What, Whether and How? Unveiling Process Reward Models for Thinking with Images Reasoning

The rapid advancement of Large Vision Language Models (LVLMs) has demonstrated excellent abilities in various visual tasks. Building upon these developments, the thinking with images paradigm has emerged, enabling models to dynamically edit and re-encode visual information at each reasoning step, mirroring human visual processing. However, this paradigm introduces significant challenges as diverse errors may occur during reasoning processes. This necessitates Process Reward Models (PRMs) for distinguishing positive and negative reasoning steps, yet existing benchmarks for PRMs are predominantly text-centric and lack comprehensive assessment under this paradigm. To address these gaps, this work introduces the first comprehensive benchmark specifically designed for evaluating PRMs under the thinking with images paradigm. Our main contributions are: (1) Through extensive analysis of reasoning trajectories and guided search experiments with PRMs, we define 7 fine-grained error types and demonstrate both the necessity for specialized PRMs and the potential for improvement. (2) We construct a comprehensive benchmark comprising 1,206 manually annotated thinking with images reasoning trajectories spanning 4 categories and 16 subcategories for fine-grained evaluation of PRMs. (3) Our experimental analysis reveals that current LVLMs fall short as effective PRMs, exhibiting limited capabilities in visual reasoning process evaluation with significant performance disparities across error types, positive evaluation bias, and sensitivity to reasoning step positions. These findings demonstrate the effectiveness of our benchmark and establish crucial foundations for advancing PRMs in LVLMs.

Modeling Cascaded Delay Feedback for Online Net Conversion Rate Prediction: Benchmark, Insights and Solutions

In industrial recommender systems, conversion rate (CVR) is widely used for traffic allocation, but it fails to fully reflect recommendation effectiveness because it ignores refund behavior. To better capture true user satisfaction and business value, net conversion rate (NetCVR), defined as the probability that a clicked item is purchased and not refunded, has been proposed.Unlike CVR, NetCVR prediction involves a more complex multi-stage cascaded delayed feedback process. The two cascaded delays from click to conversion and from conversion to refund have opposite effects, making traditional CVR modeling methods inapplicable. Moreover, the lack of open-source datasets and online continuous training schemes further hinders progress in this area.To address these challenges, we introduce CASCADE (Cascaded Sequences of Conversion and Delayed Refund), the first large-scale open dataset derived from the Taobao app for online continuous NetCVR prediction. Through an in-depth analysis of CASCADE, we identify three key insights: (1) NetCVR exhibits strong temporal dynamics, necessitating online continuous modeling; (2) cascaded modeling of CVR and refund rate outperforms direct NetCVR modeling; and (3) delay time, which correlates with both CVR and refund rate, is an important feature for NetCVR prediction.Based on these insights, we propose TESLA, a continuous NetCVR modeling framework featuring a CVR-refund-rate cascaded architecture, stage-wise debiasing, and a delay-time-aware ranking loss. Extensive experiments demonstrate that TESLA consistently outperforms state-of-the-art methods on CASCADE, achieving absolute improvements of 12.41 percent in RI-AUC and 14.94 percent in RI-PRAUC on NetCVR prediction. The code and dataset are publicly available at https://github.com/alimama-tech/NetCVR.

Delayed Feedback Modeling for Post-Click Gross Merchandise Volume Prediction: Benchmark, Insights and Approaches

The prediction objectives of online advertisement ranking models are evolving from probabilistic metrics like conversion rate (CVR) to numerical business metrics like post-click gross merchandise volume (GMV). Unlike the well-studied delayed feedback problem in CVR prediction, delayed feedback modeling for GMV prediction remains unexplored and poses greater challenges, as GMV is a continuous target, and a single click can lead to multiple purchases that cumulatively form the label. To bridge the research gap, we establish TRACE, a GMV prediction benchmark containing complete transaction sequences rising from each user click, which supports delayed feedback modeling in an online streaming manner. Our analysis and exploratory experiments on TRACE reveal two key insights: (1) the rapid evolution of the GMV label distribution necessitates modeling delayed feedback under online streaming training; (2) the label distribution of repurchase samples substantially differs from that of single-purchase samples, highlighting the need for separate modeling. Motivated by these findings, we propose RepurchasE-Aware Dual-branch prEdictoR (READER), a novel GMV modeling paradigm that selectively activates expert parameters according to repurchase predictions produced by a router. Moreover, READER dynamically calibrates the regression target to mitigate under-estimation caused by incomplete labels. Experimental results show that READER yields superior performance on TRACE over baselines, achieving a 2.19% improvement in terms of accuracy. We believe that our study will open up a new avenue for studying online delayed feedback modeling for GMV prediction, and our TRACE benchmark with the gathered insights will facilitate future research and application in this promising direction. Our code and dataset are available at https://github.com/alimama-tech/OnlineGMV .

AM$^3$Safety: Towards Data Efficient Alignment of Multi-modal Multi-turn Safety for MLLMs

Multi-modal Large Language Models (MLLMs) are increasingly deployed in interactive applications. However, their safety vulnerabilities become pronounced in multi-turn multi-modal scenarios, where harmful intent can be gradually reconstructed across turns, and security protocols fade into oblivion as the conversation progresses. Existing Reinforcement Learning from Human Feedback (RLHF) alignment methods are largely developed for single-turn visual question-answer (VQA) task and often require costly manual preference annotations, limiting their effectiveness and scalability in dialogues. To address this challenge, we present InterSafe-V, an open-source multi-modal dialogue dataset containing 11,270 dialogues and 500 specially designed refusal VQA samples. This dataset, constructed through interaction between several models, is designed to more accurately reflect real-world scenarios and includes specialized VQA pairs tailored for specific domains. Building on this dataset, we propose AM$^3$Safety, a framework that combines a cold-start refusal phase with Group Relative Policy Optimization (GRPO) fine-tuning using turn-aware dual-objective rewards across entire dialogues. Experiments on Qwen2.5-VL-7B-Instruct and LLaVA-NeXT-7B show more than 10\% decrease in Attack Success Rate (ASR) together with an increment of at least 8\% in harmless dimension and over 13\% in helpful dimension of MLLMs on multi-modal multi-turn safety benchmarks, while preserving their general abilities.

Flow2GAN: Hybrid Flow Matching and GAN with Multi-Resolution Network for Few-step High-Fidelity Audio Generation

Existing dominant methods for audio generation include Generative Adversarial Networks (GANs) and diffusion-based methods like Flow Matching. GANs suffer from slow convergence and potential mode collapse during training, while diffusion methods require multi-step inference that introduces considerable computational overhead. In this work, we introduce Flow2GAN, a two-stage framework that combines Flow Matching training for learning generative capabilities with GAN fine-tuning for efficient few-step inference. Specifically, given audio's unique properties, we first improve Flow Matching for audio modeling through: 1) reformulating the objective as endpoint estimation, avoiding velocity estimation difficulties when involving empty regions; 2) applying spectral energy-based loss scaling to emphasize perceptually salient quieter regions. Building on these Flow Matching adaptations, we demonstrate that a further stage of lightweight GAN fine-tuning enables us to obtain one-step generator that produces high-quality audio. In addition, we develop a multi-branch network architecture that processes Fourier coefficients at different time-frequency resolutions, which improves the modeling capabilities compared to prior single-resolution designs. Experimental results indicate that our Flow2GAN delivers high-fidelity audio generation from Mel-spectrograms or discrete audio tokens, achieving better quality-efficiency trade-offs than existing state-of-the-art GAN-based and Flow Matching-based methods. Online demo samples are available at https://flow2gan.github.io, and the source code is released at https://github.com/k2-fsa/Flow2GAN.

LLM-Auction: Generative Auction towards LLM-Native Advertising

The rapid advancement of large language models (LLMs) necessitates novel monetization strategies, among which LLM-native advertising has emerged as a promising paradigm by naturally integrating advertisement within LLM-generated responses. However, this paradigm fundamentally shifts the auction object from discrete ad slots to the distribution over LLM outputs, posing new challenges for designing auction mechanisms. Existing mechanisms for LLM-native advertising adopt frameworks that decouple auction and generation, which either ignore externalities or require multiple LLM inferences for ad allocation, rendering them impractical for industrial scenarios. To address these challenges, we propose LLM-Auction, which to the best of our knowledge is the first learning-based generative auction mechanism that integrates auction and LLM generation for LLM-native advertising. By formulating the allocation optimization as a preference alignment problem between LLM outputs and the mechanism's objective which reflects both advertisers' expected value and user experience, we introduce Iterative Reward-Preference Optimization (IRPO) algorithm that alternately optimizes the reward model and the LLM. This approach enables the LLM to inherently model allocation externalities without any extra inference cost. We further identify the allocation monotonicity and continuity of LLM-Auction, which allows us to prove that a simple first-price payment rule exhibits favorable incentive properties. Additionally, we design an LLM-as-a-judge simulation environment to facilitate large-scale data construction and enable comprehensive quantitative evaluation of the mechanism's performance. Extensive quantitative and qualitative experiments demonstrate that LLM-Auction significantly outperforms existing baselines in allocation efficiency, while achieving the desired mechanism properties.

MUSE: A Simple Yet Effective Multimodal Search-Based Framework for Lifelong User Interest Modeling

Lifelong user interest modeling is crucial for industrial recommender systems, yet existing approaches rely predominantly on ID-based features, suffering from poor generalization on long-tail items and limited semantic expressiveness. While recent work explores multimodal representations for behavior retrieval in the General Search Unit (GSU), they often neglect multimodal integration in the fine-grained modeling stage -- the Exact Search Unit (ESU). In this work, we present a systematic analysis of how to effectively leverage multimodal signals across both stages of the two-stage lifelong modeling framework. Our key insight is that simplicity suffices in the GSU: lightweight cosine similarity with high-quality multimodal embeddings outperforms complex retrieval mechanisms. In contrast, the ESU demands richer multimodal sequence modeling and effective ID-multimodal fusion to unlock its full potential. Guided by these principles, we propose MUSE, a simple yet effective multimodal search-based framework. MUSE has been deployed in Taobao display advertising system, enabling 100K-length user behavior sequence modeling and delivering significant gains in top-line metrics with negligible online latency overhead. To foster community research, we share industrial deployment practices and open-source the first large-scale dataset featuring ultra-long behavior sequences paired with high-quality multimodal embeddings. Our code and data is available at https://taobao-mm.github.io.

AIF: Asynchronous Inference Framework for Cost-Effective Pre-Ranking

In industrial recommendation systems, pre-ranking models based on deep neural networks (DNNs) commonly adopt a sequential execution framework: feature fetching and model forward computation are triggered only after receiving candidates from the upstream retrieval stage. This design introduces inherent bottlenecks, including redundant computations of identical users/items and increased latency due to strictly sequential operations, which jointly constrain the model's capacity and system efficiency. To address these limitations, we propose the Asynchronous Inference Framework (AIF), a cost-effective computational architecture that decouples interaction-independent components, those operating within a single user or item, from real-time prediction. AIF reorganizes the model inference process by performing user-side computations in parallel with the retrieval stage and conducting item-side computations in a nearline manner. This means that interaction-independent components are calculated just once and completed before the real-time prediction phase of the pre-ranking stage. As a result, AIF enhances computational efficiency and reduces latency, freeing up resources to significantly improve the feature set and model architecture of interaction-independent components. Moreover, we delve into model design within the AIF framework, employing approximated methods for interaction-dependent components in online real-time predictions. By co-designing both the framework and the model, our solution achieves notable performance gains without significantly increasing computational and latency costs. This has enabled the successful deployment of AIF in the Taobao display advertising system.

Think before Recommendation: Autonomous Reasoning-enhanced Recommender

The core task of recommender systems is to learn user preferences from historical user-item interactions. With the rapid development of large language models (LLMs), recent research has explored leveraging the reasoning capabilities of LLMs to enhance rating prediction tasks. However, existing distillation-based methods suffer from limitations such as the teacher model's insufficient recommendation capability, costly and static supervision, and superficial transfer of reasoning ability. To address these issues, this paper proposes RecZero, a reinforcement learning (RL)-based recommendation paradigm that abandons the traditional multi-model and multi-stage distillation approach. Instead, RecZero trains a single LLM through pure RL to autonomously develop reasoning capabilities for rating prediction. RecZero consists of two key components: (1) "Think-before-Recommendation" prompt construction, which employs a structured reasoning template to guide the model in step-wise analysis of user interests, item features, and user-item compatibility; and (2) rule-based reward modeling, which adopts group relative policy optimization (GRPO) to compute rewards for reasoning trajectories and optimize the LLM. Additionally, the paper explores a hybrid paradigm, RecOne, which combines supervised fine-tuning with RL, initializing the model with cold-start reasoning samples and further optimizing it with RL. Experimental results demonstrate that RecZero and RecOne significantly outperform existing baseline methods on multiple benchmark datasets, validating the superiority of the RL paradigm in achieving autonomous reasoning-enhanced recommender systems.

See Beyond a Single View: Multi-Attribution Learning Leads to Better Conversion Rate Prediction

Conversion rate (CVR) prediction is a core component of online advertising systems, where the attribution mechanisms-rules for allocating conversion credit across user touchpoints-fundamentally determine label generation and model optimization. While many industrial platforms support diverse attribution mechanisms (e.g., First-Click, Last-Click, Linear, and Data-Driven Multi-Touch Attribution), conventional approaches restrict model training to labels from a single production-critical attribution mechanism, discarding complementary signals in alternative attribution perspectives. To address this limitation, we propose a novel Multi-Attribution Learning (MAL) framework for CVR prediction that integrates signals from multiple attribution perspectives to better capture the underlying patterns driving user conversions. Specifically, MAL is a joint learning framework consisting of two core components: the Attribution Knowledge Aggregator (AKA) and the Primary Target Predictor (PTP). AKA is implemented as a multi-task learner that integrates knowledge extracted from diverse attribution labels. PTP, in contrast, focuses on the task of generating well-calibrated conversion probabilities that align with the system-optimized attribution metric (e.g., CVR under the Last-Click attribution), ensuring direct compatibility with industrial deployment requirements. Additionally, we propose CAT, a novel training strategy that leverages the Cartesian product of all attribution label combinations to generate enriched supervision signals. This design substantially enhances the performance of the attribution knowledge aggregator. Empirical evaluations demonstrate the superiority of MAL over single-attribution learning baselines, achieving +0.51% GAUC improvement on offline metrics. Online experiments demonstrate that MAL achieved a +2.6% increase in ROI (Return on Investment).