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.

Online food delivery (OFD) services, known for offering varied meals at home, have gained global popularity. Meituan has recently ventured into the affordable market segment with its "Pinhaofan'' service, highlighting the imperative to delivery efficiency. To achieve this, delivery scope is regarded as one of the most effective operational tools. The delivery scope of a merchant refers to the geo-graphical area where they can serve customers. Current methods for generating delivery scopes primarily focus on optimizing a single merchant's efficiency or rely on manual delineated from the merchant's perspective, neglecting the merchant substitution effect and potentially resulting in order loss. In this paper, we propose a novel method, named Collaborative Scope, which views the delivery scope as an assortment optimization problem, considering the substitution effect between merchants from the user's perspective. We introduce the discrete choice model of econometrics and use the Enhanced Multinomial Logit Model to predict user conversion rates in the merchant list. Next, we formulate the delivery scope optimization problem of multiple merchants as a mixed integer programming problem. The city-wide solution of this problem, owing to the large-scale combinatorial optimization triggered by high-dimensional decision variables, incurs high computational complexity. To address this, we propose an approximate solution to the original problem through a first-order Taylor series approximation, which significantly reduces the computation complexity at the expense of a slight decrease in solution accuracy. Offline and online A/B test results indicate that, compared to existing methods, Collaborative Scope significantly improves delivery efficiency by reducing delivery difficulty without hurt of order volume. Notably, Collaborative Scope is currently deployed on "Pinhaofan'', serving tens of millions of online users.

The recent past has witnessed a notable surge in on-demand food delivery (OFD) services, offering delivery fulfillment within dozens of minutes after an order is placed. In OFD, pooling multiple orders for simultaneous delivery in real-time order assignment is a pivotal efficiency source, which may in turn extend delivery time. Constructing high-quality order pooling to harmonize platform efficiency with the experiences of consumers and couriers, is crucial to OFD platforms. However, the complexity and real-time nature of order assignment, making extensive calculations impractical, significantly limit the potential for order consolidation. Moreover, offline environment is frequently riddled with unknown factors, posing challenges for the platform's perceptibility and pooling decisions.

On-demand food delivery (OFD) service has gained fast-growing popularity all around the world. Order dispatching is instrumental to large-scale OFD platforms, such as Meituan, which continuously match food order requests to couriers at a scale of tens of millions each day to satisfy the needs of consumers, couriers, and merchants. However, due to high dynamism and inevitable uncertainties in the real-world environment, it is not an easy task to achieve long-term global objective optimization through continuous isolated optimization decisions at each dispatch moment. Our work proposes the concept of "courier occupancy" (CO) to precisely quantify the impact of order assignment on the courier's delivery efficiency, realizing a decomposition of long-term and macro goals into various dispatch moments and micro decision-making dimensions. Then in the prediction phase, an improved and universally applicable distribution estimation method is designed to quantify CO which is a stochastic variable and contains future information, combining Monte Carlo dropout and knowledge distillation. In the optimization phase, we use CO to model the objective function at each dispatch moment to introduce future information and extend dispatch decisions from merely who to assign the order to both when and who to assign it, significantly enhancing the long-term optimization capability of dispatching decisions and avoiding local greed. We conduct extensive offline simulations based on real dispatching data as well as online AB tests through Meituan's platform. Results show that our method consistently improves the couriers' delivery efficiency and consumers' satisfaction.

Pick-up and delivery (PD) services such as online food ordering are playing an increasingly important role in serving people's daily demands. Accurate PD route prediction (PDRP) is important for service providers to efficiently schedule riders to improve service quality. It is crucial to model the decision-making process behind the route choice of riders for PDRP. Recent years have witnessed the success of utilizing imitation learning (IL) to model user decision-making process. Therefore, we propose to deploy an IL framework to solve the PDRP problem. However, there still exist three main challenges: (1) the rider's route decision is affected by multi-source and heterogeneous features and the complex relationships among these features make it hard to explore how they influence the rider's route decision-making; (2) the large route decision-making space make it easy to explore and predict unreasonable routes; (3) the rider's personalized preference is important in modeling the route decision-making process but cannot be fully explored. To tackle the above challenges, we propose ILRoute, a Graph-based imitation learning method for PDRP. ILRoute utilizes a multi-graph neural network (multi-GNN) to extract the multi-source and heterogeneous features and model their complex relationships. To address the large route decision-making space, ILRoute introduces a mobility regularity-aware constraint as prior route choice knowledge to reduce the exploration route decision-making space. To model the personalized preferences of the rider, ILRoute utilizes a personalized constraint mechanism to enhance the personalization of the rider's route decision-making process. Offline experiments conducted on three real-world datasets and online comparisons demonstrate the superiority of our proposed model.

Online food delivery (OFD) services have become popular globally, serving people's daily needs. Precise area-of-interest (AOI) boundaries help OFD platforms determine customers' exact locations, which is crucial for maintaining consistency in delivery difficulty and providing a uniform customer experience within an AOI. Existing AOI generation methods primarily rely on predefined shapes or density-based clustering, which limits the quality of the contours. Recently, Meituan has treated the AOI contours as a binary semantic segmentation problem. Their approach involves a multi-step post-process to address the issues with boundary breaks caused by semantic segmentation models, leading to decreased quality and inefficiency in the learning process. In this paper, we propose a novel method for AOI contour generation called C-AOI (Contour-based Area-of-Interest). C-AOI is an instance segmentation model that focuses on generating high-quality AOI contours. Unlike the former method, which relies on pixel-by-pixel classification, C-AOI starts from the center point of the AOI and regresses the boundary. This approach results in a higher-quality boundary and is less computationally intensive. C-AOI first corrects errors on the contour using a local aggregation mechanism. Then, we propose a novel deforming module called the contour transformer, which captures the global geometry of the object. To enhance the positional relationship among vertices, we introduce a learnable cyclic positional encoding applied to the contour transformer. Finally, to improve the boundary details, we propose the Adaptive Matching Loss (AML) that eliminates over-smoothed boundaries and promotes optimized convergence pathways. Experimental results on real-world datasets collected from Meituan have demonstrated that C-AOI significantly improves the mask and boundary quality compared to Meituan's previous work. Moreover, Its inference speed is comparable to that of E2EC, a state-of-the-art real-time contour-based method. It is noteworthy that C-AOI has been deployed in the Meituan platform for producing AOIs.

Online meal delivery is undergoing explosive growth, as this service is becoming increasingly popular. A meal delivery platform aims to provide excellent and stable services for customers and restaurants. However, in reality, several hundred thousand orders are canceled per day in the Meituan meal delivery platform since they are not accepted by the crowd soucing drivers. The cancellation of the orders is incredibly detrimental to the customer's repurchase rate and the reputation of the Meituan meal delivery platform. To solve this problem, a certain amount of specific funds is provided by Meituan's business managers to encourage the crowdsourcing drivers to accept more orders. To make better use of the funds, in this work, we propose a framework to deal with the multi-stage bonus allocation problem for a meal delivery platform. The objective of this framework is to maximize the number of accepted orders within a limited bonus budget. This framework consists of a semi-black-box acceptance probability model, a Lagrangian dual-based dynamic programming algorithm, and an online allocation algorithm. The semi-black-box acceptance probability model is employed to forecast the relationship between the bonus allocated to order and its acceptance probability, the Lagrangian dual-based dynamic programming algorithm aims to calculate the empirical Lagrangian multiplier for each allocation stage offline based on the historical data set, and the online allocation algorithm uses the results attained in the offline part to calculate a proper delivery bonus for each order. To verify the effectiveness and efficiency of our framework, both offline experiments on a real-world data set and online A/B tests on the Meituan meal delivery platform are conducted. Our results show that using the proposed framework, the total order cancellations can be decreased by more than 25% in reality.

Online food ordering and delivery service has widely served people's daily demands worldwide, e.g., it has reached a number of 34.9 million online orders per day in Q3 of 2020 in Meituan food delivery platform. For the food delivery service, accurate estimation of the driver's delivery route and time, defined as the FD-RTP task, is very significant to customer satisfaction and driver experience. In the paper, we apply deep learning to the FD-RTP task for the first time, and propose a deep network named FDNET. Different from traditional heuristic search algorithms, we predict the probability of each feasible location the driver will visit next, through mining a large amount of food delivery data. Guided by the probabilities, FDNET greatly reduces the search space in delivery route generation, and the calculation times of time prediction. As a result, various kinds of information can be fully utilized in FDNET within the limited computation time. Careful consideration of the factors having effect on the driver's behaviors and introduction of more abundant spatiotemporal information both contribute to the improvements. Offline experiments over the large-scale real-world dataset, and online A/B test demonstrate the effectiveness of our proposed FDNET.

In recent years, the online food ordering (OFO) platforms have arose fast and brought huge convenience to people in daily life. Under the scenario of a realistic OFO platform, this paper addresses an online meal delivery problem (OMDP). To reduce the search space, the OMDP is decomposed into two sub-problems, i.e., the pickup and delivery problem and the order dispatching problem. To solve each sub-problem effectively, a hybrid differential evolution algorithm is proposed, which is fused by the DE-based phase to plan routes and the heuristic-based phase to determine order dispatching schemes. In the DE-based routing phase, a heuristic considering the urgency of orders is designed to generate the initial population with certain quality. Besides, a mutation operator is developed to enhance the exploration and a crossover operator embedded with local search is designed to enhance the exploitation. In the heuristic-based dispatching phase, a regret heuristic is presented to produce good dispatching solutions by introducing the influences between delivery persons. Numerical tests have been carried out and computational results demonstrate the effectiveness of the proposed algorithm.

Route planning serves as the most fundamental part in modern food ordering and delivery system, which is also a classical combinatorial optimization problem. In this paper, we study an online extension of traditional route planning problem where a near-optimal solution has to be generated in a very short time. A simple and effective iterated greedy algorithm is presented along with problem-specific initialization rules, destruction procedure and construction procedure. We also propose a local search method, consisting of two adjustment operators and two neighborhood search operators. With experimental results, we show that our algorithm outperforms the compared evolutionary algorithms and has the capability of providing high-quality solutions within milliseconds.

Considering the uncertainty in the real world online food delivery applications, this paper addresses an online order dispatching problem with fuzzy preparation times (FOODP). According to the characteristics of the problem, the FOODP is decomposed into two sub-problems, an order assignment problem and a fuzzy traveling salesman problem with pickup and delivery. To deal with the two sub-problems efficiently, a two-stage algorithm is proposed by reasonably fusing a modified greedy search (mGS) and the fruit fly optimization algorithm (FOA). In the mGS phase, agreement index is employed in the multi-stage decision to search for robust optimal solutions. In the FOA-based search phase, a modified heuristic is used to generate the initial route. To enhance the exploration of the algorithm, an olfactory search is designed by the cooperation of several problem-specific search operators. Moreover, a specific local intensification is employed to further improve the performance of solutions. Numerical tests and statistical analysis demonstrate the effectiveness and efficiency of the proposed algorithm.

Order dispatching, a key part in real-time food delivery system, has received a lot of attention over the last decade. Under constraints of on-time rate and some other practical constraints, our goal is to maximize the total efficiency of the whole system. As an online algorithm, the running time of the algorithm for order dispatching is limited within milliseconds. At the same time, the problem is highly dynamic and the complexity of the solution space is huge. In this paper, we design two fast heuristics for order dispatching in real-time food delivery. We compare our algorithm with two state-of-the-art algorithms. With numerical results, we show our algorithms are faster than algorithms in the literature and with similar or better solution quality.

This paper tackles a large-scale and uncertain scheduling problem of semiconductor wafer fabrication system. Two kinds of uncertainties are considered here, including the machine breakdowns and the fluctuations of processing times. Instead of using the traditional rigid schedule, we propose a novel concept of operation-group-based soft schedule which is used as the new decision variable of the studied problem. The inherent idea of this concept is to make some critical scheduling decisions at the beginning of the scheduling horizon, and allow the remaining decisions to be made during the execution of the initial soft schedule. Furthermore, an operation-group-based soft scheduling approach (OGSSA) is proposed to deal with the uncertainties, which contains an offline optimization layer and an online dispatching layer. In the offline optimization layer, a prediction-based decomposition scheduling method is proposed to generate a soft schedule, and a global scheduling objective prediction model is constructed to evaluate the soft schedule. Then, the initial soft schedule is released to the online dispatching layer, and an online heuristic rule is designed to decide in real time, which operations and when to process. The computational results on the practical production data demonstrate the effectiveness of OGSSA under uncertain production environments.

In this paper, an uncertain scheduling problem arising from the steelmaking-continuous casting (SCC) production system is investigated. For the practical SCC production system, it is difficult to obtain a schedule with better performance using traditional deterministic scheduling methods since there exists uncertainty in processing times. According to the analysis on characteristics of the uncertain SCC scheduling problem (SCCSP), we construct a soft-form schedule which includes slack ratios as characteristic indexes and the job sequence at the casting stage as key decision variables to cope with the uncertainty in processing times, and propose a two-phase soft optimization method to solve the uncertain SCCSP with the just-in-time and the waiting time objectives under the break probability. In the first phase, the continuous estimation distribution algorithm (EDA) with the ordinal optimization policy is proposed to optimize slack ratios under the chance constraint, in which the optimal computing budget allocation with constrained optimization is applied to reduce the computational burden. In the second phase, based on the above optimized characteristic indexes, the discrete EDA with a local search procedure is proposed to optimize the job sequence at the casting stage. Finally, computational experiments with various scales and noise levels are performed to validate the effectiveness of the proposed algorithm.

Job deterioration and learning co-exist in many realistic scheduling situations. This paper introduces a general scheduling model that considers the effects of position-dependent learning and time-dependent deterioration simultaneously. In the proposed model, the actual processing time of a job depends not only on the total processing time of the jobs already processed but also on its scheduled position. This paper focuses on the single-machine scheduling problems with the objectives of minimizing the makespan, total completion time, total weighted completion time, discounted total weighted completion time, and maximum lateness based on the proposed model, respectively. It shows that they are polynomially solvable and optimal under certain conditions. Additionally, it presents some approximation algorithms based on the optimal schedules for the corresponding single-machine scheduling problems and analyzes their worst case error bound.