Skip to content

Language-Grounded Multi-Agent Planning for Personalized and Fair Participatory Urban Sensing

🕒 Published (v1): 2026-03-25 07:19 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

MAPUS is an LLM-based multi-agent framework for participatory urban sensing (PUS) that replaces centralized combinatorial optimization with a coordinator-plus-participant-agents architecture. Participant agents reason over individual profiles and urban attributes to generate personalized routes, while a coordinator agent enforces fairness and mediates conflict-reduction negotiations. On two real-world taxi/ride-hailing datasets, MAPUS achieves competitive sensing coverage and 20–30% higher path satisfaction than classical baselines.

Problem

Existing PUS methods model recruitment and routing as centralized coverage-maximization problems and treat participants as homogeneous, ignoring heterogeneous preferences (land-use affinity, risk aversion, time sensitivity) and long-term workload fairness. Soft, semantic participant constraints cannot be encoded cleanly as numerical optimization variables, and greedy coverage maximization repeatedly burdens the same participants.

Method

MAPUS operates in three sequential stages, all orchestrated via LangGraph with GPT-4.1-mini-2025-04-14:

  1. Preference-Aware Route Generation (PRG): Each participant agent (PA) receives the task description and decides to accept or reject. Accepted PAs iteratively refine a classically computed feasible baseline route using LLM reasoning over their profile and grid-level land-use/crime attributes. Route quality is measured by the Path Satisfaction Score (PSS): \(S(r_u) = \text{sim}(p_u, h(r_u)) - \mu\,\text{risk}(r_u)\), where \(h(r_u)\) is the empirical land-use distribution along the route and \(\text{sim}(\cdot,\cdot)\) is cosine similarity.

  2. Fairness-Aware Participant Selection (FPS): The coordinator greedily selects participants under budget \(B\) using a composite score \(J(u) = \beta\,\widetilde{\Delta\phi}(u) + (1-\beta)\,\tilde{F}(u)\), where \(\Delta\phi(u)\) is the marginal coverage gain and \(F(u) = 1/(1+h_u)\) down-weights historically over-recruited participants. An LLM tie-breaker handles near-equal scores.

  3. Negotiation-Based Route Refinement (NRR): The coordinator repeatedly finds the highest-overlap route pair using Jaccard overlap \(\text{Overlap}(r_u, r_v)\), proposes overlap-reducing revisions, and both PAs consent or reject; the process iterates until agreement or falls through to the next pair.

Key Contributions

  • Language-grounded reformulation of PUS as decentralized coordination among heterogeneous LLM agents with individual profiles and schedules.
  • PSS metric that measures preference alignment via land-use distribution cosine similarity plus a risk penalty.
  • Composite fairness-coverage selection score that balances marginal sensing utility against historical participation counts.
  • Consent-driven negotiation protocol for route overlap reduction that preserves participant autonomy.

Results

  • T-Drive (PSS): MAPUS achieves PSS of 0.875/0.875/0.898 (Small/Medium/Large) vs. best classical baseline (GraphDP) at 0.659/0.651/0.628 — roughly 20–30% improvement across scales.
  • T-Drive (Coverage): MAPUS leads at Medium (5.115 vs. GraphDP 5.102) and Large (5.251 vs. 5.214); GraphDP slightly ahead at Small (4.761 vs. 4.677).
  • Grab-Posisi (PSS): MAPUS at 0.854/0.834/0.831 vs. MSAGI 0.752/0.714/0.653 — consistent advantage, smaller margin than T-Drive due to fewer urban attribute signals.
  • Grab-Posisi (Coverage): MAPUS best at Large (5.701 vs. TCPG 5.587); competitive at Small and Medium.
  • vs. single-LLM baseline: MAPUS uniformly outperforms GPT-4.1-mini used monolithically on both coverage and PSS, with substantially lower variance.
  • NRR ablation (T-Drive Medium): Negotiation reduces pairwise route overlap from 3.482% to 1.267%; removing NRR causes the largest coverage drop among all ablations.
  • FPS ablation: Removing fairness selection slightly increases coverage but reduces PSS; removing PRG collapses PSS while coverage is near-competitive with classical planners.

Limitations

  • LLM-based agents introduce higher variance than deterministic planners; MAPUS has larger variance than all classical baselines (though smaller than single-LLM).
  • Participant profiles (land-use preferences, economic status) are synthetically generated from heuristics rather than measured from real participants.
  • PSS computation in Grab-Posisi uses only land-use attributes (no crime data), limiting preference signal diversity.
  • The negotiation stage adds sequential LLM call overhead whose wall-clock cost is not reported, potentially limiting real-time scalability.
  • Budget and grid discretization are fixed per experiment; sensitivity to these hyperparameters is not systematically evaluated.

Relevance to Harnesses / Meta-Harnesses

MAPUS is a domain-specific multi-agent harness: a coordinator agent orchestrates a typed pipeline of sub-agents (dispatch → selection → negotiation refinement) using LangGraph as the underlying orchestration layer, making it a concrete exemplar of a three-stage agentic harness with structured inter-agent communication protocols. The consent-driven negotiation loop — where the coordinator iteratively proposes, collects structured accept/reject feedback, and either commits or escalates — is a micro-harness pattern applicable beyond urban sensing to any multi-agent negotiation workflow. The ablation results directly validate the value of each harness stage in isolation, providing empirical evidence for the modular-pipeline design principle that underlies most meta-harness architectures. The comparison of monolithic single-LLM vs. decomposed multi-agent orchestration further quantifies the benefit of harness-style task decomposition over flat prompting.