On the Planning Abilities of Large Language Models : A Critical Investigation

Intrigued by the claims of emergent reasoning capabilities in LLMs trained on general web corpora, in this paper, we set out to investigate their planning capabilities. We aim to evaluate (1) the effectiveness of LLMs in generating plans autonomously in commonsense planning tasks and (2) the potential of LLMs in LLM-Modulo settings where they act as a source of heuristic guidance for external planners and verifiers. We conduct a systematic study by generating a suite of instances on domains similar to the ones employed in the International Planning Competition and evaluate LLMs in two distinct modes: autonomous and heuristic. Our findings reveal that LLMs' ability to generate executable plans autonomously is rather limited, with the best model (GPT-4) having an average success rate of ~12% across the domains. However, the results in the LLM-Modulo setting show more promise. In the LLM-Modulo setting, we demonstrate that LLM-generated plans can improve the search process for underlying sound planners and additionally show that external verifiers can help provide feedback on the generated plans and back-prompt the LLM for better plan generation.

Paper

References (40)

Scroll for more · 28 remaining

Similar papers

Peer review

Reviewer Y3FL5/10 · confidence 3/52023-07-05

Summary

This paper conducts a systematic study by generating a suite of instances on domains similar to the ones employed in the International Planning Competition and evaluate LLMs in two distinct modes: autonomous and heuristic. The experiments show that LLMs' ability to generate executable plans autonomously is rather limited, while the results in the heuristic mode show more promise.

Strengths

1. The paper is well-written and easy to understand. 2. The paper provides a detailed investigation of GPT's planning abilities in different domains and presents some interesting findings.

Weaknesses

1. Some of the conclusions presented in the paper, such as re-planning makes better performance, have already been widely applied in robotics task and motion planning applications, which is not considered novel. The community has developed a range of interesting algorithms to enhance the planning capabilities of LLM, including re-plan and generating feasible plans [1,2,3,4,5]. While the authors ignores these efforts. 2. The author's investigation of GPT's planning abilities in certain domains overlooks the fact that GPT's greatest strength lies in its zero-shot or few-shot capabilities across different domains, without the need for pre-defined action spaces. Additionally, the metrics used to evaluate GPT are not fair to the model itself. At least human evaluation should be introduced for a more comprehensive assessment. More detailed experiments and metrics can be found in [zero-shot]. 3. For a survey paper, it is not sufficient to only consider OpenAI's GPT-level models. Open-source models like LLaMA and Vicuna should also be included in the analysis to better understand if there are fundamental differences in planning capabilities across different levels of language models. 4. More demonstrations in the prompts seem to effectively improve planning performance. The author only conducted zero-shot and one-shot experiments, which is insufficient. A ablation experiment to explore the importance of the "how much-shot" factor would be valuable to the community. 5. Some relevant papers are not cited, which are listed in the references. References: [1] Inner Monologue: Embodied reasoning through planning with language models [2] Describe, Explain, Plan and Select: Interactive Planning with Large Language Models Enables Open-World Multi-Task Agents [3] ReAct: Synergizing Reasoning and Acting in Language Models [4] Reflexion: Language Agents with Verbal Reinforcement Learning [5] Text2Motion: From Natural Language Instructions to Feasible Plans

Questions

Answer questions in weakness section.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

N/A

Reviewer GZ9T8/10 · confidence 4/52023-07-07

Summary

This paper provides a systematic evaluation of the Planning abilities of a class of LLM (GPT series until the latest GPT-4), using standardized planning problems such as those provided in symbolic planning competitions. It analyzes LLM as autonomous planners, but also as heuristic planners providing suggestions to sound planners (heuristic mode). The experiments use a sophisticated prompting mechanism allowing both NL prompting and PDDL prompting. In the heuristic mode, LLM plans can be repaired via a local search planner (LPG), or use a verifier to send feedback to the LLM ('backprompting'). While LLMs score poorly as autonomous planners, they can achieve high scores (70-80%) in heuristic mode on standard benchmarks. Post-rebuttal comments: the authors have fully answered my questions, in particular providing additional results for another LLM (BLOOM, which appears a very reasonable choice, due to how its ecosystem differs from that of the GPT family). Their commitment to release some of the evaluation resources can only enhance the contribution of this work, which is reflected in my increasing the 'contribution' score in this review. Having also considered other reviewers' comments and the author responses, I remain very positive about this paper and maintain my original score.

Strengths

The paper introduces a comprehensive evaluation method for the Planning abilities of LLM, harnessing the full methodology of traditional Planning in terms of benchmarks and reasoners. The choice of the latter is particularly appropriate to the experiments at hand, since it includes both a local planner, well-suited to Plan repair, and a validator that can send feedback by identifying gaps or flaws in the proposed plan under the 'heuristic mode'. It is fairly impressive to have automated a process previously taking place in an interactive form with a human in the loop, which was vulnerable to the Clever Hans effect, also observed in other forms of LLM reasoning [1]. The prompting mechanism is particularly sophisticated without being over-engineered, as their is a clear rationale for supporting each option and a rather elegant design starting with PDDL domains and branching out to generate NL or PDDL prompts. In line with some claims that LLM reasoning tend to reproduce human reasoning to some extent, the choice of planning domains known to be solvable by humans is of high interest, although some of the user/human experiments have been moved back to supplementary material. The paper is highly readable and quite systematic, and has all the elements to become a reference paper on the topic, not least for the results produced and the contrast between autonomous and heuristic modes, the latter avoiding pitfalls of interactivity or CoT limitations. [1] Shapira, N., Levy, M., Alavi, S.H., Zhou, X., Choi, Y., Goldberg, Y., Sap, M. and Shwartz, V., 2023. Clever hans or neural theory of mind? stress testing social reasoning in large language models. arXiv preprint arXiv:2305.14763.

Weaknesses

The paper is technically sound with very few weaknesses. Perhaps one issue is the limited number of Planning test domains used in the experiments, especially compared to [Silver et al., 2022] (ref [24] in the paper). This might be in relation to the need to explore human solutions to Planning problems as described in the supplementary material, still it could be worth justifying explicitly. Another potential issue would be concentrating on the GPT family, as LLM may vary in their real-world knowledge depending on their training base. The paper has similarities with the following preprint: https://arxiv.org/abs/2302.06706 This is not a major issue, either in terms of novelty or in terms of anonymity, since the submission has substantial new material and there is no direct link to the authors of the preprint. Regardless of whether the preprint is from the same authors (or a subset), it would still be appropriate to reference it in the final version of the paper.

Questions

Can any of the observations be related to the autoregressive nature of the LLM explored? What variability in performance would you expect across LLMs (other than GPT versions)?

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

3 good

Contribution

4 excellent

Limitations

The conclusion section leaves little space to discuss limitations of the approach. Since CoT prompting has also been explored, it could have been interesting to discuss the proposed coupling of LLM to reasoners via PDDL exchange proposed as part of "Faithful CoT" [1]. The paper rightly identifies the potential role of the natural language semantics of predicates or operators' names in LLM's planning abilities, for which it designs various methods of obfuscation. However, further discussions would be interesting for this phenomenon reported in [2] ("semantics of the English terms used in the PDDL problems"), such semantics of PDDL contents having also been proposed as a mechanism for planning model extension and planning repair [3]. [1] Lyu, Q., Havaldar, S., Stein, A., Zhang, L., Rao, D., Wong, E., Apidianaki, M. and Callison-Burch, C., 2023. Faithful chain-of-thought reasoning. arXiv preprint arXiv:2301.13379. [2] Silver, T., Hariprasad, V., Shuttleworth, R.S., Kumar, N., Lozano-Pérez, T. and Kaelbling, L.P., 2022, November. PDDL planning with pretrained large language models. In NeurIPS 2022 Foundation Models for Decision Making Workshop. - ref [24] of the paper [3] Porteous, J., Ferreira, J.F., Lindsay, A. and Cavazza, M., 2021. Automated narrative planning model extension. Autonomous Agents and Multi-Agent Systems, 35(2), p.19.

Reviewer JKYr8/10 · confidence 4/52023-07-07

Summary

This work evaluates the planning abilities of LLMs in two distinct settings: (1) As generators of final plans, with or without feedback from a validator, and (2) as generators of seed plans which are then corrected by a standard planner. The evaluations are performed on two commonsense domains for which humans tend to produce high-quality plans: Blocksworld and Logistics. Four LLMs are tested, including GPT-4, which generates more correct plans than the other models. Still, GPT-4 is found to fail on most of the problems, even with the benefit of one-shot prompting and CoT reasoning. But feedback from a validator (VAL) dramatically boosts the solution rate to 82% in BW and 70% in logistics, after just 3-4 feedback loops on average. In the other setting, the standard planner (LPG) produces correct plans in significantly fewer steps when starting with seed plans generated by GPT-4.

Strengths

This is excellent work, carefully detailed, and clearly presented. It avoids the Clever Hans effect that often arises when humans are involved in evaluations. The different evaluation settings are very well chosen, and the results provide valuable guidance as industries work to understand how these LLMs can best be leveraged. GPT-4 is thoroughly evaluated on all benchmarks. Inclusion of three other LLMs on many of the evaluations provides additional insight.

Weaknesses

Covering more domains beyond these two would be a useful contribution. But two are sufficient to support the conclusions drawn, and they light the way for others to run similar evaluations on additional planning domains.

Questions

Line 97 says that “our approach of specifying the domain as part of the prompt ensures that the generated plans only use the actions in the domain specification.” How does domain specification provide this guarantee? Can't the LLM still hallucinate nonsense? Line 191 says: “We set the temperature for all models to be 1, thereby making them deterministic.” Is this a typo? The appendix correctly identifies zero as the temperature that produces deterministic behavior.

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

No concerns.

Reviewer NTC18/10 · confidence 5/52023-07-10

Summary

The paper investigates the (lack of) capabilities of pretrained LLM for solving classical well-known planning benchmarks. They study both the case of fully autonomous LLM without any external feedback and the case of using external tools, for validation feedback or as a seed for improving an external planner. As no fine-tuning is done, the work involves variations of prompts for the task considered. Fig 2 is an excellent summary of the studied tasks including zero and one-shot tasks with the problem in NL or in the original PDDL. GPT-4 is reported as the most powerful LLM, but it is still not satisfactory. Moreover, it's shown to be sensitive to the name description, performing better when using the standard names for well-known benchmarks.

Strengths

- Some LLMs such as GPT-4 are being used to obtain plans, so it's important to investigate their capabilities. - Classical planning benchmarks are well-understood so offer a solid ground for evaluation. - Consider explicitly the case of autonomous mode vs external but automatic feedback. - Chain of thought is investigated, answering a question that people familiar with LLMs might have.

Weaknesses

- LLMs are trained in language and human-written code. The evaluation with Randomized Disguising might be less meaningful. - However, Randomized Disguising is a small part of the work. - An alternative not explored in the paper is to add human-readable description to the domains. Even though that requires human intervention, it's reasonable to assume the ones providing the domain can also provide a description. - The domain-specific translator is not discussed.

Questions

- Did you try COT with PDDL prompts? - Are there any indications of how LLM would perform in problems with shallow plans? If the number of objects or actions is big enough, that might be challenging for classical planners. - For the interactive scenario, does it make sense to keep the temperature at 0? Perhaps randomization might help the LLM to recover to deviate from earlier commitments. - Did you attempt relaxations with other planning problems? Perhaps it's not "natural" in blocks world, but there are other problems where the plans are equivalent to their relaxed versions.

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

4 excellent

Contribution

4 excellent

Limitations

A potential limitation of this work is that it might inform about the relevant problems that required planning from LLMs. That's not the scope of the work, so that argument should be left aside. Instead, this work is a systematic investigation of well-known planning benchmarks. Those problems might be close to the distribution of the LLMs. I miss a discussion on the complexity of the planning tasks per se. There are some simple algorithms for solving block world problems. While logistics can be a complex problem, in the PDDL benchmark solving a single problem is bounded by the complexity of moving one package, a simple problem that as cities are fully connected and it takes one airplane trip for a package to the right city. The relaxed plans discussed in the paper might be related to my question about shallow plans. Other comments: - Fig 2 summarizes well the approaches studied, but that's sometimes not mentioned in the table captions. For instance, Table 1 and 2 should mention that they consider the automated approach. - What's I-GPT-3 in Table 3?

Reviewer NTC12023-08-14

thank you

Thank you for your responses. I agree that Randomized Disguising is not the same. That'd be a different experiment.

Reviewer GZ9T2023-08-16

Thanks for your detailed response, which answered my questions. It's great to have provided additional data, in particular on BLOOM, which differs sufficiently from GPT to broaden the argument. At some point, it might be interesting to investigate BLOOM's worst performance, but that is beyond the scope of this paper.

Reviewer Y3FL2023-08-18

Thank you for the author's reply. I hope the author can add the experimental results on other LLMs, including open-source models, to the final version.

Reviewer JKYr2023-08-18

Response to rebuttal

Thank you for the clarifications!

Program Chairsdecision2023-09-21

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC