REPOT: Recoverable Program-of-Thought via Checkpoint Repair

Parsa Mazaheri

arXiv:2605.30052 · 2026-05-29 공개 · arXiv · PDF

llm-calls checkpoint-repair puzzlezoo-775 planbench-blocksworld derail-550 adaptive-repot verified-prefix recovery-benchmark

Abstract

One-shot Program-of-Thought (PoT) emits a Python program that prints a primitive-action plan; a single invalid action silently invalidates the trajectory. We introduce RePoT (Recoverable PoT): a deterministic verified replay that walks the plan through the environment to its first invalid transition, then one LLM call that resumes from the verified prefix. RePoT costs at most one extra LLM call on the ~14% of problems where PoT fails. RePoT beats PoT by +3 to +11pp across four closed-model configurations on PuzzleZoo-775 and peaks at 96.9% vs 86.3% on gpt-5.4-mini-medium; against the matched-budget PoT-retry baseline, RePoT wins decisively on Gemini (+3.8pp, 95% CI [+2.2,+5.4]), is within sampling noise on GPT-medium and Claude, and loses on GPT-mini -- a capability-scaling pattern we begin to address with Adaptive RePoT, a rule-based dispatcher that routes between suffix repair and a fresh PoT retry based on verified-prefix length (preliminary). We replicate on PlanBench Blocksworld (+1.1 to +11.4pp) and on four open-weights models (+3.3 to +20.0pp on three of four). On Derail-550, our controlled recovery benchmark, every condition with access to checkpoint information clears >=30% on GPT-medium and >=70% on Gemini, vs <=3.1% for error-only feedback -- showing that checkpoint information, not the specific verified-prefix tail, is the load-bearing recovery signal.

한국어 요약

📋 한 줄 요약

**[Program-of-Thought / Checkpoint Repair]** RePoT가 PoT plan을 환경에서 verified replay로 first invalid transition까지 walk 후 single LLM call로 suffix 재개 — PuzzleZoo-775에서 PoT 대비 +3~+11pp, gpt-5.4-mini-medium 96.9% vs 86.3%.

🎯 핵심 기여도

💡 핵심 아이디어

PoT의 single-error fragility는 verified replay로 검증된 prefix를 추출하고 그 지점부터 single LLM call로 suffix를 재생성하는 deterministic recovery로 해결 가능하며, error 단서가 아닌 checkpoint 정보 자체가 load-bearing 회복 신호다.

🔬 기술적 접근법

📊 주요 결과

💭 의의 및 한계

**의의**: PoT의 단일 실패점을 minimal cost로 회복하는 일반 메커니즘 제시, error-message가 아닌 verified state가 회복의 핵심임을 정량 입증, capability-scaling 한계까지 분석한 정직성. **한계**: 환경 verifiable replay 가능성 전제(closed-form 환경 한정), GPT-mini 등 작은 모델에서는 fresh retry가 더 나은 capability 패턴, Adaptive 라우팅의 dispatcher 룰 의존.

🚀 실용적 활용