Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Building Socially-Equitable Public Models
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 an Equitable Objective (alpha-fairness-inspired, parameter q) and policy-gradient training so a shared public model yields more uniform decision-cost regret across heterogeneous downstream agents.
Problem
Public models trained only for prediction accuracy (MSE) can produce large performance disparities across downstream agents with different decision objectives and data availability, leading to unfair service.
Approach
Train the public model on an alpha-fairness-style Equitable Objective that sums powered (q+1) average decision regrets per agent, so worse-off agents get higher weight; use policy gradient to handle non-differentiable downstream costs, with theory on equity via variance and entropy.
Key results
- On data center 'different agents, different lambda', Equitable PM (q+1=10) reduced cost-regret variance from 0.0602 (Plain PM) to 0.0055 and mean from 0.4779 to 0.1444.
- Equitable PM lowered cost-regret variance and C95-C5 relative to Plain PM across all three data center settings, with both decreasing as q increases.
- Plain PM retained lower MSE on carbon prediction (e.g., 4.2013 vs 4.3819 at q+1=10), a fairness-accuracy trade-off the authors acknowledge.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Equitable PM instead of Plain PM | LSTM public model · Data centers: similar agents, different λ | Variance: 0.0085 -> 0.0002 | Helped |
| Use Equitable PM instead of Plain PM | LSTM public model · Data centers: similar agents, different λ | Mean cost regret: 0.2732 -> 0.0465 | Helped |
| Use Equitable PM instead of Plain PM | LSTM public model · Data centers: similar agents, different λ | MSE: 4.2054 -> 4.2194 | Hurt |
| Use Equitable PM instead of Plain PM | LSTM public model · Data centers: different agents, same λ | Variance: 0.0009 -> 1.71e-5 | Helped |
| Use Equitable PM instead of Plain PM | LSTM public model · Data centers: different agents, same λ | Mean cost regret: 0.0988 -> 0.0136 | Helped |
| Use Equitable PM instead of Plain PM | LSTM public model · Data centers: different agents, different λ | Variance: 0.0602 -> 0.0055 | Helped |
| Use Equitable PM instead of Plain PM | LSTM public model · Data centers: different agents, different λ | MSE: 4.2013 -> 4.3819 | Hurt |
| Use Equitable PM instead of Plain PM | Transformer public model · Diverse downstream objectives | Variance: 18.89 -> 15.72 | Helped |
| Use Equitable PM instead of Plain PM | Transformer public model · Diverse downstream objectives | MSE: 7.20 -> 9.67 | Hurt |
| Increase q+1 from 1 to 1.5 (ablation) | Equitable PM · Data centers: similar agents, different λ | Variance: 0.0029 -> 0.0002 | Helped |
| Increase q+1 from 1 to 20 (ablation) | Equitable PM · Data centers: different agents, same λ | Variance: 0.0008 -> 1.71e-5 | Helped |
| Increase q+1 from 1 to 10 (ablation) | Equitable PM · Data centers: different agents, different λ | Variance: 0.0181 -> 0.0055 | Helped |
Surprising findings
- Equitable PM achieves lower mean regret than Plain PM despite not minimizing prediction MSE.
Limitations
- Requires access to downstream agents' decision costs, raising privacy and security concerns.
- Case-study models are small (LSTM, single-layer Transformer) due to limited compute.
- Trade-off with prediction accuracy: MSE is slightly worse than Plain PM.
Experimental setup
Datasets: state-level energy fuel mix dataset Azure cloud workload dataset ACN-Data CAISO electricity prices NHTS
Benchmarks: Azure workload data center case study ACN-Data EV charging case study
Metrics: variance of cost regret mean cost regret C95-C5 percentile of cost regret MSE
Baselines: Plain PM
Builds on: alpha-fairness decision-focused learning predict-then-optimize policy gradient
Closest papers by meaning
- Addressing Polarization and Unfairness in Performative Prediction (AAAI 2026 · similarity 0.85)
- Partial Fairness Awareness: Belief-Guided Strategic Mechanism for Strategic Agents (AAAI 2026 · similarity 0.85)
- FairDICE: Fairness-Driven Offline Multi-Objective Reinforcement Learning (NeurIPS 2025 · similarity 0.84)
- Fair Participation via Sequential Policies (AAAI 2024 · similarity 0.83)
- Intersectional Fairness in Reinforcement Learning with Large State and Constraint Spaces (ICML 2025 · similarity 0.83)
- Some Optimizers are More Equal: Understanding the Role of Optimizers in Group Fairness (NeurIPS 2025 · 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.