Knowing When to Quit: Diagnosing and Training LLMs to Abort Futile Reasoning
The paper argues that LLMs should learn to refuse tasks beyond their capability instead of producing polished but empty reasoning.
The authors describe “futile reasoning” as costly, plausible-looking chains of thought that hide subtle errors on too-hard tasks. They say the main failure mode is specious reasoning, which worsens as task difficulty rises. Their proposed CaRL method uses reward shaping and hindsight refusal examples to align model behavior with capability limits. Experiments are reported to reduce futile reasoning while preserving performance across task difficulties. HF Daily Papers' note
The authors describe “futile reasoning” as costly, plausible-looking chains of thought that hide subtle errors on too-hard tasks. They say the main failure mode is specious reasoning, which worsens as task difficulty rises. Their proposed CaRL method uses reward shaping and hindsight refusal examples to align model behavior with capability limits. Experiments are reported to reduce futile reasoning while preserving performance across task difficulties. HF Daily Papers' note
score 5