TL;DR
AGORA is a novel LLM-agent framework that simulates public participation in transit planning to evaluate how different meeting compositions affect final outcomes. The study reveals that while aggregate transit metrics are surprisingly stable regardless of who is "in the room," skewed participant groups significantly increase fairness disparities and tail risks—vulnerabilities that structured deliberation and hard governance gates can successfully suppress.
What Problem Does This Paper Address?
When a city redesigns its bus routes or transit networks, the final plan isn't just a product of an optimization algorithm; it is heavily shaped by public hearings and town halls. However, real-world public participation is notoriously biased, often over-representing older, wealthier, car-owning residents.
In traditional transit planning, this "participation bias" acts as an uncontrolled variable. Because we cannot re-run a real-world public meeting with a different group of people under identical conditions, it has been impossible to answer a fundamental counterfactual question: How much do planning outcomes change based purely on who shows up?
The authors tackle this challenge by introducing AGORA (Agentic Governance for Optimization-Representation Alignment), a framework designed to isolate the effects of meeting composition by simulating structured public deliberation.
Methods
AGORA keeps the underlying transit network, travel demand, and optimization solver fixed while systematically manipulating the mix of participants and the rules of the meeting.
The framework operates in four stages:
- Persona Construction: A fixed pool of synthetic personas is generated, each assigned specific demographic attributes (income, age, commute mode) and transit preferences.
- Profile-Conditioned Sampling: Subsets of these personas are sampled to form a meeting based on specific profiles (e.g., an "Affluent Senior" group vs. a "Representative" group).
- Deliberation Loop: LLM agents (powered by GPT-4o-mini) represent the participants. The framework runs a multi-round propose–evaluate–critique/vote–revise loop. A planning agent proposes a route set, the persona agents provide text critiques and numerical votes, and the planning agent revises the proposal based on this feedback.
- Governance Gates: The final plans must pass three explicit institutional safeguards: a consensus gate (mean vote threshold), a service gate (maximum unserved demand), and a fairness gate (bounds on maximum subgroup disparities in cost and coverage).
The authors use a paired control strategy ($\ell=0$) where no LLM deliberation occurs, and the system simply selects a plan based on the group's initial aggregate preference weights. This isolates the true impact of the text-based deliberation process ($\ell=1$).
Datasets
The framework was evaluated using two standard Transit Network Design Problem (TNDP) benchmark instances with fixed Origin-Destination (OD) demand matrices:
- Mandl (Mandl1): A compact network consisting of 15 nodes, 42 directed links, and 12 routes, serving as a controlled proof-of-concept.
- Mumford0: A larger, more complex network with 30 nodes, 90 links, and 12 routes, used to test scalability and spatial heterogeneity.
The simulated meetings were drawn from a pool of 64 personas across 8 distinct demographic profiles, ranging from perfectly representative to highly skewed.
Key Insights and Ideas
- Deliberation is the Catalyst for Bias: Without the deliberation loop ($\ell=0$), changing the meeting composition produces absolutely zero variation in the final plan. Deliberation is the exact mechanism through which who is in the room actually translates into altered planning outcomes.
- Aggregate Stability Masks Tail Risks: On average, median outcomes (like average passenger cost) vary remarkably little across different skewed meeting profiles. However, looking at the tail risk (90th percentile) tells a different story: heavily skewed meetings consistently produce the worst-case fairness disparities and highest tail costs, whereas representative sampling minimizes these risks.
- Process Design Can Absorb Input Bias: Well-structured institutional design—specifically hard governance gates—can compress the variance caused by biased attendance. This reframes participation bias from an intractable input problem into a manageable process-design problem.
Key Results
- Tail Risk Reduction: On the Mandl network, the "Representative" profile achieved the lowest 90th-percentile passenger cost and cost disparity. Skewed profiles like "Transit-Dependent Residents" produced significantly worse tail outcomes under ungated conditions.
- The Rider-Operator Trade-off: Deliberation ($\ell=1$) systematically rebalanced transit designs toward operational efficiency. On Mandl, it halved operator costs ($C_o$) compared to the non-deliberative control, though at a slight cost to direct, transfer-free service coverage ($d_0$).
- Variance Compression: Governance gates successfully stabilized outcomes on the Mandl network, shrinking cross-profile variance in fairness disparity by an order of magnitude for the most sensitive profiles.
Limitations and Future Work
A key limitation highlighted by the authors is scale sensitivity. While the fixed governance gates worked perfectly to stabilize the smaller Mandl network, the exact same thresholds rejected the majority of runs on the larger Mumford0 network. This demonstrates that fixed policy thresholds do not automatically transfer across problem scales and require instance-specific calibration.
Additionally, because AGORA uses LLMs to simulate human behavior, it is a tool for controlled sensitivity analysis rather than a perfect behavioral replica of human psychology. The authors note that future work should focus on external validation against real-world public hearing records and scaling the framework to large, real-city transit networks where spatial and demand heterogeneity are much more pronounced.
Summary updated