arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2602.01709cs.CL
收藏

ARTIS: 通过迭代模拟实现代理风险感知的测试时间扩展

ARTIS: Agentic Risk-Aware Test-Time Scaling via Iterative Simulation

Huawei Technologies Co., Ltd(华为技术有限公司);Harbin Institute of Technology, Shenzhen(哈尔滨工业大学(深圳));Shanghai Jiao Tong University(上海交通大学)

Xingshan Zeng, Lingzhi Wang, Weiwen Liu, Liangyou Li, Yasheng Wang, Lifeng Shang, Xin Jiang, Qun Liu

AI总结:

ARTIS通过迭代模拟提升代理在高风险环境中的可靠性和鲁棒性,采用风险感知的模拟器改进决策效果。

摘要:

当前的测试时间扩展(TTS)技术通过在推理时间分配额外计算来增强大语言模型(LLM)的性能,但它们在代理设置中仍然不足,因为动作会直接与外部环境互动,其影响可能是不可逆且成本高昂的。我们提出了ARTIS,即通过迭代模拟实现的代理风险感知测试时间扩展框架,该框架通过在真实世界执行之前通过模拟交互实现测试时间探索,从而将推理时间的计算扩展到提高动作层面的可靠性和鲁棒性,而不会产生环境风险。我们进一步表明,基于LLM的模拟器在捕捉罕见但高影响的故障模式方面存在困难,这显著限制了其在代理决策中的有效性。为了解决这一限制,我们引入了一个风险感知的工具模拟器,通过针对性的数据生成和重新平衡训练来强调故障诱因动作的保真度。在多轮和多步骤代理基准测试中的实验表明,迭代模拟显著提高了代理的可靠性,并且风险感知的模拟对于在不同模型和任务中一致实现这些收益至关重要。

英文摘要:

Current test-time scaling (TTS) techniques enhance large language model (LLM) performance by allocating additional computation at inference time, yet they remain insufficient for agentic settings, where actions directly interact with external environments and their effects can be irreversible and costly. We propose ARTIS, Agentic Risk-Aware Test-Time Scaling via Iterative Simulation, a framework that decouples exploration from commitment by enabling test-time exploration through simulated interactions prior to real-world execution. This design allows extending inference-time computation to improve action-level reliability and robustness without incurring environmental risk. We further show that naive LLM-based simulators struggle to capture rare but high-impact failure modes, substantially limiting their effectiveness for agentic decision making. To address this limitation, we introduce a risk-aware tool simulator that emphasizes fidelity on failure-inducing actions via targeted data generation and rebalanced training. Experiments on multi-turn and multi-step agentic benchmarks demonstrate that iterative simulation substantially improves agent reliability, and that risk-aware simulation is essential for consistently realizing these gains across models and tasks.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑