Improving Language Model Negotiation with Self-Play and In-Context Learning from AI Feedback

We study whether multiple large language models (LLMs) can autonomously improve each other in a negotiation game by playing, reflecting, and criticizing. We are interested in this question because if LLMs were able to improve each other, it would imply the possibility of creating strong AI agents with minimal human intervention. We ask two LLMs to negotiate with each other, playing the roles of a buyer and a seller, respectively. They aim to reach a deal with the buyer targeting a lower price and the seller a higher one. A third language model, playing the critic, provides feedback to a player to improve the player's negotiation strategies. We let the two agents play multiple rounds, using previous negotiation history and AI feedback as in-context demonstrations to improve the model's negotiation strategy iteratively. We use different LLMs (GPT and Claude) for different roles and use the deal price as the evaluation metric. Our experiments reveal multiple intriguing findings: (1) Only a subset of the language models we consider can self-play and improve the deal price from AI feedback, weaker models either do not understand the game's rules or cannot incorporate AI feedback for further improvement. (2) Models' abilities to learn from the feedback differ when playing different roles. For example, it is harder for Claude-instant to improve as the buyer than as the seller. (3) When unrolling the game to multiple rounds, stronger agents can consistently improve their performance by meaningfully using previous experiences and iterative AI feedback, yet have a higher risk of breaking the deal. We hope our work provides insightful initial explorations of having models autonomously improve each other with game playing and AI feedback.

Paper

References (34)

Scroll for more · 22 remaining

Similar papers

Reviewer Tqe85/10 · confidence 3/52023-07-05

Summary

The paper studies the ability of LLMs to improve in a negotiation game. The find that only a subset of language models can self improve from AI feedback, a model's ability to learn from feedback depends on its role in the game, and stronger agents can go through more rounds of negotiation.

Strengths

- The paper is generally well written; it was more or less easy to understand the entire paper. - The experiment methodology is sensible. It was nice to narrow down the set of models by eliminating models based on their ability to respond to feedback or understand the problem. - The results are interesting and could be valuable to the community on understanding the role of LLMs as agents.

Weaknesses

- It seems like incorporating feedback is done by providing the feedback as context for the LLM. The paper could be made stronger by utilizing the fine-tuning APIs for the given models. - Using GPT3.5 as the fixed agent is interesting. It's clear GPT-4 is the strongest agent in this scenario and it'd be interesting to see how well these results would hold against a stronger agent. In a similar vein, seeing how well these results would hold when actually negotiating with a human. - Could buyer's and seller's be prompted better? E.g. would it be possible to prompt claude instant v1.0 more effectively to respond to multi-turn negotiations or prompt Cohere command or AI21 j2 better to respond to bargaining and feedback respectively?

Questions

Overall, I think the work is interesting, it's reasonably well designed, and has interseting implications in reinforcement learning and large language models. I think there were some decisions (e.g. incorporating feedback in the context passed to the LLM, using GPT3.5 as the fixed agent that we are buying/selling against) that make the paper's results a little less clear. However, the results are interesting and show a first step towards using LLMs as agents, in negotiations.

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

3 good

Presentation

3 good

Contribution

2 fair

Limitations

The authors partially address limitations of their work, and address societal implications of their work.

Reviewer sQTf5/10 · confidence 4/52023-07-05

Summary

This paper investigates the intriguing possibility of autonomous improvement among multiple large language models through a negotiation game. By assigning various LLMs to distinct roles and allowing them to engage in iterative improvement, the paper aims to enhance their negotiation strategies without human intervention. The study uncovers some insights into negotiation problems, encompassing the assessment of model capabilities and their responses to AI feedback. However, after careful evaluation, I am inclined to reject this paper due to its limited contribution and insufficient experimental results, which fail to adequately demonstrate the effectiveness of its findings.

Strengths

This paper studies an interesting problem: Improving large language models with each other and demanding only black-box access is a promising direction. This paper is largely clear and concise. It is easy to follow the problem setting and the negotiation process. This paper compares capabilities (especially the continue learning ability using in-context learning) of some advanced large language models in the proposed negotiation problem, which is interesting.

Weaknesses

One of the main technical novelties is the AI feedback technique used in their method. However, this technique seems like a result of random attempts plus some intuition. More detailed though on how did the authors develop this technique or the comparison between other possible candidate techniques is needed. Moreover, this technique is similar to CoT. It would be better for the authors to discuss on the relationship between these two techniques. I am curious about the performance of guiding LLMs to think step by step without relying on additional critics in negotiation problems. More thorough explanations or experimental results of this aspect would provide deeper insights into the effectiveness of the proposed approach. This paper aims to improve the ability of large language model by self-play with each other. However, to strengthen the paper's claims, it would be better to provide additional evidence regarding the transferability of the proposed framework to different types of games. Suggestions on the specific implementation process would also be valuable in applying this framework to different domains. The stability of the environment is also a concern since the results appear to heavily rely on the reliability of the moderator. This dependence raises doubts regarding the individual contributions of various components in the system. Furthermore, the claim of proposing a technique for prompt optimization for generic classification tasks in Line 120 lacks sufficient details and evidence. Elaborating on this technique and demonstrating its effectiveness would enhance the paper's credibility and address this specific weakness. It is not clear whether there exists an upper limit of improvement using ICL. It would be better for the authors to also discuss ICL with, e.g., fine-tuning or other trainable parts.

Questions

What is the exact definition of "self-play"? It seems that only one of the two players improves in the setting of this paper, which slightly different previous definition in, e.g., https://arxiv.org/pdf/2002.04017.pdf. Line 59: It would be beneficial to provide additional details, possibly with references, comparing the effectiveness of using AI feedback with RHLF. This comparison would contribute to a more comprehensive understanding of the proposed approach's advantages and distinguish it from existing techniques. Line 85: Why the LM engine behind the critic is consistently the same as the player it provides feedback. Figure 4: It lacks the experimental results of Claude-v1.3 and GPT-4 as buyers. Why these scenarios are omitted? Including these results in the figure would ensure a complete representation of the experiments conducted. Presentation: Line 118, 119: inspect -> inspecting, add -> adding Line 120: recommend -> and recommending Line 211: comparison -> comparisons Figure 5: whild -> while Ling 257: use less -> uses fewer Line 258: serve -> serves

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

2 fair

Presentation

3 good

Contribution

1 poor

Limitations

The authors discuss the limitations of the work in the last section.

Reviewer nXyn4/10 · confidence 4/52023-07-06

Summary

This paper studies whether multiple large language models (LLMs) can improve each other in a negotiation game by playing, reflecting, and criticizing, with minimal human intervention. Two LLMs play the roles of a seller and a buyer, and a third LLM plays the role of a critic who provides feedback to one of the players to improve their negotiation strategy. The authors report several intriguing findings, such as the different abilities and behaviors of various LLMs in the game, the trade-off between deal price and success rate, and the evidence of improved language complexity and strategy from iterative AI feedback.

Strengths

1. Interaction between LLMs is quite interesting and potentially important for future AI research. 2. The authors experimented with a variety of LLMs, including open-source (Cohere) and proprietary (GPT, Claude) ones.

Weaknesses

Major Issue: 1. The negotiation setting is quite contrived and not well-grounded. Specifically, there is no context about the goods being discussed. Also, the feedback and overall conversation seems very generic (Figure 2). There is no value/intrinsic motivation for the buyer to get the goods, and there is only one choice to go for. Typically, negotiation in real-life doesn't happen this way. 2. It feels like the current LLM negotiation setting can be viewed as `predicting the most probable outcome/text' given the negotiation contexts. 3. AI feedback is not the same as human feedback, the behavior may be completely different in real-world situations since there are a lot of other concerns including 'value', 'time', 'personal preference' for negotiation. This makes me wonder what this simple negotiation has to offer in terms of understanding/future research. Minor Issue: 1. The title is a bit misleading. It seems to me from the title that the paper is about an algorithm, but it is more about evaluations and understanding. [After Rebuttal] My concern still lies in the 'motivation' behind the setting for the experiment. In my understanding, negotiation requires a setting, e.g., both seller and buyer will have a stake on the product (buyer want to pay lower than the *utility* of the product and seller want to earn more than the *cost*). However, such basic setting is not present in the experiments. This makes me concerned if the observed effect is really negotiation or just the effect of instruct fine-tuning. I decided not to change my score.

Questions

1. I wonder why LLAMA based models are not considered in this work. 2. Is it possible to have both the seller and buyer receive feedback in a single experiment? I wonder what will happen. 3. Is it possible to add some story/context to the negotiation so that the LLMs have a better motivation? 4. Is the best policy to just always be fixed at starting price and ask LLM to generate some excuse (Prompt: "Give me some sentences to sell my __ for the highest possible price.")? [After Rebuttal] All questions have been addressed.

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good 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

3 good

Limitations

I was unable to find any information related to limitations.

Reviewer SEMS4/10 · confidence 4/52023-08-01

Summary

This paper studies the strategic multi-agent problem setting of two LLM Players interacting in a negotiation (or bargaining) game and proposes to use feedback from an LLM Critic to improve each Player’s expected behavior and performance in the game. Importantly, the paper aims to study how AI Feedback can enable Player improvement when playing competitive games under well-defined rules. The proposed method is to use the player dialog history and critic feedback as “in-context demonstrations” for players to self-improve over the course of the game. Experiments are conducted on a bargaining game instance negotiating the price of a balloon, investigating several LLMs as base models for the Players (and their Critics).

Strengths

Identified strengths of paper: - I really like the premise of this work: that agents playing in a strategic game may each be able to benefit and improve their negotiation strategies, not only by observing their opponents, but additionally by heeding the advice of personalized critics whose aim is to help (cooperate with) them. Like other AI feedback methods, this type of approach has the potential to scale better than using Human feedback for the same type of assistance, presuming feedback is helpful and thus desirable for this negotiation task. - The topic is timely and very relevant to the multi-agent and game theoretic communities, as it takes a long-existing and well-studied problem (multi-agent negotiation) and investigates how rational player strategies can be improved using AI feedback and LLMs.

Weaknesses

Identified concerns and suggestions to improve the paper submission: - It would be extremely helpful to see examples of what the *instructions* and *in-context learning* looks like for the Critic agents? Examples of their few-shot prompts (or instruction data for Finetuning) are important, as this is a key contribution of the work (adding AI Feedback to strategic reasoning/negotiation settings). Notably, where are the Buyer/Seller Critics getting the suggested negotiation strategies in Figure 1B from (e.g. the flinch technique, the anchoring technique, etc)? They seem to be good, general strategies to have in the player’s toolbox, but require the Critic to have knowledge of effective negotiations and how/when to employ them. Thus, how are the Critics coming up with general negotiation strategies and reasoning about when to apply them? This is important for opening the black box of the AI Feedback component of the proposed method. - While I like the overall premise, the novel technical contribution of this work seems tenuous. With that, the problem setting is strongly inspired by AlphaGo Zero, but is there any traditional learning in this setting (i.e. any updates to the agent policies based upon the AI feedback given)? That is still unclear to me. Currently the paper seems to simply add a small amount of additional context in the prompt for the Player LLMs. If there is no traditional learning, this setting is also critically different from AlphaGo Zero in that the proposed feedback signal (from Critic Agents) is **not** used to update the Player policies. This should be explicitly clarified. Also, if there are model updates, it would be useful to see the learning update rules. If not, why not do any finetuning of the model? At least as an experimental condition to compare against and understand if in-context learning is "sufficient". Perhaps more motivation and context for this design decision would be helpful. - This work explores a potentially interesting research direction (negotiation games or strategic gameplay more generally + AI Feedback for improved player strategies) but it’s not clear to me how well motivated AI feedback is for this negotiation game or how interesting the problem is in its current instantiation. In particular, it would be helpful if the paper provided some type of performance analysis to show that the negotiation game used (an instance of balloon purchasing) is interesting and challenging to solve *before* ever adding in LLMs or a Critic. How would existing multi-agent negotiation/bargaining approaches solve this game? What equilibria do the Players generally converge to without LLMs? How and why does using LLMs change the *expected* solution? Given two Player LLMs, why is a Critic *necessary* to have? In other words, if the Critic is being prompted with examples of bargaining conversations, could the Buyer/Seller agents not simply see those same examples and then converge upon the same solution (they are currently finding) *without* a Critic? Or is there some value that the Critic role *uniquely* adds (e.g. more efficient or robust convergence on an equilibrium in this negotiation game)? Motivating the game selected as an interesting and challenging problem in its own right is critical. - Furthermore, I still question the generality and significance of the paper’s empirical findings. **RE Significance:** How significant and meaningful are the differences in Seller performance in Table 1? The numbers don’t seem that different to me, but perhaps with more context, the significance becomes more clear. With that, I see a distributional shift in Figure 4, but how meaningful is this shift? What are the implications of it, regarding how good or bad these solutions/equilibria are before and after receiving AI feedback? Do the solutions to the negotiation game simply change a small amount but are comparable in terms of how preferable/desirable they are or do they get qualitatively better in some way? **RE Generality:** The paper seems to use only one evaluation domain (balloon purchasing). It is not clear to me which empirical findings/trends from the use of AI Feedback on this one, seemingly simple bargaining domain, are expected to generalize. In particular, regarding the effectiveness and impact of AI Feedback in other multi-agent negotiation domains and settings. This limited evaluation seems more like an interesting case study than general findings that transfer. - In Subsection 4.3, the paper provides analysis of each of the LLM models (gpt-3.5, gpt-4, claude-v1.0, etc) versus a fixed gpt-3.5-turbo opponent. However, a more thorough investigation examining the cross product of each LLM model as Buyer against every other LLM model as Seller (a *complete* Payoff matrix with all pairwise model comparison) could potentially provide more general insights. Why was this not done? Can it still be done to show a resulting Payoff Matrix? Not doing a pairwise comparison of *all* pairs of models might unnecessarily constrain the results and thus the insights that can be extracted.

Questions

Please see Weaknesses section.

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good 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

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

Please see Weaknesses section.

Reviewer nXyn2023-08-13

Thank you for the detailed responses. Overall, I think the setting of the experiments is not solid enough: 1. Use of GPT-3.5/GPT-4 as evaluator. If GPT-3 makes a mistake then the results may not be accurate. 2. Meaningless negotiation. There is no context for the LM seller for the 'cost' of the good and no context for the buyer for the 'utility' in the economic model. There is no practical reason where the price should be set. This makes the task more like some NLP benchmark. 3. Limited LLM formats. Different LLMs may behave differently under different prompts and different LLMs may also behave differently. Therefore, I'm concerned how instructive this work would be for future works. Overall, I would retain my score for the paper. Please let me know if there are any further misunderstandings.

Authorsrebuttal2023-08-13

We thank the reviewer for the response. We would like to further clarify certain misunderstandings: - **We do not use GPT-3 as evaluator**: there is no “GPT-3” in this paper — all experiments we use GPT-**3.5** and GPT-**4**. We use the **final deal price as the evaluation** (see our Figures 2, 4, 5, 6), which is a standard practice of AI bargaining research (e.g., see Heddaya et. al. 2023). We are well aware of how GPT-based evaluation can be biased (e.g., see Wang et. al. 2023), this is why we choose the evaluation to be the final deal price, as it is more objective. - **Our experiments clearly show how the negotiation improves in multiple meaningful ways**. The value of the product is discussed multiple times during negotiation, and we give many examples in the paper (e.g., Figure 2C: “high quality latex and handcrafted by expert artisan”; Figure 7 “made from durable material”) - **The prompts, especially after AI feedback, is of very high diversity.** We note that one should not view the initial instruction as the fixed prompt (we keep the initial instruction fixed because we need to use the same way to explain the rule of bargaining to different agents), but should **view the [initial instruction + round 1 dialog + round 1 AI feedback + round 2 dialog + round 2 AI feedback … ] collectively** **as the prompt** for the next round negotiation. To give a sense of how diverse the prompts are, in Figure 6, we run the game 5 rounds, and repeat 500 times, which in total gives 2500 different types of prompts. Because our prompts are so diverse, we observe multiple meaningful behaviors and a wide range of deal price distributions. We show some of the examples in Figure 1B where we sample four types of AI feedback, as the updated prompt for the player, and how it improves players’ strategy. References Heddaya et. al. 2023. Language of Bargaining Wang et. al. 2023. How Far Can Camels Go? Exploring the State of Instruction Tuning on Open Resources

Reviewer nXyn2023-08-14

Thank you for the clarifications. I have adjusted my comment to better reflect accuracy. By GPT-3, I meant language models in general

Authorsrebuttal2023-08-14

we would appreciate reviewer nXyn could keep their original comments instead of deleting / modifying them after we have addressed them

We thank the reviewer nXyn for their response. About evaluation, we would like to re-iterate: - **We do not use any GPT/ LLM as the evaluator**. We are well-aware the drawbacks of using LLM as evaluators, as is discussed in concurrent works like Wang et. al. 2023. - **We use the final deal price as the evaluation**, as the common practice of bargaining literature, because it is more objective. Additionally, **we would appreciate reviewer nXyn to keep their original comments instead of deleting / modifying them after we have addressed them**, specifically: About whether the players are having meaningful negotiation, initially the reviewers wrote > “There is no context for the LM seller for the 'cost' of the good and no context for the buyer for the '**value**' in the economic model” > **After we have addressed how the value of the product is discussed in the above response, the reviewer deleted the word 'value'** and changed their comments to be > There is no context for the LM seller for the 'cost' of the good and no context for the buyer for the '**utility**' in the economic model > Similarly, for the diversity point, initially the reviewer writes: > “Limited LLMs/prompt formats” > **After we have clarified the diversity of the prompt, the reviewer deleted the word 'prompts'** and changed this to be > “Limited LLM formats” > The reviewer’s original post can be found in the OpenReview’s notification email. Further on the reviewer’s updated comments: - We would like to note again that at the current stage, some model engines cannot even understand the basic commonsense like \\$8 is a lower price than \\$10. **How can we ask them to incorporate “the 'value' in the economic model” when they don’t even understand the very basics like \\$8 is lower than \\$10**? - About the concern about “limited LLM formats”, **we have already considered most leading LLMs** (Cohere, AI21, Claude-Instant-1.0, Claude-1.3, GPT-3.5, GPT-4) at the time of our experiments, note that although there are many open-source LLaMA based models, they do not exist when we started our experiments

Reviewer nXyn2023-08-14

Hello, Thank you for organizing the history/edits. Please note that I have always meant ‘value to the buyer’ or utility starting with the first version of my review.

Authorsrebuttal2023-08-14

We thank the reviewer for the response. We understand that the reviewer has been insisting on adding “value to the buyer” or ”utility” as well as other factors like “context about the goods”, “intrinsic motivation”, “personal preference” .etc. These are great directions that we very much love to explore. However, as we have been emphasizing, **how can we ask the model to incorporate such complex ideas when they don’t even understand the very basics like \\$8 is lower than \\$10?** The development of LLMs is simply not that advanced yet. We would like to note that at the current stage of science, it is more important to set up the basics to test whether LLMs can autonomously improve each other through game playing and AI feedback in a minimal setting, which is precisely the contribution of this work.

Reviewer nXyn2023-08-14

Hello, I understand that the deal price is the evaluation metric, but doesn’t the determination of the deal price involve an LLM? Also, in regards to value/utility. I regard the utility or value to the buyer as a key to making a deal, similar to the cost to the seller. If the argument is that we cannot include any context before we start negotiating because some LLMs do not understand context, then then it is unclear how much of the whole negotiation process the LLM understands

Authorsrebuttal2023-08-14

We thank the reviewer for the continued discussion. Below is further clarification: "but doesn’t the determination of the deal price involve an LLM?" * We use the language model to classify whether the two players have reached a deal. This classification has achieved 90+ accuracy after extensive prompt engineering. * Then we extract the actual price number by regular expression -- simply extracting the last number in the sentence (also the practice in many existing works using numbers as metrics, e.g., in Fu et. al. 2023) and it is accurate enough -- we generally believe this part is simple and can be achieved by any researchers who have basic experiences of prompt engineering. Regarding the motivation for making the deal * **We instruct the model to buy / sell the product to a lower / higher price** by putting "Your goal is to buy it with a low price, as low as possible" in the initial instruction. * Because the models are trained to follow human instructions, we tend to believe our **human instruction is their largest motivation / the key for them to make a deal, not their own interest** -- or in other words, they are more or the less doing the bargaining on because they are trained to follow human instructions that ask them to do so. * Again, we very much love to test how the model will behave if they bargain for their own interest, but that is beyond this work's scope -- our objective is to set up a minimal testbed to study if language models, under human instructions, can self-improve over AI feedback We hope this clarifies the misunderstandings. References Fu et. al. 2023. Complexity-Based Prompting for Multi-Step Reasoning

Reviewer sQTf2023-08-21

Response to the rebuttal

Thanks to the authors for the detailed response. I appreciate the additional clarifications from the authors.

Program Chairsdecision2023-09-21

Decision

Reject

© 2026 NYSGPT2525 LLC