Multi-Agent Collaborative Reasoning with Tool-Augmented Evidence for Urban Region Profiling¶
🕒 Published (v1): 2026-07-15 08:01 UTC · Source: Arxiv · Venue: KDD 2026 · link
Ask a follow-up
Open an assistant pre-loaded with this paper's context.
TL;DR¶
UrbanAgent reframes urban region profiling (predicting socioeconomic indicators like GDP, carbon emissions, and population from multimodal urban data) as a reasoning-driven inference problem rather than a latent-embedding fusion task. It deploys one LLM agent per data modality, trains each via SFT cold-start followed by GRPO-based reinforcement learning to actively call external tools, then connects agents through a typed interaction graph (corroboration, rectification, supervision edges) for iterative conflict resolution. On global urban datasets it achieves an average 8.1% improvement in \(R^2\) over prior baselines.
Problem¶
Existing multimodal urban region profiling methods fuse heterogeneous signals (satellite imagery, POI records, 3D building footprints, text) into a single latent representation via static alignment pipelines. This design (1) silently absorbs cross-modal inconsistencies rather than resolving them, and (2) applies correlations learned during training uniformly to unseen regions, producing confident but spurious predictions when observable cues (building height, nightlight intensity) are poor proxies for the target indicator in a given context.
Method¶
Two-stage framework:
Stage 1 — Tool-Augmented Agentic Learning. Each modality (satellite, text, POI, 3D buildings) has a dedicated Qwen3-VL-7B agent. A trajectory curation pipeline annotates tool-use sequences (Web Search, Crop-and-Zoom, Nightlight Retrieval, Historical Satellite Retrieval) with full reasoning chains. Agents are first SFT cold-started on these trajectories to inject ReAct-style tool-call priors, then optimized with GRPO using a composite reward:
where \(R_\text{outcome}\) measures numeric accuracy, \(R_\text{format}\) enforces output grammar, and \(R_\text{process}\) rewards structured tool use and penalizes degenerate invocations. GRPO normalizes rewards within a sampled group of \(G\) trajectories into group-relative advantages and updates the policy with PPO-style clipping plus KL regularization against an SFT reference policy.
Stage 2 — Multi-Agent Collaborative Reasoning. Agents form a heterogeneous interaction graph \(\mathcal{G}=(\mathcal{V},\mathcal{E})\) with three typed edge sets determined by prediction distance and confidence difference: - Corroboration \((v_j \to v_i) \in \mathcal{E}^\text{cor}\) if \(|\text{pred}_i - \text{pred}_j| < \tau_\text{cor}\) - Rectification \((v_j \to v_i) \in \mathcal{E}^\text{rec}\) if \(|\text{pred}_i - \text{pred}_j| > \tau_\text{rec}\) - Supervisory \((v_j \to v_i) \in \mathcal{E}^\text{sup}\) if \(c_j - c_i > \tau_\text{sup}\)
Agents run \(L\) rounds of relation-aware message passing: corroboration messages gate on confidence differences, rectification messages propagate hidden-state deltas \(\text{ReLU}(h_j^{(l)} - h_i^{(l)})\), and supervisory messages use confidence-weighted attention. States are updated via GRU, and a Final Judge aggregates outputs.
Key Contributions¶
- UrbanAgent framework: reframes multimodal urban profiling as structured multi-agent semantic reasoning, replacing latent fusion with explicit inter-agent interaction.
- Tool-augmented agentic learning pipeline: SFT cold-start + GRPO RL with a three-component hierarchical reward that incentivizes evidence-grounded, well-structured tool use.
- Typed multi-agent interaction graph: three semantically distinct edge types (corroboration, rectification, supervision) enable explicit conflict resolution rather than implicit absorption.
- Global benchmark evaluation: in-domain and unseen-city generalization results on Carbon emissions, GDP, and Population prediction across multiple world regions.
Results¶
- Average \(R^2\) improvement of 8.1% over existing DL and multi-agent baselines across Carbon, GDP, and Population tasks on global urban datasets.
- Strong generalization in unseen-city settings, demonstrating robustness beyond in-distribution evaluation.
- Ablations (implied by paper structure) validate both the tool-augmented agentic learning stage and the multi-agent collaborative reasoning stage individually.
(Note: detailed per-baseline numeric tables are not included in the provided text excerpt.)
Limitations¶
- Trajectory curation relies on Qwen3-VL-7B as annotation model; annotation quality caps agent performance.
- Tool suite is fixed to four categories (Web Search, Crop-and-Zoom, Nightlight, Historical Satellite); domain shifts requiring other evidence types are not covered.
- Thresholds \(\tau_\text{cor}\), \(\tau_\text{rec}\), \(\tau_\text{sup}\) governing graph topology are manually set and may require re-tuning for new cities or indicator types.
- The paper text provided is truncated before the full results tables; complete ablation and baseline comparison numbers are not available for review.
- Computational cost of running four separate MLLM agents plus iterative graph reasoning per region is not analyzed in the provided excerpt.
Relevance to Harnesses / Meta-Harnesses¶
UrbanAgent is a concrete instantiation of a typed multi-agent harness: each agent owns a modality, is trained end-to-end (SFT → GRPO) to invoke tools, and communicates through structured edge types that encode epistemic roles (corroborate, rectify, supervise) rather than raw message passing. That typed interaction graph is essentially a harness schema — it constrains which agents can override which, making the coordination layer inspectable and compositional in a way flat orchestration loops are not. For someone building meta-harnesses, the key transferable idea is using RL (GRPO) to train agents to decide when to call tools and how to resolve inter-agent conflicts, rather than hard-coding those policies in the orchestrator. The SFT cold-start → RL fine-tune recipe is also directly reusable for bootstrapping specialist agents in any domain-specific harness where ground-truth outputs exist.