Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Methods and Open Problems in Differentiable Social Choice: Learning Mechanisms, Decisions, and Alignment

Zhiyu An, Wan Du

TL;DR

This paper surveys differentiable social choice as a principled shift from static axiomatic mechanism design to trainable, data-driven institutions embedded in modern ML systems. It surveys six interconnected domains—Differentiable Economics, Neural Social Choice, Participatory Budgeting, Liquid Democracy, AI Alignment as Social Choice, and Inverse Mechanism Learning—showing how classical axioms reappear as inductive biases, loss terms, or audit criteria. The work emphasizes robust optimization, distributional considerations, and the auditability of learned rules, while highlighting 36 open problems that define a new cross-disciplinary research agenda. By reframing incentive compatibility as differentiable surrogates, architecture-level guarantees, and forensics, it argues for a future where normative principles are baked into learnable mechanisms and their governance can be audited across deployments. The practical impact spans auctions, resource allocation, federated learning, and alignment, offering a unified framework for designing, evaluating, and auditing algorithmic decision-making in society.

Abstract

Social choice is no longer a peripheral concern of political theory or economics-it has become a foundational component of modern machine learning systems. From auctions and resource allocation to federated learning, participatory governance, and the alignment of large language models, machine learning pipelines increasingly aggregate heterogeneous preferences, incentives, and judgments into collective decisions. In effect, many contemporary machine learning systems already implement social choice mechanisms, often implicitly and without explicit normative scrutiny. This Review surveys differentiable social choice: an emerging paradigm that formulates voting rules, mechanisms, and aggregation procedures as learnable, differentiable models optimized from data. We synthesize work across auctions, voting, budgeting, liquid democracy, decentralized aggregation, and inverse mechanism learning, showing how classical axioms and impossibility results reappear as objectives, constraints, and optimization trade-offs. We conclude by identifying 36 open problems defining a new research agenda at the intersection of machine learning, economics, and democratic theory.

Methods and Open Problems in Differentiable Social Choice: Learning Mechanisms, Decisions, and Alignment

TL;DR

This paper surveys differentiable social choice as a principled shift from static axiomatic mechanism design to trainable, data-driven institutions embedded in modern ML systems. It surveys six interconnected domains—Differentiable Economics, Neural Social Choice, Participatory Budgeting, Liquid Democracy, AI Alignment as Social Choice, and Inverse Mechanism Learning—showing how classical axioms reappear as inductive biases, loss terms, or audit criteria. The work emphasizes robust optimization, distributional considerations, and the auditability of learned rules, while highlighting 36 open problems that define a new cross-disciplinary research agenda. By reframing incentive compatibility as differentiable surrogates, architecture-level guarantees, and forensics, it argues for a future where normative principles are baked into learnable mechanisms and their governance can be audited across deployments. The practical impact spans auctions, resource allocation, federated learning, and alignment, offering a unified framework for designing, evaluating, and auditing algorithmic decision-making in society.

Abstract

Social choice is no longer a peripheral concern of political theory or economics-it has become a foundational component of modern machine learning systems. From auctions and resource allocation to federated learning, participatory governance, and the alignment of large language models, machine learning pipelines increasingly aggregate heterogeneous preferences, incentives, and judgments into collective decisions. In effect, many contemporary machine learning systems already implement social choice mechanisms, often implicitly and without explicit normative scrutiny. This Review surveys differentiable social choice: an emerging paradigm that formulates voting rules, mechanisms, and aggregation procedures as learnable, differentiable models optimized from data. We synthesize work across auctions, voting, budgeting, liquid democracy, decentralized aggregation, and inverse mechanism learning, showing how classical axioms and impossibility results reappear as objectives, constraints, and optimization trade-offs. We conclude by identifying 36 open problems defining a new research agenda at the intersection of machine learning, economics, and democratic theory.
Paper Structure (191 sections, 4 equations, 1 figure, 4 tables)

This paper contains 191 sections, 4 equations, 1 figure, 4 tables.

Table of Contents

  1. Introduction
  2. Differentiable Economics: From Optimal Auctions to Distributed Edge Markets
  3. From analytical optimal auctions to learned mechanisms
  4. Classical foundations and why multi-dimensionality is hard.
  5. Neural parameterizations: allocation and payment networks.
  6. Recovering known solutions and discovering new ones.
  7. Differentiable incentive constraints: regret, cyclic monotonicity, and certification
  8. Regret minimization as a differentiable surrogate.
  9. RochetNet and structure from cyclic monotonicity.
  10. Certification and verification: beyond empirical audits.
  11. Architectures and inductive biases: symmetry, sets, and attention
  12. Why inductive bias matters in economic learning.
  13. Permutation-equivariant RegretNet variants.
  14. Transformers for combinatorial and joint auctions.
  15. Domain-guided architectures in ad auctions and industrial marketplaces.
  16. ...and 176 more sections

Figures (1)

  • Figure 1: A structural shift from axiomatic social choice (closed-form mechanisms derived from normative axioms) to differentiable social choice (parameterized mechanisms optimized from data), where axioms reappear as architectural biases, loss terms, and audit criteria.