Research Identifies Task Insensitivity as Key LLM Agent Failure

A new study pinpoints "task insensitivity" as a critical flaw hindering the out-of-distribution (OOD) generalization of long-horizon language agents. This phenomenon occurs when models, despite being presented with semantically distinct tasks, continue to apply patterns learned during training, often outputting actions aligned with the original task rather than the current instruction. The researchers observed that models might produce the same action even when a task description in a trained prompt is replaced with a similar but different one. This behavior is linked to an attention drift during training, where models increasingly focus on local observations instead of task-specific tokens, indicating an optimization bias towards shortcuts. To counteract this issue, the paper introduces Task-Perturbed NLL Optimization, a lightweight contrastive regularizer. This method explicitly encourages the agent's actions to be dependent on the task instruction. Extensive evaluations demonstrate that this intervention significantly improves task sensitivity and OOD generalization, while also maintaining more stable attention on the task tokens.