AI for Global Climate Cooperation: Modeling Global Climate Negotiations, Agreements, and Long-Term Cooperation in RICE-N

We thank the reviewer for their thoughtful feedback and for pointing out the positive aspects of our submission, such as the novelty of the protocols, the realism of the components, the empirical support for our claims and the overall quality of the evaluation criteria, experimental design and results analysis. We hope our response clarifies any outstanding concerns.

Could region level analysis be provided? For example, what are a region’s actions overtime and how they are different with the two negotiation protocols and without negotiation. Are there any patterns in the actions for the 27 regions studied. Could their actions be categorized into a few common groups?

Thank you for this suggestion. Under Basic Club, all regions tend to increase their mitigation rate as much as constraints allow until the maximum mitigation rate is reached. Without negotiation acting as a coordination mechanism, regions tend to fall into different subgroups. There are ambitious mitigators, reducing 40-60% of emissions by the end of the run, moderate mitigators reducing 20-30%, and free riders reducing 0-10%. In the camera-ready version, we will include a detailed quantitative and qualitative analysis of these phenomena.

Can you summarize possible usages of the proposed approach for different types of users: ML researchers, climate scientists, government, international organizations such as IPCC?

ML researchers can make use of the modularity of the RL component to evaluate how different RL algorithms perform in a real world-calibrated environment. Climate scientists can use it to compare different climate modules and damage functions. Governments can use it to evaluate the robustness of a given climate-economic policy to strategic behavior and estimate the consequences on international trade. International organizations, such as the OECD or WTO could use the tool to analyze the economic trade-offs of different climate policies and negotiation strategies. We will add these potential uses to the text, while making sure to surface the limitations.

Figures 6-15 are not explained.

Thank you for pointing this out. These results are additional visualizations to support Section 6. In the camera-ready, we will expand the captions for each of the images such that they are sufficiently informative to understand the figures, describing the various scenarios, as well as the variables under study, and what can be concluded from the plots. We will also point to them more strongly in the main text.

We thank the reviewer for the incisive review of our submission. Below, we respond to the concerns raised.

> does the 'no negotiation' baseline match the Nordhaus-Yang results?

We can compare our results to Nordhaus’ RICE from 2010 [1]. The aforementioned has two results: baseline and optimal. The former consists of each region optimizing its own objective function and the latter of a cooperative mode with a shared objective function. We compare our results with RICE up until 2115, as that is the end year of our simulation. Looking at global temperature only for simplicity’s sake, the RICE2010 baseline and optimal temperature rises are 3.91 C and 2.92 C, respectively. Our no-negotiation baseline and basic club protocols result in increases of ~4.21 °C and ~2.97 °C, respectively. Exact matches are not expected as initial conditions differ: our rollout starts in 2015 (in line with DICE2016, which served as the basis for the implementation) and RICE2010 in 2005. Nonetheless, both no-negotiation and basic club yield temperature rises comparable to the RICE 2010 base and optimal trajectory, respectively.

[1] ​​Nordhaus, W. D. (2010). Economic aspects of global warming in a post-Copenhagen environment. Proceedings of the NAS, 107(26), 11721-11726.

> Is [the code] more modular, faster/slower?

While Python is likely slower than e.g. GAMS for model optimization, the runtime is bottlenecked by the deep RL agents, not the environment dynamics. Writing RICE-N in Python makes it simpler for the simulation code to interface with the RL agent code, which can be implemented using common Deep RL frameworks in Python and reduces barriers for ML researchers to contribute to the codebase, and extend it with novel modelling components. Finally, we are working on an efficient JAX implementation, which we plan to release alongside the paper and which will improve the runtime by multiple orders of magnitude.

> My biggest concern with the paper is the specification of binding negotiations.

In our current setup, agents cannot deviate from agreed-upon actions for the specified time lapse (5 years). While this assumption simplifies analysis and isolates the impact of the negotiation mechanism, it does not reflect the uncertainty and strategic mistrust present in real-world climate negotiations (e.g., countries withdrawing from agreements).

Future work should explore the possibility of non-binding commitments, which would enable agents to engage in strategic communication or 'cheap talk,' thereby better capturing realistic negotiation dynamics [2]. We will add this discussion to the text.

[2] Caparros, A. (2016). The Paris Agreement as a step backward to gain momentum: Lessons from and for theory. Revue d'économie politique, 126, 347.

> Nordhaus & Yang issues related to international transfers >

​​The reviewer is correct to identify concerns related to international transfers at the scale required to implement RICE’s optimal pathways. For example, large-scale international transfers risk misappropriation without oversight [3]. However, in the time since Nordhaus and Yang’s 1996 comment on the practicality of international transfers, there has been a greater focus on equity and burden sharing [4], shifting the question from whether international transfers are possible to how do international transfers fit into a larger patchwork of climate finance instruments [5]. Adding an international transfers module is on the RICE-N development road map such that we can explore different means of financing mitigation.

[3] Nest, M., et al. (2022). Corruption and climate finance. Brief Chr. Michelsen Institute U, 4, 14.

[4] Landis, T. & Bernauer, T. (2012). Transfer payments in global climate policy. Nature Climate Change, 2(8), 628-633.

[5] Pickering, J., et al. (2017). Special Issue: Managing Fragmentation and Complexity in the Emerging System of International Climate Finance. International Environmental Agreements: Politics, Law and Economics, 17(1), 1-16.

> too many of the paper's details were hidden in the appendix.

Thank you for this feedback. We will use the extra page of the camera-ready version to discuss more RL-related details in the main text.

> it could be useful to plot policy functions as well

Thank you for the suggestion. We will add to the final version.

> Nordhaus appears in the references as both "W." and "W.D." "Stern" and "AI" should be capitalised in the references.

Thank you for catching this. We have corrected this oversight.

> why are 27 fictitious countries used, rather than the full set?

Many individual countries lack comprehensive data making modeling unreliable. By aggregating countries into regions, the errors from lacking in data from these countries would be less impactful to the whole region. We believe, for a similar reason, RICE-2010 only includes 12 regions. Besides, 27-region setting makes bilateral negotiation computational feasible due to $O(n^2)$ complexity.

Thank you for the thoughtful review and for highlighting the positive aspects, such as RICE-N’s suitability for studying strategic climate negotiations, the evidence-based claims, and the originality of our MARL application. We are pleased that you consider both the paper and the code accessible, and we hope that our response clarifies any outstanding concerns.

> Suitability for ICML.

Thank you for voicing your concerns. We believe that the application of MARL to a climate economic negotiation integrated assessment model provides a relevant methodological contribution to the Application-Driven Machine Learning track of ICML. Beyond its contribution to topics mentioned in the call for papers such as social sciences, and sustainability and climate for which the ML community demonstrates interest through concurrent work on climate investment [1], as well as the many potential extensions that are interesting to the ML community, such as the integration of LLMs into negotiation [2], this work also provides a useful testbed for topics of interest to MARL researchers, such as studying multi-party cooperation between AI agents in sequential social dilemmas [3].

[1] Hou, X., et al. (2025). InvestESG: A multi-agent reinforcement learning benchmark for studying climate investment as a social dilemma. In ICLR 2025.

[2] Vaccaro, M., et al. (2025). Advancing AI negotiations: New theory and evidence from a large-scale autonomous negotiations competition. arXiv preprint arXiv:2503.06416.

[3] Leibo, J. Z., et al. (2017). Multi-agent reinforcement learning in sequential social dilemmas. In Proceedings of the 16th AAMAS. ACM.

> Additional experimental analysis on effects of variables.

Since the space of possible configurations is large, we perform a sensitivity analysis over a subset of economically relevant parameters, namely the discount factor, welfare loss weight, consumption substitution rate and relative preference for domestic goods. This analysis showcases the robustness of our findings. We will add it to the appendix and reference it in the main text. We plot heatmaps of the analysis here: https://imgur.com/a/W0194Ej . We show the percentage change in outcome variables of interest across different scenarios when critical model parameters are perturbed by a multiplication factor ranging from 0.96 to 1.04. The max percentage change is 3.16% while mean is -0.22% and medium is -0.36%. We thus conclude the dynamics are stable, corresponding to changes in critical model parameters.

We will also include a discussion on the computational complexity of RICE-N (please see our response to QTAV).

> 208, column 2: clarify that the link is to Section 6

Thank you for spotting this.

> possible unintended consequences

We wish to highlight the importance of qualifying these claims by discussing the potential unintended consequences of climate clubs being implemented without redistributive financing and technology sharing. Uniform tariffs on developing countries with lower mitigation rates would effectively serve as a tax on less developing countries, which we do not advocate for [4,5]. Our goal is to create a simulation framework where these dynamics and alternative policies can be explored and cross compared. We will give an account of these points in the main text of the camera-ready version.

[4] Goldthau, A., & Tagliapietra, S. (2022). How an open climate club can generate carbon dividends for the poor. Bruegel-Blogs.

[5] Perdana, S., & Vielle, M. (2022). Making the EU Carbon Border Adjustment Mechanism acceptable and climate friendly for least developed countries. Energy Policy, 170, 113245.

> loosely coupled

To clarify: this comment refers to the structure of the code, not any functional difference. We mean that we try to reduce as much as possible the number of assumptions that each component must make about the other components, and make the necessary dependencies between dynamics both explicit and local [6]. This makes extensions to the existing codebase much easier to implement, which makes testing new ideas, such as novel negotiation protocols, much easier.

[6] Leymann, F. (2016, September 5–7). Loose coupling and architectural implications. Keynote address presented at the ESOCC, Vienna, Austria.

> uncertainty estimates

The uncertainties arise from the stochasticity of the learned policy as we sample actions from the policy's distribution during evaluation. We estimate uncertainty across 50 rollouts with varying seeds (see Section 6, paragraph “Experimental Setup”). The relatively small uncertainty shows that the agents have learned robust policies. An increase could indicate that the learned policy is less reliable, potentially due to insufficient training or the model's inability to generalize effectively. Adding uncertainty to the environment (e.g., the climate component) or relevant parameters (see sensitivity analysis) would likely result in an increase of the uncertainty.

We thank the reviewer for their favorable assessment regarding the innovativeness and suitability of our chosen approach, as well as the constructive feedback. Below, we carefully address any outstanding concerns:

> detailed sensitivity analysis

Thank you for this suggestion. We select parameters based on economic theory, which are the discount factor, welfare loss weight, consumption substitution rate and relative preference for domestic goods. This analysis showcases the robustness of our findings. We will add it to the appendix and reference it in the main text. We plot heatmaps of the analysis here: https://imgur.com/a/W0194Ej . In the heatmap, we show the percentage change in variables of interest (including Temperature, Carbon Emissions, GDP) across different scenarios when critical model parameters are perturbed by a multiplication factor ranging from 0.96 to 1.04. The max percentage change is 3.16% while mean is -0.22% and medium is -0.36%. We thus conclude the dynamics are stable, corresponding to changes in critical model parameters.

> real-world applicability of the Basic Club protocol

Basic Club and carbon border adjustment mechanisms (CBAM) fight carbon leakage by setting the strength of the tariff proportionally to the difference in the cost of carbon between the exporting and importing countries. Basic Club relies on a uniform tariff [1]. In contrast, CBAM targets specific goods, which can exacerbate carbon leakage by leaving large swaths of emissions unaccounted for, such as those produced for non-EU exports. Climate clubs can go further than what is modeled in RICE-N at the moment, hence the term “Basic”. Uniform tariffs, coupled with technology sharing and redistribution, are theorized to be more effective at curtailing carbon leakage [2]. Future work should implement CBAM in RICE-N, which requires disaggregating production and trade by sector to allow for targeted tariffs. This is currently a work in progress and high on our agenda.

[1] Overland, I., & Huda, M. S. (2022). Climate clubs and carbon border adjustments: A review. Environmental Research Letters, 17(9), 093005.

[2] Tarr, D. G., Kuznetsov, D. E., Overland, I., & Vakulchuk, R. (2023). Why carbon border adjustment mechanisms will not save the planet but a climate club and subsidies for transformative green technologies may. Energy Economics, 122, 1066

> What are the computational requirements of the MARL approach, particularly for large-scale simulations with many regions?

The computational complexity of our MARL approach is driven by the number of regions N (i.e., agents). Since each agent’s action space scales linearly with N, the total action space across all agents grows quadratically (O(N²)). However, the number of agents in our setting is naturally bounded by the number of countries on the planet. Currently, training 27 agents for 100 thousand episodes takes approximately 3 hours on a 30 CPU cluster. To improve runtime, we are exploring more efficient implementations using JAX-based acceleration and model parallelism, which would enable large-scale sensitivity analyses and experiments.

> more detailed discussion of recent work on MARL in climate-related applications.

We extend the section on MARL in the appendix with some more references on MARL applied to real-world problems such as:

Hou, X., et al. (2025). InvestESG: A multi-agent reinforcement learning benchmark for studying climate investment as a social dilemma. In Proceedings of the Thirteenth International Conference on Learning Representations (ICLR).

May, R., & Huang, P. (2023). A multi-agent reinforcement learning approach for investigating and optimising peer-to-peer prosumer energy markets. Applied Energy, 334, 120705.

Wang, X., et al. (2020). Large-scale traffic signal control using a novel multiagent reinforcement learning. IEEE Transactions on Cybernetics, 51(1), 174–187.

> discussion of the potential policy implications

The policy impacts of climate clubs (and border adjustment mechanisms for that matter) depends on their implementation. Without some degree of redistribution and technology sharing, they risk acting as a tax on carbon locked developing countries [3; 4]. A Loss and damage fund can mitigate that risk [5].

[3] Goldthau, A., & Tagliapietra, S. (2022). How an open climate club can generate carbon dividends for the poor. Bruegel-Blogs.

[4] Perdana, S., & Vielle, M. (2022). Making the EU Carbon Border Adjustment Mechanism acceptable and climate friendly for least developed countries. Energy Policy, 170, 113245.

[5] Boyd, E., et al. (2021). Loss and damage from climate change: A new climate justice agenda. One Earth, 4(10), 1365–1370.

> more detailed discussion of the limitations of the model.

Thank you for this suggestion. We will add a clarifying paragraph to the main text describing key assumptions around binding commitments, region preferences, and power imbalances.