efficient patterns looks like without explicitly penalizing alternatives, resulting in a more stable and targeted optimization signal for knowledge boundary enhancement.
Appendix D Reliability of Knowledge Boundary Estimation
We discuss a potential concern of whether the knowledge boundary estimation based on “at least one correct” among no-tool rollouts ( in our experiment settings) is reliable or whether it is dominated by questions where the model merely guesses correctly by chance.
Specifically, we analyze the distribution of correct rollout counts for all questions classified as NT=1 across training. As shown in Figure 9, NT=1 questions have 5.0/8 correct rollouts in early training and 5.2/8 in late training on average. Only 18.1% (early) and 16.7% (late) of NT=1 questions have exactly 1/8 correct, while the majority (63.4% early, 67.1% late) achieve 4/8 correct rollouts. Notably, 30.9% (early) to 36.7% (late) of questions achieve a perfect 8/8, indicating that the model fully and reliably leverages the parametric knowledge.
These results demonstrate that the knowledge boundary estimation is in high confidence rather than noise-driven. Furthermore, the improvement from early to late training confirms that AKBE’s on-policy design progressively strengthens the reliability of boundary estimation as training progresses. For the minority of potentially noisy cases (1/8 correct), the small coefficient ensures that these weak signals cannot override the dominant RL objective, and the on-policy mechanism provides an additional indemnification.
Appendix E Supplementary Overhead Analysis
To complement the per-step overhead comparison in the main text (§5.3.5), we provide a supplementary analysis of the two underlying factors that drive AKBE’s computational efficiency: tool call frequency and response length.
Tool Call Count.
Figure 10 (Left) shows the total number of search calls per training batch across training steps. Both methods start at a similar level (320 calls per batch), but their trajectories diverge as training progresses. GRPO’s tool call count increases steadily, rising from approximately 320 to over 370 by step 300, reflecting the well-known tendency of RL-trained agents to escalate tool use when rewarded only for correctness. In contrast, AKBE maintains a relatively stable tool call count around 290–300 throughout training, with a slight decrease in the middle stages (steps 50–150). On average, AKBE reduces tool calls by 8.4% relative to GRPO across the entire training process.
Response Length.
Figure 10 (Right) reveals significant difference in mean response length. GRPO exhibits a strong upward trend, with average response length growing from 2,300 tokens to over 2,500 tokens by step 300. This increase correlates with more tool calls generating longer multi-turn interaction sequences. AKBE, by contrast, shows a decreasing trend in the early-to-middle training phase (steps 1–100), stabilizing around 1,600–1,700 tokens thereafter. The overall response length reduction is 23.0%, directly translating to lower inference latency and computational cost.
These trends confirm that AKBE’s knowledge boundary-guided signals produce a compounding efficiency effect: by teaching the model to avoid unnecessary tool calls early in training, subsequent rollouts become inherently shorter, which in turn accelerates both training and inference.
Appendix F Case Study
We present representative examples from three AKBE signal categories during training (Qwen3-4B Multi-Hop). These cases illustrate how the dual-path comparison reveals the model’s knowledge boundary and guides signal construction. For each case, we show the complete reasoning trajectories from both paths, the signal category classification, and which trajectory is selected as the supervision target.
Analysis of Case #1 (Efficiency).
This case demonstrates the efficiency signal category, where the model possesses sufficient parametric knowledge to answer correctly without external retrieval. In the with-tool path, the model explicitly states “I’m not immediately familiar with Donker Mag” and initiates a search, despite the no-tool path revealing that it can readily recall the album’s association with Die Antwoord from its internal knowledge. This discrepancy exposes a Level-1 knowledge boundary violation: the model defaults to tool use even when unnecessary. By selecting the no-tool trajectory as the supervision target, AKBE teaches the model to trust its parametric memory for well-known facts, reducing redundant tool calls and improving inference efficiency.
Analysis of Case #2 (Tool-dependent).
This case illustrates the tool-dependent category, where external retrieval is genuinely necessary. The no-tool path exposes a clear knowledge gap: the model hallucinates a plausible but incorrect answer (“Warren Buffett”), cycling through multiple guesses without arriving at the correct one (Seth Klarman). In contrast, the with-tool path formulates a targeted query (“Baupost Group founder”) that immediately retrieves the relevant passage, enabling correct reasoning. Here, the with-tool trajectory with minimum tool calls (TC=1) is selected as the target. This signal reinforces Level-2 boundary awareness: for questions beyond parametric knowledge, the model should learn efficient retrieval patterns that resolve the query in a single tool call.
Analysis of Case #3 (Hallucination).
This case reveals a subtle failure mode: retrieval-induced hallucination. The model searches for “Arline Burks Gant death date” but the retriever returns irrelevant passages about “Barbara Stoddard Burks.” Rather than recognizing the retrieval failure, the with-tool path incorrectly conflates the two individuals based on the shared surname, arriving at a wrong answer through flawed reasoning (“maybe Arline Burks Gant is the same as Barbara Stoddard Burks?”). Meanwhile, the no-tool path correctly recalls that Maurice Pialat died in 2003 and Arline Burks Gant lived until 2011. This case demonstrates that tool use can be actively harmful when retrieval results are noisy or off-target. By selecting the no-tool trajectory as the target, AKBE guides the model to avoid over-reliance on retrieval for queries where parametric knowledge is more reliable than noisy search results.
These three cases collectively illustrate how AKBE’s dual-path comparison mechanism adaptively identifies the appropriate behavior at the knowledge boundary. The signal construction does not uniformly favor either path; instead, it selects the trajectory that reflects the most appropriate tool-use decision for each specific query. This fine-grained, per-question supervision enables the model to develop nuanced knowledge boundary awareness across diverse scenarios.
| Question: Donker Mag is an album by what African hip hop group? Ground Truth: Die Antwoord |
|---|
| With-tool path (TC=1, EM=1.0) |
| <think> Okay, the user is asking about the album “Donker Mag” and wants to know which African hip hop group it belongs to. First, I need to figure out who the artists are associated with this album. I’m not immediately familiar with “Donker Mag,” so I should look it up. </think> |
| <search> Donker Mag album </search> |
| <result> |
| Page 1: “Die Antwoord” However the album was met with average to poor reviews by many critics. “Ten$Ion” was followed up with a non-album single, “XP N$IV $H1T”. Following the release of “XP N$IV $H1T”, Die Antwoord set out on a brief tour across Europe from June to July. Soon after they began touring, snippets for a new single called “Cookie Thumper!”, which was to be accompanied by a music video, were released. Around the release of their new single, Die Antwoord also announced the title of their third album, “Donker Mag”, which was released on 3 June 2014. The video for “Cookie Thumper!” was released on 27 May 2014 as the album’s lead single. |
| Page 2: “Magnet (magazine)” Magnet is a music magazine that generally focuses on alternative, independent, or out-of-the-mainstream bands. The magazine is published four times a year, and is independently owned and edited by Eric T. Miller. Music magazines with a similar focus in the 1990s era included “Option”, “Raygun”, and “Alternative Press”. The first issue of “Magnet” came out in mid-1993. |
| Page 3: “Green Monkey Records” (CDR GM1028) The OF - Escape Goat (CD GM1029) Jim McIver - Sunlight Reaches (CD GM1030) Tom Dyer’s New Pagan Gods - History Of Northwest Rock Vol. 1, 1959-1968 (CD GM1031) Fur For Fairies (CD GM1032) The Fuzz - Best Kept Secret… |
| </result> |
| <think> Hmm, the first search result mentions “Die Antwoord” in relation to the album “Donker Mag.” Die Antwoord is a South African hip hop group. The other results don’t seem relevant. So, the answer should be Die Antwoord. </think> |
| <answer> The final answer is \boxed{Die Antwoord}. </answer> |
| No-tool path (TC=0, EM=1.0) Selected as target |
| <think> Okay, the user is asking about the album “Donker Mag” and which African hip hop group it belongs to. First, I need to recall if I know any African hip hop groups that have an album with that title. “Donker Mag” sounds Dutch. Let me think… I recall that Die Antwoord is a South African hip hop group known for performing in Afrikaans and English. “Donker Mag” means “Dark Force” in Afrikaans/Dutch. Yes, I’m fairly confident that “Donker Mag” is the third studio album by Die Antwoord, released on 3 June 2014. The group consists of Ninja and Yolandi Visser, and they are from Cape Town, South Africa. </think> |
| <answer> The final answer is \boxed{Die Antwoord}. </answer> |
| Question: The Baupost Group asked which American investor, hedge fund manager, and philanthropist to run the hedge fund? Ground Truth: Seth Klarman |
|---|
| With-tool path (TC=1, EM=1.0) Selected as target (min TC) |
| <think> Okay, let’s see. The question is about the Baupost Group asking which American investor, hedge fund manager, and philanthropist to run the hedge fund. I need to figure out the answer. First, I should recall who the Baupost Group is. Baupost is a hedge fund firm, right? They might have had a notable founder or a prominent figure associated with them. The question mentions that they asked a specific person to run the fund. So, I need to find the name of that individual. Maybe I can search for “Baupost Group founder” to find the founder’s name. Let me check that. </think> |
| <search> Baupost Group founder </search> |
| <result> |
| Page 1: “Baupost Group” …and $29.9 billion as of December 31, 2013. Baupost Group The Baupost Group is a long-only hedge fund founded in 1982 by Harvard Professor William Poorvu and partners Howard Stevenson, Jordan Baruch and Isaac Auerbach. Seth Klarman, who was asked by Poorvu to help run the fund, remains at its head today. Baupost Group’s investment philosophy emphasizes risk management. The firm, one of the largest hedge funds in the world, is a value investing manager. According to Bloomberg L.P., Baupost is ranked 4th in net gains since inception. |
| Page 2: “Baupost Group” Baupost Group The Baupost Group is a long-only hedge fund founded in 1982 by Harvard Professor William Poorvu… Seth Klarman, who was asked by Poorvu to help run the fund, remains at its head today. |
| Page 3: “Baupost Group” …Baupost CEO, Seth Klarman, explained in a speech to MIT students that investment research driven by emotion is risky and can lead to a bad investment. From its founding the firm’s three private partnerships have generated an average annual return of 19%. |
| </result> |
| <think> So, from the search result, the Baupost Group was founded by William Poorvu and others. Seth Klarman was asked by Poorvu to help run the fund and remains at its head. The question is asking who the Baupost Group asked to run the hedge fund. The answer is Seth Klarman. Let me confirm: the search results consistently mention that Seth Klarman is the current head, and he was asked by Poorvu to help run the fund. So the answer should be Seth Klarman. </think> |
| <answer> The final answer is \boxed{Seth Klarman}. </answer> |
| No-tool path (TC=0, EM=0.0) |
| <think> Okay, let’s see. The question is about the Baupost Group asking which American investor, hedge fund manager, and philanthropist to run the hedge fund. Hmm. I remember that Baupost Group is a hedge fund, right? So they might have had a managing director or someone else appointed. Wait, I think the answer is someone like Warren Buffett. No, wait, Buffett is the CEO of Berkshire Hathaway. Maybe it’s another person. Let me think. There’s a person named Bill Ackman? No, he’s more of an activist investor. Or maybe Roger L. Ivelt? No, that’s not right. Wait, I think the correct answer is Dick Fuld. No, he was the CEO of Lehman Brothers. That’s not it. Maybe someone else. The Baupost Group was founded by Robert F. Draper and Jerry Rhoads. But the question is about who they asked to run the fund. Perhaps the answer is Warren Buffett after all. </think> |
| <answer> The final answer is \boxed{Warren Buffett}. </answer> |
| Question: Who died first, Maurice Pialat or Arline Burks Gant? Ground Truth: Maurice Pialat |
|---|
| With-tool path (TC=4, EM=0.0) |
| <think> Okay, so the question is asking who died first between Maurice Pialat and Arline Burks Gant. I need to find the dates of their deaths and compare them. First, I should look up when Maurice Pialat died. Let me search for that. </think> |
| <search> Maurice Pialat death date </search> |
| <result> |
| Page 1: “Maurice Pialat” Maurice Pialat (; 31 August 1925 – 11 January 2003) was a French film director, screenwriter and actor noted for the rigorous and unsentimental style of his films. His work is often described as being “realist”. Pialat was born in Cunlhat, Puy-de-Dôme, France. |
| Page 2: “Maurice Pialat” …He was always willing to bend his narratives around experience… Maurice Pialat (; 31 August 1925 – 11 January 2003) was a French film director… |
| Page 3: “Maurice Ronet” …He died in a Paris hospital, of cancer… Maurice Ronet (13 April 1927 – 14 March 1983) was a French film actor, director, and writer. |
| </result> |
| <think> So Maurice Pialat died on 11 January 2003. Now, I need to find when Arline Burks Gant died. Let me look that up. </think> |
| <search> Arline Burks Gant death date </search> |
| <result> |
| Page 1: “Barbara Stoddard Burks” …years later. Burks married Herman Ramsperger, a National Research Fellow in chemistry at Stanford, in 1927; they remained married until his death in 1932. In 1943, she became engaged to longtime friend Robert Cook, but she died before they could be married. Burks died on May 25, 1943, when she “either fell or jumped to her death from the George Washington Bridge” in New York City. |
| Page 2: “Richard Fenner Burges” …he was married to Ethel Petrie Shelton… He died in El Paso on January 13, 1945. |
| Page 3: “Abbie G. Rogers” …died unexpectedly on May 21, 1894, age 53 in New York City. She had been undergoing an operation to remove a tumor. |
| </result> |
| <think> Hmm, the first result mentions Barbara Stoddard Burks, but that’s a different person from Arline Burks Gant. The second result is about Richard Fenner Burges. The third is about Abbie G. Rogers. None of these are about Arline Burks Gant. Wait, maybe Arline Burks Gant is the same as Barbara Stoddard Burks? The name “Burks” appears in both. Alternatively, maybe there’s a mix-up. Assuming that Arline Burks Gant is related to the “Burks” who died in 1943, then she died before Maurice Pialat (2003). Therefore, Arline Burks Gant died first. </think> |
| <answer> The final answer is \boxed{Arline Burks Gant}. </answer> |
| No-tool path (TC=0, EM=1.0) Selected as target |
| <think> The question asks who died first between Maurice Pialat and Arline Burks Gant. I recall that Maurice Pialat was a well-known French film director who passed away on January 11, 2003. As for Arline Burks Gant, she is less well-known, but I believe she lived until 2011. Therefore, Maurice Pialat died first in 2003, before Arline Burks Gant. </think> |
| <answer> The final answer is \boxed{Maurice Pialat}. </answer> |