Sequential Deliberation
Independent Study | Stanford MS&E 408 | Advised by Prof. Ashish Goel | Apr 2026 – Jun 2026
Designed a behavioral experiment to test whether a social choice mechanism with strong theoretical guarantees holds up when real people negotiate a public budget. This is an experiment design; the study has not been run yet.
Context
Large groups struggle to decide together: voting needs every option listed upfront and ignores how strongly people care. Sequential deliberation (Fain, Goel, Munagala & Sakshuwong, 2017) takes a different route. Random pairs of people negotiate one after another, each starting from the previous pair's agreement. It is proven to reach Pareto-efficient outcomes with low distortion (around 1.2) within a few rounds, regardless of group size, but it had never been tested with real people.
Experiment Design
Question
Do real people negotiating in sequence behave the way the theory predicts?
Setting
Participatory budgeting: how should California's General Fund be split across six spending areas? Participants first state their ideal split privately.
Process
In each run, a random participant's ideal split is the starting point. Three random pairs then negotiate in turn over a 7-minute video call with a shared allocation slider, and each agreement becomes the next pair's starting point. 30 runs, up to 180 participants.
Hypotheses
Distortion falls and converges over rounds; agreements land close to the Nash bargaining solution; outcomes are Pareto-efficient; and people anchor on the starting point.
Analysis
Distortion per round (one-sided Wilcoxon signed-rank test), distance to the Nash solution, a Pareto-efficiency check, an anchoring index, and LLM analysis of the negotiation transcripts. A follow-up survey asks whether participants prefer the deliberated outcome over alternatives such as their own ideal, random dictatorship, or a simple average.
Platform
Revived an earlier online deliberation platform (Node.js, Socket.IO) and designed the participant flow and negotiation interface.
Next
AI-assisted deliberation: LLM nudges toward the Nash bargaining point, or an AI mediator in the style of the Habermas Machine, to reduce anchoring and peer pressure between participants.