Policy Improvement using Language Feedback Models

We introduce Language Feedback Models (LFMs) that identify desirable behaviour - actions that help achieve tasks specified in the instruction - for imitation learning in instruction following. To train LFMs, we obtain feedback from Large Language Models (LLMs) on visual trajectories verbalized to language descriptions. First, by using LFMs to identify desirable behaviour to imitate, we improve in task-completion rate over strong behavioural cloning baselines on three distinct language grounding environments (Touchdown, ScienceWorld, and ALFWorld). Second, LFMs outperform using LLMs as experts to directly predict actions, when controlling for the number of LLM output tokens. Third, LFMs generalize to unseen environments, improving task-completion rate by 3.5-12.0% through one round of adaptation. Finally, LFM can be modified to provide human-interpretable feedback without performance loss, allowing human verification of desirable behaviour for imitation learning.

Paper

Similar papers

Peer review

Reviewer B45L7/10 · confidence 3/52024-07-05

Summary

The authors introduce Language Feedback Models (LFM), a method filtering out desirable transitions (i.e. transitions considered as helping to solve a task) collected by an agent in an environment to then improve this agent by doing Imitation Learning on these transitions. The LFM method has three phases: 1) First, a fairly large number of transitions are collected by the initial policy, and GPT4 is used to determine which transitions are desirable. Then, 2) given this dataset of desirable transitions, a smaller LLM is trained to reproduce GPT4 in saying these transitions are desirable. Finally, 3) the policy is used to collect new transitions, which are then filtered out by the smaller LLM, taking only the ones considered as desirable, and the policy maximizes the likelihood of the actions chosen in these transitions. Experiments in two TextWorlds and one visual navigation environment are performed, showing that LFM successfully improves the initial policies (learned with Behavioral Cloning from expert demonstrations) and leads to better results than using GPT4 to directly propose the best action to perform and imitate it. Additionally, experiments show that the learned feedback model can be used to improve the policy in unseen test tasks which further increases the policy's performance even though the feedback model has not been trained on the test tasks. The authors also show that one can go further and not only train the smaller LLM to select desirable transitions but also to explain why these transitions are desirable (also by imitating GPT4), leading to better explainability.

Strengths

The method relies on distilling the feedback ability from GPT4 to a much smaller LLM (Flan-T5 770M) to reduce the cost of asking for feedback at every step. The authors show that they successfully transferred this ability and that the obtained feedback model generalizes to unseen tasks. The experiments also show that their method is not only less compute-effective but also much more efficient than directly imitating GPT4 used as the policy. The authors also show the method can be trained to provide explanations of the feedback given, leading to better explainability of the method.

Weaknesses

One of the main weaknesses I see is the lack of baselines to properly assess the efficiency of the method. The LFM method acts like a rejection sampling algorithm where only transitions considered desirable by the feedback model are kept to fine-tune the policy. Similarly, If all environments provide a reward if would have been interesting to have a baseline relying on this information for the rejection sampling. If the reward is sparse, one could wait for the end of the episode and only keep desirable trajectories (e.g. whose episodic reward is above some threshold or if the goal has been reached). The authors also say "In this work, we consider long-horizon settings with only sparse and delayed task-completion rewards. Consequently, we focus on imitation learning from demonstrations as opposed to reinforcement learning from rewards." It could also be interesting to see how RL performs with the same amount of collected data. Finally, it seems the authors use a dataset of expert demonstrations to first train the initial policy with BC before applying their method and the compared baselines. However, they do not provide any insights on how important this preliminary phase is. Indeed, on key advantage of LFM is that it relies on BC while not requiring an expert policy. However, results from Table 3 show that LFM mostly seems to substantially improve the initial policy learned with BC from expert demonstrations.

Questions

- In Table 3, BC results are different from "Prev SOTA" on ALFWorld and Touchdown. Why is this the case, and how did the authors collect expert demonstrations that led to a policy better than SOTA? - When describing the ACTPRED baseline l237, it is mentioned "First, we execute k steps of the base policy, then query the LLM for the next action". What are these k first steps and why are they necessary? - The authors computed the perplexity averaged over tokens of each possible action with the policy (that is an LLM) and selected the action with the minimum perplexity. Why did the authors not use the (log) probability of the sequence of tokens as in SayCan (Ahn et. al, 2022), GLAM (Carta et. al 2023) or TWOSOME (Tan et. al 2024)? - Also, computing the perplexity (or probability) of each action to follow the prompt can be very extensive when the action space grows, given that the policy is still a fairly large model (770M). Did the authors use anything to make this fast enough, especially in ScienceWorld where the set of possible actions can be very large? - In LFM, an initial phase uses the policy to collect trajectories that are then given to GPT4 in order to obtain the dataset to finetune the feedback model. In the experiments, 10k trajectories of 20 steps have been collected if I understood well. Were these trajectories also appended to the dataset used to improve the policy later on? - The authors showed an example of an explanation provided by the feedback model (based on Flan-T5 770M) in Table 5. Given the limited abilities in text generation of Flan-T5 770M, did the authors perform a deeper study of the explanations provided by the model, especially on the unseen test tasks?

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

2

Limitations

The authors discuss the limitations of the introduced method in the appendices. They identify that, while LFM improves compute efficiency by distilling feedback abilities from GPT4 into a smaller LLM, Flan-T5 770M is still a fairly large model to be called at every step, both for the policy and the feedback. Also, Appendix G shows that using Llama 2 70B instead of GPT4 can significantly impact the results. This is not a strong limitation, in my opinion, given that GPT4 is only used in the initial phase to train the smaller LLM.

Reviewer hZ534/10 · confidence 4/52024-07-12

Summary

This paper provides a policy improvement method with a Language Feedback Model (LFM) for decision making tasks. The proposed method mainly consists of two stages: (1) training a Language Feedback Model (LFM) and (2) improving a policy model with the trained LFM. In the first stage, to train a LFM, the initial policy is used to generate rollouts by interacting with the environment. Then, a LLM is used to generate text feedback on each action in the rollouts. Then, a LFM is trained on this feedback dataset. In the second stage, the trained LFM is used to generate feedback on each action generated by the initial policy. Then, actions that are predicted as desirable from the LFM are collected as a dataset. Finally, the policy is trained on the dataset that contains desirable actions. This paper evaluates the proposed method on three decision making benchmarks: Touchdown, ScienceWorld, and ALFWorld. This paper empirically demonstrates that the proposed method provides better scores than baselines such as BC or ActPred.

Strengths

S1. The idea of learning a language feedback model (LFM) to provide text feedback on each action sampled from the initial policy is interesting. S2. This paper provides experimental results on three representative benchmarks: Touchdown, SecienceWorld, and ALFWorld.

Weaknesses

W1. In the introduction section, this paper mentions that sample-efficiency and generalizability are important in instruction-following agents. I am not sure that the policy improvement with LFM is sample-efficient and generalizable. The proposed method trains the initial policy on a rollout dataset where desirable actions are collected by the LFM. Training on many rollouts does not seem sample-efficient. Also, the policy trained on a specific environment may lose generalizability. W2. In Section 3.2 (i.e., Learning a feedback model), it is rather unclear how to collect desirable behavior D_k. W3. It would be better to provide ablation study that shows effectiveness of the proposed method. For example, the authors may compare the policy improvement with an expert LLM to the policy improvement with the LFM.

Questions

Q1. Regarding the weakness W1, how many rollouts are needed to train a LFM? Also, how many revised rollouts are need to train an initial policy? Q2. Regarding the weakness W2, if we only select desirable actions from a rollout, does the revised trajectory correctly reflect or follow the environment dynamics or the transition function?

Rating

4

Confidence

4

Soundness

2

Presentation

2

Contribution

2

Limitations

The authors provide limitations of their work in Appendix A (i.e., Limitations).

Authorsrebuttal2024-08-12

Dear Reviewer, We wanted to send a friendly reminder that we are awaiting your response. With the deadline for the reviewer-author discussion approaching on August 13, we would greatly appreciate it if you could provide feedback at your earliest convenience. We hope we have addressed your concerns and questions. If so, would you please consider increasing your score to show support for our work? Sincerely, Authors

Reviewer hZ532024-08-13

After the Author Response

Thank you for providing thoughtful responses to my comments. For now, I maintain my initial rating. However, I am open to AC's decision on this paper.

Authorsrebuttal2024-08-13

Thank you for your acknowledgement! Is there anything unsatisfactory about our response, such that the you would not consider increasing your score? Specifically, 1) we provided clarifications regarding sample-efficiency and generalizability, 2) we elaborated on how desirable behaviour is collected, and 3) we noted that the existing manuscript does contain comparing to expert LLMs as the reviewer requested. Is there anything else the you would like to discuss in order for you to support our work?

Reviewer u2Y56/10 · confidence 4/52024-07-12

Summary

The papers present a method to essentially filter which actions should be used to learn a policy via imitation learning. The method follows the online imitation learning setting and replaces the expert policy with a language feedback model (LFM) distilled from a LLM. The LFM evaluates which transitions from the policy's rollouts should be used to train the policy. A LFM is used instead of LLM to reduce the computational complexity of the task, and both consumes and produces text. The method is evaluated on several benchmarks. The method is compared against several ablations over the design decisions.

Strengths

- The paper proposes an interesting idea to filter out the data that should be used for imitation learning, and then proposes to do this with a language model. - The method demonstrates clear improvements to methods evaluated against.

Weaknesses

- The authors claim that their LFM method is better than using a LLM as an expert policy, because it scan provide human interpretable feedback. However, the authors do not provide any results to suggest that the LFM can produce outputs humans can do something with. - The only comparisons are to ablations of the proposes method. At a minimum, some of the mentioned related work (like MOTIF (Klissarov et al [22])) should probably be a baseline. - The LMFA 1 rnds and LMFA 2 rnds in Table 3 do not seem to be discussed in the main body. - There are some gaps writing that have left me with a lot of questions/uncertainties (see below). - Small things: - you have places in the PDF with weird formatting, e.g. lines 156 - 157. - The reference to Figure 1(b) in Section 2 line 72 should probably be 1(c)

Questions

- It is unclear how the method is marking states as desirable in section "learning from language feedback" is different from that in "naively learning from LLM feedback". Is there a connection missing that links how the LFM is "efficiently" learned and how the policy learns? Instead of querying the LLM at each step, the LLM is queried of a trajectory of steps and responds by saying which steps were desirable? - Is there also a reward function? - What is an iteration/round in this method? i.e. "In round k, we rollout the base policy...." (Section "Learning from language feedback") - You state, "Unlike these works ... we consider settings where training and evaluation goals are different." It is not clear to me how the training and evaluation goals differ in your set up. - How do you decide when to summarize what has happened with a "Before" in the LLM prompt? e.g. Table 2 - LFM prompt. - What does it mean that you "...limit the amount of LLM usage to 100k GPT-2 tokens"? Did you constrain the number of input tokens? The number of output tokens? How did you go about the constraints? - How did you select 20 as the number of timesteps over which LLM feedback was collected? - You say you subselect feedback data to have an even split of productive and non-productive action. How much data do you actually end up with? What were the originally ratios? How diverse are the samples? - For the results section on generalization to new environments, what makes a new environment new? Table 4 holds results for ALFWord, ScienceWorld, and Touchdown, which are the environments you report training on.

Rating

6

Confidence

4

Soundness

3

Presentation

2

Contribution

3

Limitations

There are no limitations in the main body.

Reviewer 8DkG5/10 · confidence 4/52024-07-22

Summary

The paper proposes to train a language feedback model and leverage the language feedback model to conduct policy improvement in language-based task. The authors also propose a pipeline that can apply CLIP to convert images into textual description. Experiments over ALFWorld, ScienceWorld and Touchdown validate the effectiveness of proposed algorithms.

Strengths

1, The paper is overall clear and easy to follow. 2. The idea of training a separate language model to serve as language critic function for policy learning is overall novel in decision-making task. 3. The experiments cover a large range of tasks, including one visual task -- which is great to demonstrate the generality of the proposed algorithms.

Weaknesses

1. The improvement is mainly from distilling language feedback from stronger models (such as GPT-4), which somehow limits the technical contribution of proposed algorithm. Is it possible to derive language feedback from exactly the same model? (FLAN-770M might be impossible, but what about larger one like llama-3-8b?) 2. The experiments are not comprehensive enough. For instance, the table 3 presents ALF-world's SOTA from results in paper from 2021 -- which I believe is quite out-dated and there are plenty of work that improve ALF's performance a lot. Also, the authors miss several important baselines, 1. RL -- since the policy evaluation + policy improvement is the basic foundation for RL. It is essential to compare the proposed algorithms with RL. 2. a bunch of work starting after reflection, that also uses verbal feedback to improve the performance.

Questions

1. There are a few literatures that might be highly relevant to the paper, including: [1] Shinn, Noah, et al. "Reflexion: Language agents with verbal reinforcement learning." Advances in Neural Information Processing Systems 36 (2024), which also studies how to use verbal descriptions as feedback to guide the LLM's policy. [2] Feng, Xidong, et al. "Natural Language Reinforcement Learning." arXiv preprint arXiv:2402.07157 (2024), which is also motivated by verbalizing policy learning process. [3] A bunch of work covering llm-as-judge, like: Wang, Yidong, et al. "Pandalm: An automatic evaluation benchmark for llm instruction tuning optimization." arXiv preprint arXiv:2306.05087 (2023). 2. Why choose GPT-4-0315 to conduct experiments? This is a relatively old GPT4 model considering it's mid 2024 now? And is there any other ablation studies covering different types of llm? 3. By checking the prompt template shown in table 2, I have a question about: why there is no COT process between the model judge yes or no? Is it done on purpose or is there any explanation for that since in most setting COT can enhance GPT-4's performance -- and I believe this yes/no is very important since it directly influence the performance of policy improvement. 4. The comparison between LFMD and LFD or other baselines seems a bit unfair? Since you still require to run rollouts on test set while the other results are zero-shot.

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

See Weakness and Questions.

Authorsrebuttal2024-08-12

Dear Reviewer, We wanted to send a friendly reminder that we are awaiting your response. With the deadline for the reviewer-author discussion approaching on August 13, we would greatly appreciate it if you could provide feedback at your earliest convenience. We hope we have addressed your concerns and questions. If so, would you please consider increasing your score to show support for our work? Sincerely, Authors

Authorsrebuttal2024-08-13

Last day of discussion

Dear Reviewer 8DkG, Today is the last day for discussion. Would you please take a look at the author response before the discussion ends? We hope we have addressed your concerns and questions. If so, would you please consider increasing your score to show support for our work? Thank you! Authors

Reviewer B45L2024-08-09

Response to rebuttal

The authors properly answered all my questions and concerns. I will keep my score as it is.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC