The article identifies a ‘no-recovery bottleneck’ in long-horizon reasoning for LLMs where extreme decomposition leads to irreversible errors on hard steps.

The authors propose LEAD (Lookahead-Enhanced Atomic Decomposition) which uses short-horizon future validation and aggregated overlapping rollouts to balance stability and error correction.

On the Checkers Jumping puzzle, LEAD enables the o4-mini model to solve complexities up to n=13, whereas extreme decomposition fails beyond n=11.