arXiv cs.CL· Minghan Wang, Boyuan Wang, Jinhang Zuo, Yuxin Tao, Fang Kong·· 9 天前AI 评分30
语言模型能否选择性依赖外部指导？Box²-Bench 基准发布
Thinking Outside the Box: Can Language Models Rely on External Guidance Selectively?
AI 导读
研究者提出 Box²-Bench，在固定模型与任务的前提下改变工作流可靠性，以隔离模型对指导的依赖调节能力，并将这种"从有用指导中获益、同时覆盖不可靠指导"的能力称为 thinking outside the box。
正文 · 原文
Abstract:Agent harnesses often improve language models with human-designed workflows, but as models grow more capable, unreliable guidance can increasingly constrain their execution. We call the ability to benefit from useful guidance while overriding unreliable guidance thinking outside the box. We introduce Box$^2$-Bench, which holds the model and task fixed while varying workflow reliability to isolate how models regulate their reliance on guidance. On Box$^2$-Bench, frontier models often benefit from reliable guidance but remain vulnerable when it is misleading or becomes unreliable. To test whether this capability can be learned, we train two open-weight models using bad workflows, reserving good workflows for evaluation. We explore two complementary training strategies: counterfactual supervised fine-tuning improves robustness, while outcome-based reinforcement learning can shift the balance toward greater use of helpful workflows. We further find that this behavior extends beyond workflows to other forms of external information, improving peer correction and robustness to corrupted memory. Together, our results identify selective reliance on fallible external information as a dimension of agent reliability not captured by task performance alone.
| Subjects: | Computation and Language (cs.CL) |
| Cite as: | arXiv:2609.39578 [cs.CL] |
| (or arXiv:2609.39578v2 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2609.39578 arXiv-issued DOI via DataCite |
Submission history
From: Boyuan Wang [view email]
[v1]
Wed, 30 Sep 2026 12:10:24 UTC (16,005 KB)
[v2]
Thu, 8 Oct 2026 04:08:50 UTC (16,005 KB)
来源：arXiv cs.CL · arxiv.org