AI Paper Atlas

Research map / Large language models / Agents, code and math

USTBench: Benchmarking and Dissecting Spatiotemporal Reasoning Capabilities of LLMs as Urban Agents

Siqi Lai, Yansong Ning, Zirui Yuan, Zhixi Chen, Hao Liu

ICLR 2026 · Large language models › Agents, code and math · approach: Agent instruction-following benchmarks

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use reasoning-post-trained models (R1, QwQ, GLM-Z1) instead of their non-reasoning base models for spatiotemporal understandingLLM urban agents · USTBench spatiotemporal understandingaccuracy: gains of 7-20%Improved
use reasoning model DeepSeek-R1 instead of non-reasoning Llama3.3-70B on temporal trend understandingDeepSeek-R1 · USTBench Trend understandingaccuracy: 0.3203 -> 0.2502Hurt
use reasoning variants (QwQ, DeepSeek-R1) instead of non-reasoning bases on congestion predictionQwQ-32B / DeepSeek-R1 · USTBench congestion forecastingaccuracy: Qwen2.5-32B 0.6900 -> QwQ-32B 0.6290; Llama3.3-70B 0.7493 -> DeepSeek-R1 0.6020Hurt
use reasoning variants instead of non-reasoning bases on planning overallDeepSeek-R1 vs Llama3.3-70B · USTBench planningaccuracy: 0.4503 -> 0.3968Hurt
post-train Qwen2.5-7B on synthetic spatiotemporal understanding data (Qwen2.5-7B-ST)Qwen2.5-7B · USTBench downstream reasoningaccuracyImproved
add reflection mechanism to agent with strong reflection ability (ablation)DeepSeek-R1 agent · urban downstream taskstask performanceImproved
add reflection mechanism to agent with moderate reflection ability (ablation)Qwen2.5-32B agent · urban downstream taskstask performanceNo clear effect
add reflection mechanism to agent with weak reflection ability (ablation)Qwen2.5-7B agent · urban downstream taskstask performanceHurt
distill DeepSeek-R1 into Qwen2.5-7B (R1-Distill-Qwen-7B) instead of Qwen2.5-7BQwen2.5-7B · USTBench spatiotemporal understandingaccuracy: 0.4552 -> 0.2936Hurt
scale model size from Qwen2.5-7B to Qwen2.5-32BQwen2.5 · USTBench spatiotemporal understandingaccuracy: 0.4552 -> 0.6224Improved
use LLM agents instead of classic domain baselines on downstream urban tasksLLM urban agents · Congestion predictionaccuracy (%): 17.18% -> 75.73% (o4-mini)Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy