Long-Horizon Task
A core concept for evaluating the capabilities of AI agents, referring to complex tasks requiring multiple consecutive steps in which each step depends on the result of the previous one and errors are likely along the way. The difficulty lies in error accumulation: even with a high per-step success rate, overall success rate declines rapidly as the number of steps increases.
Timeline (last 90 days)
The red team acknowledges that the reliability of current models on long-chain tasks is still limited.
A long-chain task is a complex task requiring multiple consecutive steps, where each step depends on the result of the previous one and errors are very likely along the way; its difficulty lies in error accumulation.
All Facts (3)
The red team acknowledges that the reliability of current models on long-chain tasks is still limited.
50%UnverifiedA long-chain task is a complex task requiring multiple consecutive steps, where each step depends on the result of the previous one and errors are very likely along the way; its difficulty lies in error accumulation.
50%Unverified级联失败是指单次工具调用错误会污染后续上下文，导致模型在错误的中间状态上继续推理，最终使整个任务偏轨
50%