Just for FUNS: LLM-Guided Spatio-Temporal Graph Node Generation for Forecasting Unobserved Node States

Shuhao Li; Weidong Yang; Changan Liu; Wei Zhuo; Yingbo Zhou; Fan Zhang; Siqiang Luo

cs.LGSep 24, 2026

Just for FUNS: LLM-Guided Spatio-Temporal Graph Node Generation for Forecasting Unobserved Node States

Authors: Shuhao Li, Weidong Yang, Changan Liu, Wei Zhuo, Yingbo Zhou, Fan Zhang, Siqiang Luo

Organizations: Fudan University Shanghai, China · Fudan University Shanghai, China Zhuhai Fudan Innovation Research Institute Zhuhai, China · College of Computing and Data Science Nanyang Technological University Singapore · Shanghai Key Laboratory of Data Science Fudan University Shanghai, China · GZHU-SCHB Intelligent Transportation Joint Lab Guangzhou University Guangzhou, China

Abstract

Spatio-temporal forecasting is a cornerstone of logistics, urban planning, and intelligent transportation systems. However, constrained by deployment costs and maintenance resources, sensor networks often lack comprehensive spatial coverage, rendering Forecast Unobserved Node States (FUNS) a critical yet formidable challenge. Conventional models rely on historical observations and typically falter when encountering nodes without prior records. To address this, we redefine the problem as a conditional generation task on spatio-temporal graphs and propose GenST, a framework that introduces Large Language Models (LLMs) as a semantic bridge, leveraging a pre-trained LLM fine-tuned to extract rich semantic features from node descriptions, such as functional zones and road network structures, to compensate for missing spatio-temporal signals. Specifically, we design a two-stage generative architecture: a Spatio-Temporal VAE first compresses spatio-temporal dynamics into a latent space, followed by a Generative Transformer (GenT) that reconstructs the future states of unobserved nodes from noise, guided by multi-modal conditions including semantics, geographic coordinates, and neighborhood contexts. Experiments on six traffic and two non-traffic datasets show GenST significantly outperforms existing baselines in zero-shot prediction tasks, demonstrating the practical potential of semantic-guided generation for mitigating spatio-temporal data sparsity.

Figures & tables

Figure 1 . Illustrative scenarios and spatial heterogeneity of unobserved nodes for FUNS.

Figure 2 . Reformulating FUNS as unobserved node generation based on observed graphs.

Figure 3 . The GenST architecture. The architecture features a unified generative pipeline centered on Stage 2, where latent foundation is unified with multi-modal context alignment to drive the primary engine for direct Stage 3 zero-shot inference.

Models	Datasets	METR-LA			PeMS-Bay			PeMS03			PeMS04			PeMS07			PeMS08
Models	Tasks	MAE	RMSE	MAPE	MAE	RMSE	MAPE	MAE	RMSE	MAPE	MAE	RMSE	MAPE	MAE	RMSE	MAPE	MAE	RMSE	MAPE
K-NN	Global	16.84	26.76	30.38%	13.79	19.05	27.50%	81.28	124.99	91.34%	107.81	141.79	62.36%	107.62	176.30	79.96%	101.17	135.04	148.28%
K-NN	Unobserved	20.90	33.01	35.72%	14.83	20.62	34.78%	147.77	230.74	275.44%	180.46	247.05	136.68%	243.90	323.19	268.02%	194.22	259.33	316.14%
Kriging	Global	15.50	21.09	28.21%	13.21	22.43	28.42%	82.89	124.68	92.26%	99.81	128.27	82.80%	132.57	159.28	141.37%	107.12	141.78	170.59%
Kriging	Unobserved	18.82	31.07	32.41%	14.77	24.03	29.48%	116.74	141.40	210.65%	122.15	157.50	144.07%	165.09	206.04	170.74%	145.50	181.34	297.98%
HA	Global	20.97	24.87	32.56%	12.44	22.43	27.26%	92.78	116.64	214.90%	111.10	138.07	165.64%	133.26	162.86	150.08%	98.42	122.49	123.92%

Table 1 . Main performance comparison. Best Global/Unobserved results are highlighted in bold with dark blue and dark orange colors, respectively; second-best results are indicated with light blue and light orange backgrounds.

Panel of Figure 4 — Figure 4 . Varying unobserved ratios robustness evaluation.

Panel of Figure 5 — Figure 5 . Backend LLM comparison.

Variants	Datasets	METR-LA			PeMS-Bay
Variants	Metrics	MAE	RMSE	MAPE	MAE	RMSE	MAPE
w/o	Global	5.23 ( $\uparrow$ 1.08)	9.30 ( $\uparrow$ 1.15)	12.65% ( $\uparrow$ 2.34%)	3.90 ( $\uparrow$ 1.00)	7.26 ( $\uparrow$ 0.91)	8.20% ( $\uparrow$ 1.06%)
S.Alig.	Unob.	7.68 ( $\uparrow$ 2.11)	13.22 ( $\uparrow$ 3.33)	27.01% ( $\uparrow$ 10.36%)	6.23 ( $\uparrow$ 1.85)	12.20 ( $\uparrow$ 3.31)	21.07% ( $\uparrow$ 8.67%)
w/o	Global	4.73 ( $\uparrow$ 0.58)	8.81 ( $\uparrow$ 0.66)	11.58% ( $\uparrow$ 1.27%)	3.55 ( $\uparrow$ 0.65)	6.87 ( $\uparrow$ 0.52)	7.67% ( $\uparrow$ 0.53%)
ST-VAE	Unob.	6.26 ( $\uparrow$ 0.69)	10.56 ( $\uparrow$ 0.67)	17.94% ( $\uparrow$ 1.29%)	5.63 ( $\uparrow$ 1.25)	9.52 ( $\uparrow$ 0.63)	13.29% ( $\uparrow$ 0.89%)
w/o	Global	4.61 ( $\uparrow$ 0.46)	8.60 ( $\uparrow$ 0.45)	11.03% ( $\uparrow$ 0.72%)	3.40 ( $\uparrow$ 0.50)	6.66 ( $\uparrow$ 0.31)	7.54% ( $\uparrow$ 0.40%)

Table 2 . Components ablation study.

Dataset	Backend	Training		Sampler	Inference
		T. Time	T.Mem. (Peak)		In. Time	In.Mem. (Peak)	Unobs. MAE
METR-LA	LLaMA3.1-8B	6.72s	27.59G	DDPM (K=1000)	2.15s	17.49G	5.57
				DDIM (K=50)	0.73s	17.49G	6.33
				DPM-Solver++ (K=20)	0.53s	17.49G	6.27
	Qwen2.5-1.5B	1.69s	6.92G	DDPM (K=1000)	0.49s	3.20G	6.41
PeMS04	LLaMA3.1-8B	9.93s	34.09G	DDPM (K=1000)	3.17s	19.12G	27.40

Table 3 . Training and inference cost.

Models		K-NN		GWNet		AGCRN		ST-LLM+		RoadDiff		GenST
Datasets	Tasks	Global	Unob.	Global	Unob.	Global	Unob.	Global	Unob.	Global	Unob.	Global	Unob.
KownAir	MAE	28.07	47.99	20.72	42.16	22.42	45.03	20.25	39.22	20.84	41.26	18.56	19.78
	RMSE	43.13	68.86	32.63	53.20	34.85	56.18	32.49	51.17	33.14	52.49	25.32	30.14
	MAPE	34.32%	79.92%	29.21%	69.38%	32.20%	76.17%	28.70%	64.33%	29.54%	67.64%	18.82%	22.53%
Electricity	MAE	3.13	5.68	2.58	5.78	2.72	6.41	2.11	4.66	2.26	5.12	1.87	2.18
	RMSE	25.68	38.91	22.04	42.13	22.47	42.63	17.50	31.79	18.76	34.64	15.72	19.37

Table 4 . Cross-domain generalization.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Metric	Full	Generic template			Shuffled descriptions
Dataset	Metric	Semantic	20%	50%	80%	20%	50%	80%
	MAE	5.57	6.13	6.61	7.24	6.38	6.92	7.72
METR-LA	RMSE	9.89	10.76	11.5	12.58	11.08	11.93	13.25
	MAPE	16.65%	18.28%	19.83%	21.74%	19.04%	20.66%	23.18%
	MAE	21.29	23.85	24.62	26.31	24.48	25.82	27.24
PeMS03	RMSE	32.39	36.05	37.86	41.08	36.99	38.11	40.64

Appendix

Table 5 . Results of semantic degradation robustness.

Figure 6 . Performance sensitivity analysis across varying diffusion iteration counts.

Dataset	Seed 1	Seed 2	Seed 3	Seed 4	Seed 5	Mean±Std
METR-LA	5.57	5.42	5.71	5.63	5.49	5.56 ± 0.11
PeMS04	27.4	27.12	27.78	27.95	27.31	27.51 ± 0.34
PeMS03	21.29	20.94	21.66	21.48	21.12	21.30 ± 0.28
PeMS08	19.76	19.42	20.08	19.89	19.60	19.75 ± 0.25

Appendix

Table 6 . Stability under different random seeds.

Panel of Figure 7 — Figure 7 . Performance across varying forecasting horizons.

Panel of Figure 8 — Figure 8 . Supplemental robustness evaluation under varying unobserved node ratios.

Panel of Figure 9 — Figure 9 . Supplemental backend LLM comparison across additional real-world traffic datasets.

Variants	Datasets	PeMS07
Variants	Metrics	MAE	RMSE	MAPE
w/o Semantic	Global	40.05 ( $\uparrow$ 4.08)	63.54 ( $\uparrow$ 8.63)	20.10% ( $\uparrow$ 2.77%)
Alignment	Unobserved	43.83 ( $\uparrow$ 11.29)	66.86 ( $\uparrow$ 19.07)	38.24% ( $\uparrow$ 17.74%)
w/o	Global	37.39 ( $\uparrow$ 1.42)	59.61 ( $\uparrow$ 4.70)	18.49% ( $\uparrow$ 1.16%)
ST-VAE	Unobserved	33.54 ( $\uparrow$ 1.00)	51.08 ( $\uparrow$ 3.29)	21.82% ( $\uparrow$ 1.32%)
w/o Cross-	Global	38.10 ( $\uparrow$ 2.13)	59.60 ( $\uparrow$ 4.69)	17.99% ( $\uparrow$ 0.66%)

Appendix

Table 7 . PeMS07 and PeMS08 ablation results.

Panel of Figure 10 — Figure 10 . Per-node MAE distribution on PeMS03, dashed lines indicate subset means.

Dataset	METR-LA	PeMS-Bay	PeMS03	PeMS04
Start Time	2012/3/1	2017/1/1	2018/9/1	2018/1/1
End Time	2012/6/30	2017/5/31	2018/11/30	2018/2/28
Time Interval	5 min	5 min	5 min	5 min
# Object	207	325	358	307
# Timestamps	34272	52116	26208	16992
Dataset	PeMS07	PeMS08	KnowAir	Electricity

Appendix

Table 8 . Statistics of all datasets.

Explore similar work

U-STS-LLM A Unified Spatio-Temporal Steered Large Language Model for Traffic Prediction and Imputation
May 12, 2026Yichen Zhang, Jun LiLong-Term Time Series Forecasting Time Series Imputation
School of Information Science and Engineering, Southeast University, Nanjing, 210096, CHINA.
LLMODE: Aligning ODEs with LLMs via Gated Token Injection for Irregular Spatio-Temporal Forecasting
Aug 30, 2026Di Zhang, Jingyang Zhang, Ziqian Wang +4Irregular Time-Series Modeling Zero-Shot Time Series Forecasting
Beihang University Beijing, China
Addressing Spatial Indistinguishability in Spatiotemporal Prediction via Optimal Transport-Guided Masking
Sep 28, 2026Guangyu Wang, Jiawei TongSelf-Supervised Time Series Representation Learning Spatiotemporal Forecasting
Data Science and Artificial Intelligence, Dongbei University of Finance and Economics, Dalian, Liaoning, China · Institute of Systems, Molecular & Integrative Biology, University of Liverpool, Liverpool, United Kingdom

May 12, 2026, cs.LGJ/K move · Enter open · S save

U-STS-LLM A Unified Spatio-Temporal Steered Large Language Model for Traffic Prediction and Imputation

Yichen Zhang, Jun Li

School of Information Science and Engineering, Southeast University, Nanjing, 210096, CHINA.

Long-Term Time Series Forecasting Time Series Imputation LLM-Based Time Series Forecasting Wireless Communications Spatiotemporal Forecasting

Open paper AskPDF arXiv

The efficient operation of modern cellular networks hinges on the accurate analysis of spatio-temporal traffic data. Mastering these patterns is essential for core network functions, chiefly forecasting future load to pre-empt congestion and imputing missing values caused by sensor failures or transmission errors to ensure data continuity. While deeply connected, forecasting and imputation have historically evolved as separate sub-fields. The dominant paradigm, Spatio-Temporal Graph Neural Networks (STGNNs), while effective, are often specialized, computationally intensive, and exhibit limited generalization. Concurrently, adapting large pre-trained language models (LLMs) offers a powerful alternative for sequence modeling, yet existing approaches provide weak structural guidance, leading to unstable convergence and a narrow focus on forecasting. To bridge these gaps, we propose U-STS-LLM, a unified framework built on a spatio-temporally steered LLM. Our core innovation is a Dynamic Spatio-Temporal Attention Bias Generator that synthesizes a persistent functional graph with transient nodal states to explicitly steer the LLM's attention. Coupled with a partially frozen backbone tuned via Low-Rank Adaptation (LoRA) and a Gated Adaptive Fusion mechanism, the model achieves stable, parameter-efficient adaptation. Trained under a unified multi-task objective, U-STS-LLM learns a holistic data representation. Extensive experiments on real-world cellular datasets demonstrate that U-STS-LLM establishes new state-of-the-art performance in both long-horizon forecasting and high-missing-rate imputation, while maintaining remarkable training efficiency and stability, offering a novel blueprint for harnessing foundation models in structured, non-linguistic domains.

Abstract

Explore similar work

U-STS-LLM A Unified Spatio-Temporal Steered Large Language Model for Traffic Prediction and Imputation

LLMODE: Aligning ODEs with LLMs via Gated Token Injection for Irregular Spatio-Temporal Forecasting

Addressing Spatial Indistinguishability in Spatiotemporal Prediction via Optimal Transport-Guided Masking