Research map / Multi-agent systems and games / Games and equilibria
Operationalising Rawlsian Ethics for Fairness in Norm Learning Agents
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.
Proposes RAWL·E, DQN-based norm-learning agents that add a maximin (Rawlsian) ethics module providing self-directed sanctions, and shows in two simulated berry-harvesting environments that resulting norms yield higher minimum well-being, lower inequality, higher social welfare and slightly longer survival than baseline DQN societies.
Problem
Norms learned by self-interested agents can emerge that subjugate some agents; learning from existing behaviour risks propagating unethical norms (is-ought gap), so agents need a normative basis for fair norm emergence.
Approach
Operationalise Rawls' maximin principle as an ethics module that emits a self-directed sanction (reward shaping) according to whether an action improves the minimum well-being, so that the norm-learning module records and promotes fair cooperative behaviours.
Key results
- In the allotment harvest, RAWL·E societies raise minimum well-being from 7.18 to 10.82 (Cohen's d 3.09, p<0.01).
- Gini index of well-being falls from 0.2 (baseline) to 0.1 (RAWL·E) in allotment harvest (d=1.58).
- Robustness (survival length) rises from 47.36 to 48.19 steps, but the effect is negligible (d=0.11).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add a Rawlsian maximin ethics module | DQN society · Capabilities Harvest | Gini index (well-being inequality): 0.12 -> 0.07 | Helped |
| Add a Rawlsian maximin ethics module | DQN society · Capabilities Harvest | Gini index (resource inequality): 0.09 -> 0.04 | Helped |
| Add a Rawlsian maximin ethics module | DQN society · Capabilities Harvest | Minimum experience (well-being): 5.81 -> 7.69 | Helped |
| Add a Rawlsian maximin ethics module | DQN society · Capabilities Harvest | Minimum experience (resources): 3.62 -> 3.98 | No clear effect |
| Add a Rawlsian maximin ethics module | DQN society · Capabilities Harvest | Social welfare: 35.25 -> 37.05; 16.52 -> 16.96 | Mixed |
| Add a Rawlsian maximin ethics module | DQN society · Capabilities Harvest | Episode length: 38.07 -> 40.5 | Helped |
| Add a Rawlsian maximin ethics module | DQN society · Allotment Harvest | Gini index (well-being inequality): 0.2 -> 0.1 | Helped |
| Add a Rawlsian maximin ethics module | DQN society · Allotment Harvest | Gini index (resource inequality): 0.14 -> 0.06 | Helped |
| Add a Rawlsian maximin ethics module | DQN society · Allotment Harvest | Minimum experience (well-being): 7.18 -> 10.82 | Helped |
| Add a Rawlsian maximin ethics module | DQN society · Allotment Harvest | Minimum experience (resources): 3.79 -> 4.5 | No clear effect |
| Add a Rawlsian maximin ethics module | DQN society · Allotment Harvest | Social welfare: 51.5 -> 59.8; 18.94 -> 20.6 | Mixed |
| Add a Rawlsian maximin ethics module | DQN society · Allotment Harvest | Episode length: 47.36 -> 48.19 | Helped |
Surprising findings
- Minimum resource-based experience showed no significant difference between societies despite significant well-being improvements.
- RAWL·E agents received lower raw environmental rewards than baseline agents by design, yet societies still performed better on fairness metrics.
Limitations
- Evaluated only on two simplified grid-world harvesting scenarios, limiting real-world applicability.
- Robustness improvement is negligible in effect size; many metrics not significant for resource-based variables.
- Only a single ethical principle (maximin) is operationalised and only compared to a non-ethical baseline.
Experimental setup
Datasets: simulated capabilities harvest environment simulated allotment harvest environment
Benchmarks: capabilities harvest simulation allotment harvest simulation
Metrics: Gini index inequality minimum experience social welfare robustness (episode length) Mann-Whitney U p-value Cohen's d
Baselines: Baseline Cooperative DQN
Builds on: DQN Rawlsian ethics (maximin) Leben normative ethics utility function self-directed sanctions
Closest papers by meaning
- Moral Alignment for LLM Agents (ICLR 2025 · similarity 0.86)
- Many LLMs Are More Utilitarian Than One (NeurIPS 2025 · similarity 0.85)
- MoralReason: Generalizable Moral Decision Alignment for LLM Agents Using Reasoning-Level Reinforcement Learning (AAAI 2026 · similarity 0.84)
- Evolving Interpretable Constitutions for Multi-Agent Coordination (ICML 2026 · similarity 0.84)
- Cooperate or Collapse: Emergence of Sustainable Cooperation in a Society of LLM Agents (NeurIPS 2024 · similarity 0.84)
- Investigating Prosocial Behavior Theory in LLM Agents Under Policy-Induced Inequities (AAAI 2026 · 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.