AdaPlanner: Adaptive Planning from Feedback with Language Models

Large language models (LLMs) have recently demonstrated the potential in acting as autonomous agents for sequential decision-making tasks. However, most existing methods either take actions greedily without planning or rely on static plans that are not adaptable to environmental feedback. Consequently, the sequential decision-making performance of LLM agents degenerates with problem complexity and plan horizons increase. We propose a closed-loop approach, AdaPlanner, which allows the LLM agent to refine its self-generated plan adaptively in response to environmental feedback. In AdaPlanner, the LLM agent adaptively refines its plan from feedback with both in-plan and out-of-plan refinement strategies. To mitigate hallucination, we develop a code-style LLM prompt structure that facilitates plan generation across a variety of tasks, environments, and agent capabilities. Furthermore, we propose a skill discovery mechanism that leverages successful plans as few-shot exemplars, enabling the agent to plan and refine with fewer task demonstrations. Our experiments in the ALFWorld and MiniWoB++ environments demonstrate that AdaPlanner outperforms state-of-the-art baselines by 3.73% and 4.11% while utilizing 2x and 600x fewer samples, respectively.

Paper

Similar papers

Peer review

Reviewer aJpk7/10 · confidence 3/52023-06-25

Summary

LLMs have shown success as autonomous agents that make and execute plans in sequential decision problems. Existing methods either make open-loop plans, limiting adaptability to the environment, or closed-loop plans. Existing closed-loop methods, apart from DEPS, keep the plan static but simply modify immediate actions according to environment feedback, leading to potentially sub-optimal policies. The authors introduce AdaPlanner, a closed-loop LLM planner that additionally allows for *plan* refinement during the episode. The success of their method not only relies on this, but additionally code-style prompts and a skill-discovery mechanism for few-shot exemplars. AdaPlanner outperforms existing works while relying on far fewer demonstration examples from similar tasks.

Strengths

- Empirically the authors show strong results with respect to sample efficiency and asymptotic performance. - Many ablations make it easy to understand which components of the model lead to overall success. - Conceptually simple approach.

Weaknesses

- In the evaluation section, the baselines are glossed over. This makes it hard to comprehend the distinction between their approach and the baselines. - I’d recommend adding some of the Appendix descriptions to the evaluation section, and potentially referencing Table 1 more often. - The authors use the term ‘hallucination’ a lot but do not define it. - The authors discuss in- and out-of- plan refiners a lot before providing intuitive examples for when either would be necessary. Could the authors provide more examples earlier on in the paper? - DEPS appears to be a relevant baseline. Could the authors include it or at least delve deeper into its limitations and why it is not appropriate? - It appears that the largest contributor to the success of AdaPlanner, over existing approaches, is code style prompts and skill prompts. Wouldn’t it be worthwhile to apply those modifications to existing approaches, like Reflextion (Fig 4), and contrast? - AdaPlanner prompts the LLM to correct any syntax errors. How important is this? Would be nice to include this ablation.

Questions

- Line 80, could you define the output of pi, in the same way that you did for the planner? - Line 81, shouldn’t it be P_t rather than P_{t - 1}? - Lines 114 - 144 I think you’ve repeated the sentence twice. - Line 216, what are the 6 task types? - Line 132, how is N chosen and what’s its effect on performance?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

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

3 good

Presentation

2 fair

Contribution

3 good

Limitations

AdaPlanner still requires demonstrations for learning. Would be worthwhile comparing with RL agents trained directly on the task, without any expert demonstrations.

Reviewer oy346/10 · confidence 4/52023-06-29

Summary

Briefly summarize the paper and its contributions. This is not the place to critique the paper; the authors should generally agree with a well-written summary. The paper proposes AdaPlanner, an LLM-based adaptive planner for text-based sequential decision-making tasks. The planner is adaptive in the sense that it can refine the generated plan/policy based on feedback. The contributions made in this paper include the following 1. interacting with the environment with LLM in the loop 2. a code-style prompt is engineered for LLMs to output a policy 3. refining the LLM policy for the current task based on feedback 4. prompt tuning for new tasks based on previous interaction (termed skill discovery) The proposed AdaPlanner is evaluated on two text-based sequential decision-making environments ALFWorld and MiniWoB++. Their experiments indicate that with feedback, LLMs can adapt the plan.

Strengths

* The paper is well written. * The paper focuses on extremely relevant and signifcant problems.

Weaknesses

* I find the paper lacks significant details. Please see the next section for the list of questions. * The paper employs sloppy mathematical notations. * The paper lacks the rigor of scientific evaluation. * Paper misses all references to LLM-based approaches for planning with PDDL. The one that I find most relevant for code generation is "Generalized Planning in PDDL Domains with Pretrained Large Language Models, Tom Silver, Soham Dan, Kavitha Srinivas, Joshua B. Tenenbaum, Leslie Pack Kaelbling, Michael Katz”

Questions

**Major** 1. How is the programmatic response from LLM converted to action responses? Did the conversion require manual intervention? For instance, Figure 2 has an indentation error which would result in a wrong plan. Were such indentation errors evaluated manually? Can authors provide a list of errors made by LLMs? 1. In line 167, what does an alignment between ‘anticipated plan’ and environment mean? How does the AdaPlanner observe the alignment? 1. Can authors provide details about the size of the task used in the prompt (for samples) vs the size of the task that was successfully solved by AdaPlanner? To establish the claim of sample efficiency, it is important to understand if the planner is able to efficiently plan for tasks that are significantly different from the prompts. 1. The X-axis in Figure 3 indicates `# Samples per task`. Is this the number of samples provided for each trajectory? Or sum? 1. What was the length of plans or length of trajectories generated by AdaPlanner vs other approaches? To claim the effectiveness of the AdaPlanner, it is important to compare the length of successful trajectories. 1. For skill discovery, how is the solution converted to the skill? How are skills represented? How large is the skill memory? Were the discovered skills included in the count of samples used for training as they are training samples for the next set of trajectories? 1. It is not clear how skills are filtered and what criteria are used for the evaluation and ranking of skills. 1. What is the connection between skill discovery and prompt tuning? 1. The success rate of "With SD" in Figure 4d looks significantly reduced from Figure 4a. Were different settings used for theses experiments? 1. At various places, the paper mentions "environment feedback". In my opinion, this is a misnomer. The feedback is not from the environment. The environment just provides the next observation, the feedback is generated by the agent itself. And the use of observation to refine a plan or next action is quite standard practice in RL. I would highly recommend dropping the term feedback from the title. 1. The use of term plan and policy is a little confusing. A plan is a sequence of actions. A policy is a mapping from states to actions. By this definition, the `solution()` function is as a policy. In preliminaries, the planning policy ($\rho$) is conditioned on a previous plan $P_t$. However, the appendix describes the refinement prompt using the assertion error (instead of `solution()`). Isn't the assertion error providing information about the policy (the `solution()` function)? So I am confused by the terminologies. Is the $\rho$ refined conditioned on the policy or the plan? The usage of these terms is also confusing in the Preliminary section. Request authors to precisely define the mathematical notations and highlight what they represent in the examples. **Minor** 12. In line 387, there are extra curly braces. 12. The notation $\rho$ is used in line 73 but introduced much later. 12. As the context $c_t$ is defined as a sequence of action and observations from time step $0$ to $t$, it is not clear what $c_{>t}$ means (in line 116). 12. Open-Loop system in Figure 1 should have an arrow going from env to planner with $o_1$. 12. Statement in Line 144 "To generate a plan .." looks like a repetition of Line 141 "To generate an initial plan..." 12. In line 116, if $h_t$ is obtained from $c_t$ then would it not be captured in $c_{>t}$? An example of $h_t$ would help better understand the proposed update. 12. In line 73, as $\rho$ is defined using $\Delta(A^{T})$. But the length $T$ is not fixed. 12. In line 73 $\rho$ is defined where a plan is conditioned only on observation and goal. However, later it is conditioned on the context, plan, and goal.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, 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

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

* The evaluations are restricted to text-based sequential decision-making problems and task where the inadmissible actions do not cause drastic changes in the environment. On the contrary, inadmissible actions are like no-ops. Further, the paper does not present analysis of plan length. Hence, the analysis is limited to zero risk environments. * The claim made in the abstract about skill discovery mechanism enabling agent to plan with fewer task demonstration is not substantiated in the evaluations. Evaluation in Fig. 4d only established improvement in success rate, not sample efficiency.

Reviewer GDYQ5/10 · confidence 4/52023-07-04

Summary

The paper presents AdaPlanner, a closed-loop planning method that uses a large language model (LLM) to solve tasks in text-based environments. AdaPlanner operates by decomposing a complex task into manageable sub-goals and predicting environmental feedback for each. During execution, it refines its actions based on the feedback received from the environment. AdaPlanner operates solely via prompting, eliminating the need for a dedicated training phase and reducing its computational cost. The paper demonstrates that AdaPlanner consistently outperforms existing baselines, achieving state-of-the-art performance in ALFWorld tasks and MiniWoB++ tasks.

Strengths

- AdaPlanner introduces a novel approach to task-solving in text-based environments using a large language model. It stands out for its closed-loop planning method and its ability to decompose tasks into manageable sub-goals. - The paper is well-written and clear. The authors have done a good job of explaining complex concepts and methodologies in an understandable manner. - The work presents a new way of leveraging large language models for task-solving in text-based environments. The results show that AdaPlanner can effectively leverage feedback to refine its plans and enhance its performance.

Weaknesses

- The part about skill discovery is not described very clearly, and I still cannot understand the details of the skill discovery module well. - The author compared the version without a code interface in the experiment, but it seems that they did not specifically show the prompt after removing the code interface. At the same time, as an ablation experiment, it is also necessary to analyze the effects of specific components in the code interface. - The phenomenon that GPT-3 performs better than GPT-3.5 is interesting, but it seems that the paper only compares GPT-3 and GPT-3.5 in Alfworld, without conducting the same experiments in MiniWoB++ to further support the conclusion. And the author's hypotheses about this phenomenon (the smaller scale of GPT3.5) lacks specific analysis or literature references to support it.

Questions

- In the experiment, what is the proportion of in-plan and out-of-plan occurrences? How will this proportion change over time? This should be a necessary indicator for understanding the two refiners. - On MiniWoB++, will there be better performance from GPT-3 than GPT-3.5? - Is there still a necessity for AdaPlanner in larger-scale LLMs, such as models like GPT4 with better self-refining capabilities?

Rating

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

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

3 good

Presentation

3 good

Contribution

2 fair

Limitations

- As mentioned above, this paper still needs more experiments and analysis to further validate the rationality of its methods, as well as the observed phenomena and corresponding hypotheses.

Reviewer 4ZaS6/10 · confidence 3/52023-07-07

Summary

This paper looks at explicit closed-loop systems with LLMs for adaptive planning utilizing environmental feedback. They showcase better planning performance on ALFWorld and MiniWOB++ environments over existing state-of-the-art works like ReAct and Reflexion.

Strengths

The paper is well written and the experiments are thorough. They present an interesting improvement over the current works like ReAct and Reflexion.

Weaknesses

1. The kind of tasks in these domains don’t seem to have interaction resolution where there are multiple conflicting causal links from the initial to the goal state which have to be resolved (including negative interactions between subgoals). This could also lead to the human demonstrations helping significantly with the It would be useful to analyze the performance of AdaPlanner specifically in such cases. 2. I think non-ergodic environments could clearly pose danger to such agents. It would be interesting to see how AdaPlanner can perform against ReAct or Reflexion in such environments.

Questions

1. Given that the LLM seems to verify the plan to determine its feasibility, what was its efficiency in those assessments? Are there any results pertaining to that? 2. Is there any classification of the tasks with respect to their hardness? 3. For how many of these tasks did the human expert demonstration solve the task?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

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

3 good

Presentation

3 good

Contribution

3 good

Limitations

The authors have addressed some of the limitations. I have provided some limitations in the weaknesses section.

Authorsrebuttal2023-08-12

**Please check the remaining part of our rebuttal to Reviewer oy34 as follows:** **[Q1.4]** We summarized the errors that have occurred through the evaluation in ALFWorld as follows: | Type | Description | | ----------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | Indentation error | Occasionally, the code may contain indentation errors. This error can be addressed by the Code Check. | | Reference error | The ```start_from``` may be misconfigured and the last breakpoint are not properly loaded. This error can be mitigated by out-of-plan refinement. | | Incomplete code | Due to the token limit, generated code might sometimes be truncated. We can adopt newer models that support extended context lengths, such as gpt-3.5-turbo-16k. | **[Q6.1/6.2]** We use each solution as a demonstration in the prompt to solve similar tasks. If including the solution boosts the success rate, we will keep the solution in the skill memory and add it to the prompt for future task-solving. **[Q6.3]** The skill memory can potentially be large, depending on the task complexity. The automatic adjustment of the skill memory's size is an important feature that we will investigate in the future. **[Q6.4]** AdaPlanner is purely prompting-based and there is no training process involved. We add the discovered skills to the prompts. The skills are actively discovered by the agent, thus not included in the count of expert demonstrations. **[Q7]** As illustrated in line 209-213, we assess the effect of using each discovered skill as the demonstration in the prompt to solve similar tasks. If adding a candidate solution boosts the success rate on these tasks, it's added to the memory of discovered skills; if not, it's discarded. **[Q8]** The successful skills discovered by AdaPlanner will be added to the prompt. It could be understood as a method for prompt tuning. However, we emphasize that this procedure is automatically completed by LLM itself during the planning. **[Q9]** Yes. The detailed settings for these two figures were mentioned in line 227-229, 442-445. In Figure 4a, we adopt the setting of prompted samples as in Table 4. In Figure 4d, we only provide one sample of the simplest task (pick) and use skill discovery to explore skills for the rest five tasks. The difference in the success rate of “with SD” in Figure 4d originates from this setting difference. **[Q10]** We use “environment feedback” to indicate any outcome provided by the environment, including the observations. **[Q11]** The assertion error provides information about the plan rather than the policy. For example, in ALFWorld, the error message reports which action within the plan has been executed and the error occurs (e.g., ```Error in [Step X], …```). This information corresponds to $P_{t-1}$ as in the definition of $\rho(P_t|g, c_t, P_{t-1})$ in line 119. **[Typos] Q12, 13, 14, 15, 16, 19** Thank you for the comments. We will revise these typos and notations in the updated version of our paper. **[Q17]** The $h_t$ would be included in $c_{>t}$. For example, the agent identifies the target object ```book 1``` from the environment observation ```On drawer 2, you see a book 2, and a keychain 1.``` Here $h_t$ is the identifier information which is then used for future actions. **[Q18]** We fix $T$ as the step limit of the environments. For example, in ALFWorld, this number is set to 50.

Reviewer oy342023-08-16

Thank you for your thorough response. I still find that literature that uses LLM with PDDL is relevant to this work and should be cited. This is a fast-moving field and hence keeping track of all work & its chronology is clearly impossible, but **not** citing **any** of the work done in that field is un-scholarly. I appreciate the addition of the empirical results that study sample efficiency. I believe the response to Limitation 2 (and other additional results) addresses Weakness 3. I will update my score accordingly.

Authorsrebuttal2023-08-19

Thank you for taking the time to review our rebuttal and offering insightful feedback. We will update our paper with the additional results and discussions and add the literature review on methods leveraging LLM with PDDL.

Reviewer aJpk2023-08-14

Response to Rebuttal

I thank the authors for their detailed rebuttal. I appreciate the additional ablation and the comparisons made with methods that do not leverage demonstrations. In line with my review, I believe this is a good paper and will keep my score as is.

Authorsrebuttal2023-08-19

Thanks for reading our rebuttal and providing valuable feedback. We will incorporate these additional results and discussions into the updated version of our paper.

Reviewer GDYQ2023-08-14

Thanks for the author's response. Most of my concerns have been addressed, for now I will maintain my current score and continue to pay attention to other reviews and discussions.

Authorsrebuttal2023-08-19

Thank you for offering feedback on our rebuttal. We will update our paper and include the additional results and discussions.

Reviewer 4ZaS2023-08-21

Thank you for the detailed response. Most of my concerns have been addressed.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC