Language agents select actions based on task descriptions, environmental observations, and interaction histories. We investigate whether task requirements reliably constrain these decisions when the surrounding context already suggests how to proceed. Across five models and three interactive environments, we find that models, including GPT-5.6-Sol, can continue to act when task descriptions are replaced with the meaningless placeholder xxxx, even when prompted to ask for clarification. Action generation is more frequent in states that suggest a concrete next operation, although these cues do not establish whether that operation serves the user's goal. These findings suggest a possible training-based explanation: contextual cues can make demonstrated responses predictable, allowing models to fit them with limited reliance on task requirements. Tracking seven models during fine-tuning, we observe an overall decrease in attention to task-description tokens and an increase in attention to current observations. To strengthen task dependence, we introduce Task-Perturbed Negative Log-Likelihood (TPNLL), which contrasts the prediction costs of the same response under different task descriptions while holding the non-task context fixed. TPNLL increases attention to task-description tokens and improves complete-episode success in out-of-distribution settings.

翻译：暂无翻译