REPOT: Recoverable Program-of-Thought via Checkpoint Repair

Researchers introduced RePoT, a checkpoint-repair version of Program-of-Thought reasoning that verifies each Python step deterministically and repairs only the first invalid step.

On PuzzleZoo-775, RePoT beat vanilla PoT by about 3 to 11 percentage points across several models, reaching 96.9% vs. 86.3% on gpt-5.4-mini-medium.

The approach also reproduced gains on PlanBench Blocksworld and adds little overhead because it usually needs at most one extra model call.

The team also released Derail-550 to benchmark recovery methods fairly with the failure point fixed.