Unverified50% confidenceFactExact time
Large language models are prone to goal drift during multi-step long-duration tasks, gradually deviating from the original core objective
1
Sources
50%
Confidence
Long-term
Relevance
9/22/2026
First Seen
Sources
Related Entities
Related Claims
Unverified「分步拆解」策略（将大任务拆成多个明确子任务逐步执行）比一次性发出模糊宏观指令效果更好76% similarVerifiedIn multi-step automation scenarios, single-step errors caused by model hallucinations are amplified exponentially along the task chain.73% similarUnverifiedA single execution-type agent exhibits three typical problems during long-running tasks: goal drift, shortcut-taking, and mid-task cessation72% similarUnverifiedPaths explored by researchers to mitigate goal drift include periodic goal restatement, hierarchical task planning, and the introduction of explicit task state tracking modules in Agent frameworks71% similarUnverified长任务中的漂移可拆解为目标漂移、状态漂移、策略漂移三种近似正交的失败模式71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/940413API
curl https://kongchang.com/api/v1/knowledge/claims/940413MCP
get_claim(id=940413)