Reproduce “Enhancing LLM Planning Capabilities through Intrinsic Self-Critique” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Bernd Bohnet, Pierre-Alexandre Kamienny, Hanie Sedghi, Dilan Gorur, Pranjal Awasthi, Aaron Parisi, Kevin Swersky, Rosanne Liu and 2 more
The authors present an LLM-based method that iteratively critiques plans against task rules and reuses failed plans and critiques as context. In Blocksworld 3–5, Gemini 1.5 Pro accuracy rises from 49.8±4.0% without critique to 89.3±2.5% with an eight-shot critic and five-vote self-consistency. VAL scores final plans but does not guide revision; this formal-task result does not establish comparable real-world performance.
Key insights
- Intrinsic self-critique, when implemented with a carefully structured prompt, is highly effective for enhancing LLM planning capabilities, directly challenging prior negative conclusions about its viability.
- Including a comprehensive domain definition and explicit, step-by-step verification instructions in the self-critique prompt is crucial for the LLM's ability to accurately evaluate its own plans.
- Combining iterative self-critique with Self-Consistency significantly boosts the reliability of the LLM's self-evaluation, allowing it to reduce false positives and approach oracle-level performance in identifying correct plans.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.