Thinking Outside the Box: Can Language Models Rely on External Guidance Selectively?
Frontier models follow unreliable external guidance; training improves selective reliance, identifying it as a key agent reliability dimension.
Published Sep 30, 2026▲ 67 on Hugging FacearXiv ↗
Only vote on papers you've read. Sign in with GitHub to vote.
Box²-Bench cleanly isolates selective reliance as a core agent failure mode and shows it can be trained, though the benchmark never defines unreliable guidance, omits failure rates, and leaves unclear whether models override bad guidance by…
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.