This paper studies why large language models keep generating long, expensive reasoning chains on problems beyond their real capability, producing plausible-looking but wrong answers instead of admitting failure. The authors trace the dominant failure mode to specious reasoning — output that looks superficially valid but contains subtle errors that grow more common as task difficulty increases — and propose CaRL (Capability-aligned Reinforcement Learning), which reshapes rewards to favor refusal over futile reasoning and converts past failures into refusal training data. Experiments show the method substantially reduces flawed reasoning on tasks beyond the model’s capability while preserving performance on tasks it can actually solve.
