Conceptreasoning loop / 行动缺失 / action starvation
Reasoning Loop
A characteristic failure mode of reasoning large language models in agentic settings, in which the model continuously generates textual descriptions of actions it intends to perform without ever triggering actual tool calls or action instructions, causing task execution to stall indefinitely.
Timeline (last 90 days)
Sep 5
The reasoning loop contains three phases—Thought, Action, and Observation—and each iteration requires feeding the complete historical context back into the model, causing token consumption to grow in a stepwise fashion.
Unverified50%