Yan Liu, Feng Zhang, Zhanyu Ma, Jun Xu, Jiuchong Gao, Jinghua Hao, Renqing He, Han Liu, Yangdong Deng. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Weitao Ma, Xiaocheng Feng, Lei Huang, Xiachong Feng, Zhanyu Ma, Jun Xu, Jiuchong Gao, Jinghua Hao, Renqing He, Bing Qin. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Reinforcement Learning with Verifiable Rewards (RLVR) is a key paradigm for improving large-scale reasoning models.Unlike supervised fine-tuning (SFT), RLVR exhibits distinct optimization dynamics and is sensitive to the preservation of pre-trained geometric structures.However, existing parameterefficient methods face key limitations in this regime.Low-rank adaptation methods, such as PiSSA, are primarily designed for Supervised Fine-Tuning (SFT) and do not account for the distinct optimization dynamics and geometric structures of RLVR.Conversely, directly fine-tuning the unstructured sparse parameter subspace favored by RLVR encounters efficiency bottlenecks on modern hardware.To address these challenges, we propose GeoRA (Geometry-Aware Low-Rank Adaptation), a low-rank adaptation method tailored for RLVR.Specifically, GeoRA exploits the anisotropic and compressible structure of RL update subspace, and extracts its principal directions via Singular Value Decomposition (SVD) to initialize low-rank adapters, while freezing residual components as a structural anchor during training.This design preserves the pre-trained structure and enables efficient dense computation.Experiments on Qwen and Llama models from 1.5B to 32B parameters show that GeoRA consistently outperforms strong lowrank baselines across RLVR settings in mathematics, medicine, and coding, while showing stronger generalization and less forgetting on out-of-domain tasks.

Road network generation from massive trajectories has mainly focused on the mining of urban arterial roads while the demand for refined road networks in communities full of low-grade roads is increasing. Trajectory GPS signals in communities tend to be very noisy and hard to label for general modeling. We propose Style-Transfer-Roadnet-Generation(STRG), a method based on un-supervised image style transfer to discover roads from trajectories in these communities. First, we convert the trajectory data into raster images. Then we train a style transferer to transform the raster images into road network style images. At last, the road networks are recovered and refined from predicted images. Experiments on the large-scale data and online deployment results show that STRG can effectively model the texture style of road networks in communities and discover new road networks from trajectories in a low-cost manner, outperforming strong baseline approaches.

A Point-Of-Interest (POI) represents a specific point location that may be useful or interesting for people, and therefore each and every building footprint in a topographic map can be recognized as a POI. Automatic extraction of building footprints using remote sensing images has become a challenging and important research topic which is in demand for urban planning and development. Extensive studies have explored a variety of semantic segmentation methods using deep learning algorithms to achieve better performance in building footprint extraction, however the existing algorithms were shown to have some limitations which lead to poor segmentation results. Building roofs were recognized as building footprints in the previous studies. This is prone to error especially for high-rise buildings due to different sensor view angles. In this paper, we propose a multi-task Res-U-Net model with attention mechanism for the extraction of the building roofs and the whole building shapes from remote sensing images, then use an offset vector method to detect the footprints of the high-rise buildings based on the boundaries of the corresponding building roofs and shapes. We also apply the online food delivery (OFD) data to parse the POI name of every building footprint. Several strategies are also developed in combination with the proposed model, including data augmentation and post-processing. We conduct numerical experiments using real data of remote sensing images and OFD historical order data. Results demonstrate that our proposed model achieves a total F1-score of 77.05% and intersection over union (IoU) of 63.55% in terms of the building roof segmentation, and an overall F1-score of 79.02% and IoU of 66.05% for the whole building shape segmentation, which both achieve the best performance among all baseline models.