Research map / Multi-agent systems and games / Cooperative multi-agent RL
InvestESG: A multi-agent reinforcement learning benchmark for studying climate investment as a social dilemma
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces InvestESG, an open-source MARL benchmark (PyTorch and JAX) simulating companies and ESG-conscious investors facing an intertemporal climate social dilemma under ESG disclosure mandates, trained with independent PPO.
Problem
Policy effects of ESG disclosure mandates on corporate climate mitigation are hard to test with empirical or traditional theoretical models, which are limited in agent count, horizon, or periods; the paper targets this intertemporal social dilemma between profit-driven companies and investors.
Approach
A multi-agent simulation where companies allocate capital to mitigation, greenwashing, and resilience and investors choose portfolios based on ESG preference; a climate-risk model with Bernoulli climate events; agents trained with independent PPO, with Schelling diagrams used to show the social dilemma structure.
Key results
- With ESG-conscious investors (alpha=10), mitigation spending increases and climate risk falls relative to status quo with mandate, which stays similar to status quo.
- Providing climate-risk information to companies with no investors yields final climate risk of about 0.7, comparable to the alpha=1 scenario.
- Greenwashing was quickly abandoned by IPPO agents in favor of mitigation, so greenwashing did not significantly undermine mitigation under learned behavior.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Disclose ESG scores while investors remain profit-driven | InvestESG with IPPO · Status Quo with Mandate vs. Status Quo | corporate mitigation | No clear effect |
| Increase investor ESG-consciousness | InvestESG with IPPO · Mandate with ESG-conscious investors | final climate risk | Helped |
| Increase investor ESG-consciousness | InvestESG with IPPO · Mandate with ESG-conscious investors | market wealth | Helped |
| Add climate event probabilities and occurrences to observations (ablation) | InvestESG with IPPO · Default 5-company, 3-investor scenario | final climate risk | Helped |
| Add climate event probabilities and occurrences to observations (ablation) | InvestESG with IPPO · No investors | final climate risk: ≈ 0.7 | Helped |
| Allow companies to spend on resilience (ablation) | InvestESG with IPPO · Default 5-company, 3-investor scenario | corporate mitigation spending | Helped |
| Allow companies to spend on resilience (ablation) | InvestESG with IPPO · Default 5-company, 3-investor scenario | final climate risk | No clear effect |
| Allow companies to spend on greenwashing (ablation) | InvestESG with IPPO · Mandate with ESG-conscious investors (α = 1) | corporate mitigation | No clear effect |
| Lock agent decisions for five years (ablation) | InvestESG with IPPO · 5-year lock-in experiments | learning-curve volatility | Hurt |
| Seed company actions with real-world corporate behaviors (ablation) | InvestESG with IPPO · Real-world-seeded experiments | total corporate mitigation | No clear effect |
| Make climate-event damage uncertain (ablation) | InvestESG with IPPO · Uncertain damage vs. default mandate with no ESG-conscious investors | total corporate mitigation | No clear effect |
| Make climate-event damage uncertain (ablation) | InvestESG with IPPO · Uncertain damage vs. default mandate with no ESG-conscious investors | market wealth | Hurt |
Surprising findings
- Greenwashing does not significantly undermine mitigation: although hard-coded greenwashing attracts ESG investors, IPPO agents quickly drop greenwashing as their main strategy.
- Mandatory ESG disclosure alone (with profit-driven investors) yields almost no change in climate risk versus status quo.
Limitations
- Simplifies real-world climate, financial markets, and regulation; a first-principles model not intended to fully represent reality.
- Investors make binary investment choices; companies and investors are limited in heterogeneity and scope.
- Learned results are sensitive to design choices like fixed random seed for climate events and lock-in periods; results mainly directional.
Experimental setup
Benchmarks: InvestESG
Metrics: final climate risk P_100 total market wealth W_100 mitigation spending greenwashing spending number of bankrupt companies
Baselines: Status Quo (profit-only agents) Status Quo with Mandate hard-coded greenwashing vs mitigation policies
Builds on: Independent PPO Stable Baselines 3 Sequential social dilemmas Pastor et al. (2021) ESG equilibrium model DICE/RICE
Closest papers by meaning
- AI for Global Climate Cooperation: Modeling Global Climate Negotiations, Agreements, and Long-Term Cooperation in RICE-N (ICML 2025 · similarity 0.86)
- EcoGym: Evaluating LLMs for Long-Horizon Plan-and-Execute in Interactive Economies (NeurIPS 2026 · similarity 0.85)
- TwinMarket: A Scalable Behavioral and Social Simulation for Financial Markets (NeurIPS 2025 · similarity 0.84)
- FinCon: A Synthesized LLM Multi-Agent System with Conceptual Verbal Reinforcement for Enhanced Financial Decision Making (NeurIPS 2024 · similarity 0.83)
- EconGym: A Scalable AI Testbed with Diverse Economic Tasks (NeurIPS 2025 · similarity 0.83)
- Cooperate or Collapse: Emergence of Sustainable Cooperation in a Society of LLM Agents (NeurIPS 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.