{"id":2,"url":"https://pm.philipcastiglione.com/papers/2.json","title":"Position: LLMs Can’t Plan, But Can Help Planning in LLM-Modulo Frameworks","read":false,"authors":"Subbarao Kambhampati, Karthik Valmeekam, Lin Guan, Mudit Verma, Kaya Stechly, Siddhant Bhambri, Lucas Saldyt, Anil Murthy","year":2024,"auto_summary":"The paper \"Position: LLMs Can’t Plan, But Can Help Planning in LLM-Modulo Frameworks\" by Subbarao Kambhampati et al. argues that large language models (LLMs), such as GPT-4, cannot independently perform planning or self-verification tasks, which involve reasoning capabilities. The authors discuss the reasons behind common misconceptions in the literature regarding LLMs' abilities in these areas. They propose that LLMs should be seen as universal approximate knowledge sources that can significantly contribute to planning and reasoning tasks when integrated into a framework they term \"LLM-Modulo Frameworks.\"\n\nThe paper critiques the notion that LLMs possess inherent planning and reasoning capabilities, highlighting that LLMs are better understood as pseudo System 1 entities (using Kahneman's dual-system theory) that excel in tasks like linguistic completion but lack the structured reasoning required for planning (System 2 tasks). The authors emphasize that the excitement around LLMs' anecdotal performance on reasoning tasks is often misplaced, as empirical studies reveal their limitations in generating executable plans and verifying them.\n\nThe proposed LLM-Modulo Framework combines the strengths of LLMs with external model-based verifiers in a bi-directional interaction regime. This framework allows LLMs to generate candidate plans and approximate domain models, which are then critiqued and refined by external verifiers or humans, ensuring formal correctness where possible. The framework aims to leverage LLMs' capabilities in generating ideas and converting formats, while external critics handle the verification of correctness, thus providing a more robust approach to planning that extends beyond simple format translation.\n\nThe paper provides evidence from various studies showing that LLMs struggle with planning tasks, particularly when used in autonomous modes. For instance, experiments reveal that LLMs generate correct plans in only a small fraction of cases, and their performance does not significantly improve with iterative prompting or fine-tuning. Additionally, LLMs are found to be ineffective at self-verification, as they cannot reliably critique or improve their own outputs without external validation.\n\nThe authors argue that the LLM-Modulo Framework offers a more flexible and expressive approach to planning by allowing LLMs to contribute approximate knowledge and candidate solutions, while external verifiers ensure soundness. This approach avoids the expressiveness and search-complexity limitations of traditional symbolic planners, making it more applicable to real-world planning problems where complete domain models and problem specifications are not always available.\n\nIn conclusion, the paper positions LLMs as valuable tools for generating approximate knowledge and candidate solutions in planning tasks, but emphasizes the need for external verification to ensure correctness. The LLM-Modulo Framework is presented as a principled method for integrating LLMs into planning processes, leveraging their strengths while mitigating their limitations.","notes":{"id":2,"name":"notes","body":null,"record_type":"Paper","record_id":2,"created_at":"2024-12-10T03:49:32.591Z","updated_at":"2024-12-10T03:49:32.591Z"},"created_at":"2024-12-10T03:49:23.094Z","updated_at":"2024-12-10T03:49:43.715Z"}