On-demand ride-hailing services like DiDi, Uber, and Lyft have transformed urban transportation, offering unmatched convenience and flexibility. In this paper, we introduce DiMA, an LLM-powered ride-hailing assistant deployed in DiDi Chuxing. Its goal is to provide seamless ride-hailing services and beyond through a natural and efficient conversational interface under dynamic and complex spatiotemporal urban contexts. To achieve this, we propose a spatiotemporal-aware order planning module that leverages external tools for precise spatiotemporal reasoning and progressive order planning. Additionally, we develop a cost-effective dialogue system that integrates multi-type dialog repliers with cost-aware LLM configurations to handle diverse conversation goals and trade-off response quality and latency. Furthermore, we introduce a continual fine-tuning scheme that utilizes real-world interactions and simulated dialogues to align the assistant's behavior with human prefered decision-making processes. Since its deployment in the DiDi application, DiMA has demonstrated exceptional performance, achieving 93% accuracy in order planning and 92% in response generation during real-world interactions. Offline experiments further validate DiMA's capabilities, showing improvements of up to 70.23% in order planning and 321.27% in response generation compared to three state-of-the-art agent frameworks, while reducing latency by 0.72x to 5.47x. These results establish DiMA as an effective, efficient, and intelligent mobile assistant for ride-hailing services.

With the advancement of large language models (LLMs), solving complex tasks (e.g., math problems, code generation, etc.) has garnered increasing attention. Inference-time computation methods (e.g., Best-of-N, MCTS, etc.) are of significant importance, as they have the potential to enhance the reasoning capabilities of LLMs without requiring external training computation. However, due to the inherent challenges of this technique, most existing methods remain proof-of-concept and are not yet sufficiently effective. In this paper, we investigate and benchmark strategies for improving inference-time computation across a wide range of reasoning tasks. Since most current methods rely on a pipeline that first generates candidate solutions (e.g., generating chain-of-thought candidate solutions) and then selects them based on specific reward signals (e.g., RLHF reward, process reward, etc.), our research focuses on strategies for both candidate solution generation (e.g., instructing prompts, hyperparameters: temperature and top-p, etc.) and reward mechanisms (e.g., self-evaluation, reward types, etc.). The experimental results reveal that several previously overlooked strategies can be critical for the success of inference-time computation (e.g., simplifying the temperature can improve general reasoning task performance by up to 5%). Based on extensive experiments (more than 20,000 A100-80G GPU hours with over 1,000 experiments) across a variety of models (e.g., Llama, Qwen, and Mistral families) of various sizes, our proposed strategies outperform the baseline by a substantial margin in most cases, providing a stronger foundation for future research.

Recently, long-thought reasoning models achieve strong performance on complex reasoning tasks, but often incur substantial inference overhead, making efficiency a critical concern. Our empirical analysis reveals that the benefit of using Long-CoT varies across problems: while some problems require elaborate reasoning, others show no improvement—or even degraded accuracy. This motivates adaptive reasoning strategies that tailor reasoning depth to the input. However, prior work primarily reduces redundancy within long reasoning paths, limiting exploration of more efficient strategies beyond the Long-CoT paradigm. To address this, we propose a novel two-stage framework for adaptive and efficient reasoning. First, we construct a hybrid reasoning model by merging long and short CoT models to enable diverse reasoning styles. Second, we apply bi-level preference training to guide the model to select suitable reasoning styles (group-level), and prefer concise and correct reasoning within each style group (instance-level). Experiments demonstrate that our method significantly reduces inference costs compared to other baseline approaches, while maintaining performance. Notably, on five mathematical datasets, the average length of reasoning is reduced by more than 50\%, highlighting the potential of adaptive strategies to optimize reasoning efficiency in large language models.

Harmful fine-tuning attack introduces significant security risks to the fine-tuning services. Main-stream defenses aim to vaccinate the model such that the later harmful fine-tuning attack is less effective. However, our evaluation results show that such defenses are fragile-- with a few fine-tuning steps, the model still can learn the harmful knowledge. To this end, we do further experiment and find that an embarrassingly simple solution-- adding purely random perturbations to the fine-tuned model, can recover the model from harmful behaviors, though it leads to a degradation in the model’s fine-tuning performance. To address the degradation of fine-tuning performance, we further propose \methodname, which optimizes an adaptive perturbation that will be applied to the model after fine-tuning. \methodname maintains model's safety alignment performance without compromising downstream fine-tuning performance. Comprehensive experiments are conducted on different harmful ratios, fine-tuning tasks and mainstream LLMs, where the average harmful scores are reduced by up-to 21.2%, while maintaining fine-tuning performance. As a by-product, we analyze the adaptive perturbation and show that different layers in various LLMs have distinct safety coefficients. Source code available at https://github.com/w-yibo/Panacea.

Recent years have witnessed the prevalent integration of Large Language Models (LLMs) in various Web applications, such as search engines and recommender systems. As an emerging technique, instruction tuning aims to align pre-trained LLMs as capable chatbots that excel at following human instructions. Previous research indicates that selecting an appropriate subset of a large instruction dataset can enhance the capabilities of LLMs and reduce training costs. However, existing works tend to overlook external correlations between instruction examples during data selection process, which can introduce potential bias and lead to sub-optimal performance. To bridge this gap, we formalize this problem from graph influence maximization perspective and propose Uncertainty-aware influence Maximization (UniMax), a data selection framework that explicitly incorporates the complex inter-dependencies within instruction data. Specifically, we first define a latent instruction graph, treating each instruction example as a graph node and representing their implicit relations as graph edges. Instead of solely relying on heuristic metrics for graph construction, we develop a self-supervised graph learner to uncover the latent structure beyond surface-level feature correlations. After that, we propose an uncertainty-aware influence function to score each example on the instruction graph, allowing a simple greedy algorithm to select a valuable subset that embodies both high influence and uncertainty with an approximation guarantee. Extensive experiments on public datasets show that the proposed approach can significantly enhance model capabilities, underscoring the importance of exploiting data dependencies in instruction data selection.

Rapid urbanization has significantly escalated traffic congestion, underscoring the need for advanced congestion prediction services to bolster intelligent transportation systems.As one of the world's largest ride-hailing platforms, DiDi places great emphasis on the accuracy of congestion prediction to enhance the effectiveness and reliability of their real-time services, such as travel time estimation and route planning.Despite numerous efforts have been made on congestion prediction, most of them fall short in handling heterogeneous and dynamic spatio-temporal dependencies (e.g., periodic and non-periodic congestions), particularly in the presence of noisy and incomplete traffic data.In this paper, we introduce a Congestion Prediction Mixture-of-Experts, CP-MoE, to address the above challenges.We first propose a sparsely-gated Mixture of Adaptive Graph Learners (MAGLs) with congestion-aware inductive biases to improve the model capacity for efficiently capturing complex spatio-temporal dependencies in varying traffic scenarios.Then, we devise two specialized experts to help identify stable trends and periodic patterns within the traffic data, respectively.By cascading these experts with MAGLs, CP-MoE delivers congestion predictions in a more robust and interpretable manner.Furthermore, an ordinal regression strategy is adopted to facilitate effective * Corresponding author.

Spatio-Temporal Graph Neural Network (STGNN) has been used as a common workhorse for traffic forecasting. However, most of them require prohibitive quadratic computational complexity to capture long-range spatio-temporal dependencies, thus hindering their applications to long historical sequences on large-scale road networks in the real-world. To this end, in this paper, we propose BigST, a linear complexity spatio-temporal graph neural network, to efficiently exploit long-range spatio-temporal dependencies for large-scale traffic forecasting. Specifically, we first propose a scalable long sequence feature extractor to encode node-wise long-range inputs ( e.g. , thousands of time-steps in the past week) into low-dimensional representations encompassing rich temporal dynamics. The resulting representations can be pre-computed and hence significantly reduce the computational overhead for prediction. Then, we build a linearized global spatial convolution network to adaptively distill time-varying graph structures, which enables fast runtime message passing along spatial dimensions in linear complexity. We empirically evaluate our model on two large-scale real-world traffic datasets. Extensive experiments demonstrate that BigST can scale to road networks with up to one hundred thousand nodes, while significantly improving prediction accuracy and efficiency compared to state-of-the-art traffic forecasting models.

Indoor localization availability is still sporadic in industry, especially at the black-hole, i.e., there only exist cellular signals, no GPS or WiFi signals. Based on our 2-year observations at the DiDi ride-hailing platform in China, there are$ 68\,\text{k}$orders everyday created at black-hole. In this paper, we presentTransparentLoc, a large-scale cellular localization system for pickup position recommendation of the DiDi platform. Specifically, we design a CNN model for real-time localization based on a crowdsourcing fingerprint set constructed by outdoor trajectories and abnormal cell tower detection. Then we leverage a DeepFM model to recommend an optimal pickup position for passengers. We share our 2-year experience with 50 million orders across 13 million devices in 4541 cities to address practical challenges including sparse cell towers, unbalanced user fingerprints, temporal variations, and abnormal cell towers in terms of four major service metrics, i.e., pickup position error, over-30-meters ratio, cancel ratio, and call ratio. The large-scale evaluations show that our system achieves a$ 0.54\,\text{m}$lower median pickup position error compared to the iOS built-in cellular localization system, regardless of environmental changes, smartphone brands/models, time, and cellular providers. Additionally, the over-30-meters ratio, cancel ratio, and call ratio have significant reductions of 0.88%, 0.88%, and 5.13%, respectively.

Point-of-interest (POI) search is important for location-based services, such as navigation and online ride-hailing service. The goal of POI search is to find the most relevant destinations from a large-scale POI database given a text query. To improve the effectiveness and efficiency of POI search, most existing approaches are based on a multi-stage pipeline that consists of an efficiency-oriented retrieval stage and one or more effectiveness-oriented re-rank stages. In this article, we focus on the first efficiency-oriented retrieval stage of the POI search. We first identify the limitations of existing first-stage POI retrieval models in capturing the semantic-geography relationship and modeling the fine-grained geographical context information. Then, we propose a Geo-Enhanced Dense Retrieval framework for POI search to alleviate the above problems. Specifically, the proposed framework leverages the capacity of pre-trained language models (e.g., BERT) and designs a pre-training approach to better model the semantic match between the query prefix and POIs. With the POI collection, we first perform a token-level pre-training task based on a geographical-sensitive masked language prediction and design two retrieval-oriented pre-training tasks that link the address of each POI to its name and geo-location. With the user behavior logs collected from an online POI search system, we design two additional pre-training tasks based on users’ query reformulation behavior and the transitions between POIs. We also utilize a late-interaction network structure to model the fine-grained interactions between the text and geographical context information within an acceptable query latency. Extensive experiments on the real-world datasets collected from the Didichuxing application demonstrate that the proposed framework can achieve superior retrieval performance over existing first-stage POI retrieval methods.

Time-of-arrival estimation or Estimated Time of Arrival (ETA) has become an indispensable building block of modern intelligent transportation systems. While many efforts have been made for time-of-arrival estimation, most of them have scalability and robustness issues when dealing with real-world large-scale ETA scenarios, where billions of vehicle trajectories and ETA requests have been continuously generating every day. To this end, in this paper, we propose a robust and scalable incremental ETA learning framework, iETA, to continuously exploit spatio-temporal traffic patterns from massive floating-car data and thus achieve better estimation performances. Specifically, we first build an incremental travel time predictor that can be incrementally updated based on newly generated traffic data. The incremental travel time predictor not only reduces the overall learning overhead but also improves the model's robustness toward urban traffic distribution shifts. Then, we propose a historical traffic knowledge consolidation module to preserve critical spatio-temporal knowledge from previous ETA predictors under the incremental learning setting. Moreover, to reduce interference induced by low-quality traffic data, we propose an adversarial training module to improve the learning robustness by proactively mitigating and resisting traffic noise perturbations. Finally, extensive experiments demonstrate the effectiveness and efficiency of the proposed system against state-of-the-art baselines in large-scale ETA scenarios. Most importantly, iETA has been deployed on the Didi Chuxing platform, handling real-time billions of ETA queries every day, and substantially improves the prediction accuracy.

With the increasing popularity of location-based services, the point-of-interest (POI) search has received considerable attention in recent years. Existing studies on POI search mostly focus on how to construct better retrieval models to retrieve the relevant POI based on query-POI matching. However, user behavior in POI search, i.e., how users examine the search engine result page (SERP), is mostly underexplored. A good understanding of user behavior is well-recognized as a key to develop effective user models and retrieval models to improve the search quality. Therefore, in this paper, we propose to investigate user behavior in POI search with a lab study in which users' eye movements and their implicit feedback on the SERP are collected. Based on the collected data, we analyze (1) query-level user behavior patterns in POI search, i.e., examination and interactions on SERP; (2) session-level user behavior patterns in POI search, i.e., query reformulation, termination of search, etc. Our work sheds light on user behavior in POI search and could potentially benefit future studies on related research topics.

Travel time estimation is a crucial task in practical transportation applications, while providing the reliability of estimation is important in many working scenarios. Most existing studies do not consider the dynamics of traffic status for different road segments in real time, thus yielding unsatisfactory results. To address the problem, we propose to formulate the traffic network as a temporal attributed graph and perform node representation learning on it. The learned representation is capable of jointly exploiting the dynamic traffic conditions and the topology of the road network, which is then fed into a route-based spatio-temporal dependence learning module to estimate the travel time. By incorporating a distribution loss function, our proposed model is able to predict the distribution of travel time. In the meantime, we design an auxiliary local task of predicting the congestion status of each road segment, which further enhances the generalization performance of the representation learning. Extensive experiments on real-world large-scale datasets demonstrated the superiority of our method compared with the state-of-the-arts.