Trajectory Generation Using Negative World Model
Author：Mengkun Gao
Affiliation：Graduate School of Information Science and Technology, The University of Osaka
Abstract：This study trains an LLM using both negative and positive check-in transitions to better capture nuances in user behavior. First, Llama-3.2-3B is LoRA fine-tuned on negative samples for next-POI and time prediction. It is then further fine-tuned on positive samples to infer POI transition motivations. The final model generates full trajectories for a given date. Experimental results show that mixing positive and negative samples improves performance, with the best positive sample ratio being 0.6.
Affiliation：Graduate School of Information Science and Technology, The University of Osaka
Abstract：This study trains an LLM using both negative and positive check-in transitions to better capture nuances in user behavior. First, Llama-3.2-3B is LoRA fine-tuned on negative samples for next-POI and time prediction. It is then further fine-tuned on positive samples to infer POI transition motivations. The final model generates full trajectories for a given date. Experimental results show that mixing positive and negative samples improves performance, with the best positive sample ratio being 0.6.
Posted : March 31,2026
