Reassessing the Role of Supervised Fine-Tuning: An Empirical Study in VLM Reasoning

Recent advances in vision-language models (VLMs) reasoning have been largely attributed to the rise of reinforcement Learning (RL), which has shifted the community's focus away from the supervised fine-tuning (SFT) paradigm. Many studies suggest that introducing the SFT stage not only fails to improve reasoning ability but may also negatively impact model training. In this study, we revisit this RL-centric belief through a systematic and controlled comparison of SFT and RL on VLM Reasoning. Using identical data sources, we find that the relative effectiveness of SFT and RL is conditional and strongly influenced by model capacity, data scale, and data distribution. Contrary to common assumptions, our findings show that SFT plays a crucial role across several scenarios: (1) Effectiveness for weaker models. SFT more reliably elicits reasoning capabilities in smaller or weaker VLMs. (2) Data efficiency. SFT with only 2K achieves comparable or better reasoning performance to RL with 20K. (3) Cross-modal transferability. SFT demonstrates stronger generalization across modalities. Moreover, we identify a pervasive issue of deceptive rewards, where higher rewards fail to correlate with better reasoning accuracy in RL. These results challenge the prevailing "RL over SFT" narrative. They highlight that the role of SFT may have been underestimated and support a more balanced post-training pipeline in which SFT and RL function as complementary components.

AI-Salesman: Towards Reliable Large Language Model Driven Telemarketing

Goal-driven persuasive dialogue, exemplified by applications like telemarketing, requires sophisticated multi-turn planning and strict factual faithfulness, which remains a significant challenge for even state-of-the-art Large Language Models (LLMs). A lack of task-specific data often limits previous works, and direct LLM application suffers from strategic brittleness and factual hallucination. In this paper, we first construct and release TeleSalesCorpus, the first real-world-grounded dialogue dataset for this domain. We then propose AI-Salesman, a novel framework featuring a dual-stage architecture. For the training stage, we design a Bayesian-supervised reinforcement learning algorithm that learns robust sales strategies from noisy dialogues. For the inference stage, we introduce the Dynamic Outline-Guided Agent (DOGA), which leverages a pre-built script library to provide dynamic, turn-by-turn strategic guidance. Moreover, we design a comprehensive evaluation framework that combines fine-grained metrics for key sales skills with the LLM-as-a-Judge paradigm. Experimental results demonstrate that our proposed AI-Salesman significantly outperforms baseline models in both automatic metrics and comprehensive human evaluations, showcasing its effectiveness in complex persuasive scenarios.

You Only Evaluate Once: A Tree-based Rerank Method at Meituan

Reranking plays a crucial role in modern recommender systems by capturing the mutual influences within the list. Due to the inherent challenges of combinatorial search spaces, most methods adopt a two-stage search paradigm: a simple General Search Unit (GSU) efficiently reduces the candidate space, and an Exact Search Unit (ESU) effectively selects the optimal sequence. These methods essentially involve making trade-offs between effectiveness and efficiency, while suffering from a severe \textbf{inconsistency problem}, that is, the GSU often misses high-value lists from ESU. To address this problem, we propose YOLOR, a one-stage reranking method that removes the GSU while retaining only the ESU. Specifically, YOLOR includes: (1) a Tree-based Context Extraction Module (TCEM) that hierarchically aggregates multi-scale contextual features to achieve "list-level effectiveness", and (2) a Context Cache Module (CCM) that enables efficient feature reuse across candidate permutations to achieve "permutation-level efficiency". Extensive experiments across public and industry datasets validate YOLOR's performance, and we have successfully deployed YOLOR on the Meituan food delivery platform.

NGA: Non-autoregressive Generative Auction with Global Externalities for Advertising Systems

Online advertising auctions are fundamental to internet commerce, demanding solutions that not only maximize revenue but also ensure incentive compatibility, high-quality user experience, and real-time efficiency. While recent learning-based auction frameworks have improved context modeling by capturing intra-list dependencies among ads, they remain limited in addressing global externalities and often suffer from inefficiencies caused by sequential processing. In this work, we introduce the Non-autoregressive Generative Auction with global externalities (NGA), a novel end-to-end framework designed for industrial online advertising. NGA explicitly models global externalities by jointly capturing the relationships among ads as well as the effects of adjacent organic content. To further enhance efficiency, NGA utilizes a non-autoregressive, constraint-based decoding strategy and a parallel multi-tower evaluator for unified list-wise reward and payment computation. Extensive offline experiments and large-scale online A/B testing on commercial advertising platforms demonstrate that NGA consistently outperforms existing methods in both effectiveness and efficiency.

One Model to Rank Them All: Unifying Online Advertising with End-to-End Learning

Modern industrial advertising systems commonly employ Multi-stage Cascading Architectures (MCA) to balance computational efficiency with ranking accuracy. However, this approach presents two fundamental challenges: (1) performance inconsistencies arising from divergent optimization targets and capability differences between stages, and (2) failure to account for advertisement externalities - the complex interactions between candidate ads during ranking. These limitations ultimately compromise system effectiveness and reduce platform profitability. In this paper, we present UniROM, an end-to-end generative architecture that Unifies online advertising Ranking as One Model. UniROM replaces cascaded stages with a single model to directly generate optimal ad sequences from the full candidate ad corpus in location-based services (LBS). The primary challenges associated with this approach stem from high costs of feature processing and computational bottlenecks in modeling externalities of large-scale candidate pools. To address these challenges, UniROM introduces an algorithm and engine co-designed hybrid feature service to decouple user and ad feature processing, reducing latency while preserving expressiveness. To efficiently extract intra- and cross-sequence mutual information, we propose RecFormer with an innovative cluster-attention mechanism as its core architectural component. Furthermore, we propose a bi-stage training strategy that integrates pre-training with reinforcement learning-based post-training to meet sophisticated platform and advertising objectives. Extensive offline evaluations on public benchmarks and large-scale online A/B testing on industrial advertising platform have demonstrated the superior performance of UniROM over state-of-the-art MCAs.

Beyond Cascaded Architectures: An End-to-end Generative Framework for Industrial Advertising

Traditional online industrial advertising systems suffer from the limitations of multi-stage cascaded architectures, which often discard high-potential candidates prematurely and distribute decision logic across disconnected modules. While recent generative recommendation approaches provide end-to-end solutions, they fail to address critical advertising requirements of key components for real-world deployment, such as explicit bidding, creative selection, ad allocation, and payment computation. To bridge this gap, we introduce End-to-End Generative Advertising (EGA), the first unified framework that holistically models user interests, point-of-interest (POI) and creative generation, ad allocation, and payment optimization within a single generative model. Our approach employs hierarchical tokenization and multi-token prediction to jointly generate POI recommendations and ad creatives, while a permutation-aware reward model and token-level bidding strategy ensure alignment with both user experiences and advertiser objectives. Additionally, we decouple allocation from payment using a differentiable ex-post regret minimization mechanism, guaranteeing approximate incentive compatibility at the POI level. Through extensive offline evaluations and large-scale online experiments on real-world advertising platforms, we demonstrate that EGA significantly outperforms traditional cascaded systems in both performance and practicality. Our results highlight its potential as a pioneering fully generative advertising solution, paving the way for next-generation industrial ad systems.

NLGR: Utilizing Neighbor Lists for Generative Rerank in Personalized Recommendation Systems

Reranking plays a crucial role in modern multi-stage recommender systems by rearranging the initial ranking list. Due to the inherent challenges of combinatorial search spaces, some current research adopts an evaluator-generator paradigm, with a generator generating feasible sequences and an evaluator selecting the best sequence based on the estimated list utility. However, these methods still face two issues. Firstly, due to the goal inconsistency problem between the evaluator and generator, the generator tends to fit the local optimal solution of exposure distribution rather than combinatorial space optimization. Secondly, the strategy of generating target items one by one is difficult to achieve optimality because it ignores the information of subsequent items. To address these issues, we propose a utilizing Neighbor Lists model for Generative Reranking (NLGR), which aims to improve the performance of the generator in the combinatorial space. NLGR follows the evaluator-generator paradigm and improves the generator's training and generating methods. Specifically, we use neighbor lists in combination space to enhance the training process, making the generator perceive the relative scores and find the optimization direction. Furthermore, we propose a novel sampling-based non-autoregressive generation method, which allows the generator to jump flexibly from the current list to any neighbor list. Extensive experiments on public and industrial datasets validate NLGR's effectiveness and we have successfully deployed NLGR on the Meituan food delivery platform.

Gaussian Mixture based Evidential Learning for Stereo Matching

In this paper, we introduce a novel Gaussian mixture based evidential learning solution for robust stereo matching. Diverging from previous evidential deep learning approaches that rely on a single Gaussian distribution, our framework posits that individual image data adheres to a mixture-of-Gaussian distribution in stereo matching. This assumption yields more precise pixel-level predictions and more accurately mirrors the real-world image distribution. By further employing the inverse-Gamma distribution as an intermediary prior for each mixture component, our probabilistic model achieves improved depth estimation compared to its counterpart with the single Gaussian and effectively captures the model uncertainty, which enables a strong cross-domain generation ability. We evaluated our method for stereo matching by training the model using the Scene Flow dataset and testing it on KITTI 2015 and Middlebury 2014. The experiment results consistently show that our method brings improvements over the baseline methods in a trustworthy manner. Notably, our approach achieved new state-of-the-art results on both the in-domain validated data and the cross-domain datasets, demonstrating its effectiveness and robustness in stereo matching tasks.

Off-Policy Primal-Dual Safe Reinforcement Learning

Primal-dual safe RL methods commonly perform iterations between the primal update of the policy and the dual update of the Lagrange Multiplier. Such a training paradigm is highly susceptible to the error in cumulative cost estimation since this estimation serves as the key bond connecting the primal and dual update processes. We show that this problem causes significant underestimation of cost when using off-policy methods, leading to the failure to satisfy the safety constraint. To address this issue, we propose \textit{conservative policy optimization}, which learns a policy in a constraint-satisfying area by considering the uncertainty in cost estimation. This improves constraint satisfaction but also potentially hinders reward maximization. We then introduce \textit{local policy convexification} to help eliminate such suboptimality by gradually reducing the estimation uncertainty. We provide theoretical interpretations of the joint coupling effect of these two ingredients and further verify them by extensive experiments. Results on benchmark tasks show that our method not only achieves an asymptotic performance comparable to state-of-the-art on-policy methods while using much fewer samples, but also significantly reduces constraint violation during training. Our code is available at https://github.com/ZifanWu/CAL.

Deep Automated Mechanism Design for Integrating Ad Auction and Allocation in Feed

E-commerce platforms usually present an ordered list, mixed with several organic items and an advertisement, in response to each user's page view request. This list, the outcome of ad auction and allocation processes, directly impacts the platform's ad revenue and gross merchandise volume (GMV). Specifically, the ad auction determines which ad is displayed and the corresponding payment, while the ad allocation decides the display positions of the advertisement and organic items. The prevalent methods of segregating the ad auction and allocation into two distinct stages face two problems: 1) Ad auction does not consider externalities, such as the influence of actual display position and context on ad Click-Through Rate (CTR); 2) The ad allocation, which utilizes the auction-winning ad's payment to determine the display position dynamically, fails to maintain incentive compatibility (IC) for the advertisement. For instance, in the auction stage employing the traditional Generalized Second Price (GSP) , even if the winning ad increases its bid, its payment remains unchanged. This implies that the advertisement cannot secure a better position and thus loses the opportunity to achieve higher utility in the subsequent ad allocation stage. Previous research often focused on one of the two stages, neglecting the two-stage problem, which may result in suboptimal outcomes...