Reproduce “Generating Fair Consensus Statements with Social Choice on Token-Level MDPs” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Carter Blair, Kate Larson
Researchers from Harvard and the University of Waterloo integrated social choice theory with large language models by modeling consensus generation as a token-level Markov Decision Process (MDP), enabling the creation of fair statements with provable guarantees. Their approach achieved superior egalitarian welfare in single-statement generation, with Beam Search consistently yielding the lowest average Egalitarian Perplexity (2.87), and also demonstrated a method for generating stochastic policies provably within the ex-ante core.
Key insights
- It is possible to derive agent-specific token-level rewards from personalized LLM policies, and these rewards exhibit localized credit assignment, meaning they respond predictably to token-level changes in alignment with agent profiles.
- Maximizing Nash Welfare for stochastic policies provably yields outcomes in the ex-ante core, extending a strong social choice fairness concept to the domain of sequential text generation.
- Token-level search algorithms (Beam Search and Finite Lookahead), when guided by egalitarian welfare objectives, effectively generate single consensus statements that are more agreeable to the least satisfied agents than various baselines.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.