arXiv subjects

Shuai Zhu

Publications and source records attributed to Shuai Zhu.

10 recordsLinked to original sources

arXiv2026

CAP: A Scalable Benchmark for Evaluating Cross-Site Browser Agents with Complex Actions and Perception

Large language models are increasingly deployed as autonomous agents that interact with the web through browsers. While recent progress has been driven by benchmarks that evaluate end-to-end task success, these evaluations largely overlook two fundamental sources of difficulty in real web browsing: complex actions over rich user interfaces and visual perception of dynamically rendered content, especially in workflows that span multiple websites. We introduce CAP, a scalable benchmark for evaluating browser agents on cross-site, human-like web tasks that require non-trivial UI interactions and visual understanding. Specifically, we adopt a decomposition-and-recomposition pipeline that first abstracts each website into a structured site card capturing user-facing functions, complex execution operations, and perceptual requirements, and then recomposes these components into realistic cross-site workflows. Each task is therefore grounded in multiple specific operations on each website, enabling fine-grained diagnosis. Built on this framework, we construct 420 tasks across 108 real-world websites and 24 domains under careful quality control. Experiments on state-of-the-art browser agents using our verifiable agent-as-a-judge evaluation framework show low success rates and reveal that perception-heavy interactions remain a major bottleneck, exposing substantial gaps between current agents and real-world web browsing demands.

cs.AI↗

arXiv2026

CriPO: Enhancing Rubric-based RL via Self-Distillation

Rubric-based RL has recently shown promise in improving LLMs on open-ended tasks. A widely recognized limitation of rubric-based RL is limited exploration: criteria that no rollout manages to satisfy (Unexplored Criteria, UC) receive no optimization signal. Recent methods address this by incorporating rubric information as external guidance during rollout, yet they introduce a train-inference mismatch: the policy is optimized on rollouts produced under external guidance while this guidance is absent at inference time, causing error accumulation through autoregressive decoding. Moreover, these exploration-focused approaches overlook a fundamentally different failure mode that we term Suppressed Criteria (SC) -- criteria that are satisfied by some rollouts yet whose learning signals are lost during optimization because scalar reward aggregation assigns them non-positive aggregate advantages. Our analysis reveals that SC are remarkably prevalent: over 57% of samples exhibit this failure mode throughout training, with an average of 1.8 SC per sample. To simultaneously address both UC and SC without introducing training-inference mismatch, we propose Criterion-Distilled Policy Optimization (CriPO), which enhances rubric-based RL via on-policy self-distillation. For UC, CriPO constructs a criterion-injection self-teacher and computes a localized forward-KL loss to inject missing behaviors into the policy. For SC, CriPO employs a counterfactual self-teacher to locate criterion-relevant tokens in negative-advantage rollouts and flips their token-level advantages to positive values, preserving useful patterns that would otherwise be suppressed. Experiments on medicine and science benchmarks demonstrate that CriPO consistently outperforms rubric-based RL, achieving stronger final performance with approximately $2\times$ fewer optimization steps.

cs.LG↗

arXiv2026

SODA: Semantic-Oriented Distributional Alignment for Generative Recommendation

Generative recommendation has emerged as a scalable alternative to traditional retrieve-and-rank pipelines by operating in a compact token space. However, existing methods mainly rely on discrete code-level supervision, which leads to information loss and limits the joint optimization between the tokenizer and the generative recommender. In this work, we propose a distribution-level supervision paradigm that leverages probability distributions over multi-layer codebooks as soft and information-rich representations. Building on this idea, we introduce Semantic-Oriented Distributional Alignment (SODA), a plug-and-play contrastive supervision framework based on Bayesian Personalized Ranking, which aligns semantically rich distributions via negative KL divergence while enabling end-to-end differentiable training. Extensive experiments on multiple real-world datasets demonstrate that SODA consistently improves the performance of various generative recommender backbones, validating its effectiveness and generality. Codes will be available upon acceptance.

cs.IR↗

arXiv2026

UniGRec: Unified Generative Recommendation with Soft Identifiers for End-to-End Optimization

Generative recommendation has recently emerged as a transformative paradigm that directly generates target items, surpassing traditional cascaded approaches. It typically involves two components: a tokenizer that learns item identifiers and a recommender trained on them. Existing methods often decouple tokenization from recommendation or rely on asynchronous alternating optimization, limiting full end-to-end alignment. To address this, we unify the tokenizer and recommender under the ultimate recommendation objective via differentiable soft item identifiers, enabling joint end-to-end training. However, this introduces three challenges: training-inference discrepancy due to soft-to-hard mismatch, item identifier collapse from codeword usage imbalance, and collaborative signal deficiency due to an overemphasis on fine-grained token-level semantics. To tackle these challenges, we propose UniGRec, a unified generative recommendation framework that addresses them from three perspectives. UniGRec employs Annealed Inference Alignment during tokenization to smoothly bridge soft training and hard inference, a Codeword Uniformity Regularization to prevent identifier collapse and encourage codebook diversity, and a Dual Collaborative Distillation mechanism that distills collaborative priors from a lightweight teacher model to jointly guide both the tokenizer and the recommender. Extensive experiments on real-world datasets demonstrate that UniGRec consistently outperforms state-of-the-art baseline methods. Our codes are available at https://github.com/Jialei-03/UniGRec.

cs.IR↗

arXiv2024

MOBIUS: Towards the Next Generation of Query-Ad Matching in Baidu's Sponsored Search

Baidu runs the largest commercial web search engine in China, serving hundreds of millions of online users every day in response to a great variety of queries. In order to build a high-efficiency sponsored search engine, we used to adopt a three-layer funnel-shaped structure to screen and sort hundreds of ads from billions of ad candidates subject to the requirement of low response latency and the restraints of computing resources. Given a user query, the top matching layer is responsible for providing semantically relevant ad candidates to the next layer, while the ranking layer at the bottom concerns more about business indicators (e.g., CPM, ROI, etc.) of those ads. The clear separation between the matching and ranking objectives results in a lower commercial return. The Mobius project has been established to address this serious issue. It is our first attempt to train the matching layer to consider CPM as an additional optimization objective besides the query-ad relevance, via directly predicting CTR (click-through rate) from billions of query-ad pairs. Specifically, this paper will elaborate on how we adopt active learning to overcome the insufficiency of click history at the matching layer when training our neural click networks offline, and how we use the SOTA ANN search technique for retrieving ads more efficiently (Here ``ANN'' stands for approximate nearest neighbor search). We contribute the solutions to Mobius-V1 as the first version of our next generation query-ad matching system.

cs.IR↗

arXiv2024

Generating Human Motion in 3D Scenes from Text Descriptions

Generating human motions from textual descriptions has gained growing research interest due to its wide range of applications. However, only a few works consider human-scene interactions together with text conditions, which is crucial for visual and physical realism. This paper focuses on the task of generating human motions in 3D indoor scenes given text descriptions of the human-scene interactions. This task presents challenges due to the multi-modality nature of text, scene, and motion, as well as the need for spatial reasoning. To address these challenges, we propose a new approach that decomposes the complex problem into two more manageable sub-problems: (1) language grounding of the target object and (2) object-centric motion generation. For language grounding of the target object, we leverage the power of large language models. For motion generation, we design an object-centric scene representation for the generative model to focus on the target object, thereby reducing the scene complexity and facilitating the modeling of the relationship between human motions and the object. Experiments demonstrate the better motion quality of our approach compared to baselines and validate our design choices.

cs.CV↗

arXiv2023

Multi-view Hypergraph Contrastive Policy Learning for Conversational Recommendation

Conversational recommendation systems (CRS) aim to interactively acquire user preferences and accordingly recommend items to users. Accurately learning the dynamic user preferences is of crucial importance for CRS. Previous works learn the user preferences with pairwise relations from the interactive conversation and item knowledge, while largely ignoring the fact that factors for a relationship in CRS are multiplex. Specifically, the user likes/dislikes the items that satisfy some attributes (Like/Dislike view). Moreover social influence is another important factor that affects user preference towards the item (Social view), while is largely ignored by previous works in CRS. The user preferences from these three views are inherently different but also correlated as a whole. The user preferences from the same views should be more similar than that from different views. The user preferences from Like View should be similar to Social View while different from Dislike View. To this end, we propose a novel model, namely Multi-view Hypergraph Contrastive Policy Learning (MHCPL). Specifically, MHCPL timely chooses useful social information according to the interactive history and builds a dynamic hypergraph with three types of multiplex relations from different views. The multiplex relations in each view are successively connected according to their generation order.

cs.IR↗

arXiv2023

Towards Hierarchical Policy Learning for Conversational Recommendation with Hypergraph-based Reinforcement Learning

Conversational recommendation systems (CRS) aim to timely and proactively acquire user dynamic preferred attributes through conversations for item recommendation. In each turn of CRS, there naturally have two decision-making processes with different roles that influence each other: 1) director, which is to select the follow-up option (i.e., ask or recommend) that is more effective for reducing the action space and acquiring user preferences; and 2) actor, which is to accordingly choose primitive actions (i.e., asked attribute or recommended item) that satisfy user preferences and give feedback to estimate the effectiveness of the director's option. However, existing methods heavily rely on a unified decision-making module or heuristic rules, while neglecting to distinguish the roles of different decision procedures, as well as the mutual influences between them. To address this, we propose a novel Director-Actor Hierarchical Conversational Recommender (DAHCR), where the director selects the most effective option, followed by the actor accordingly choosing primitive actions that satisfy user preferences. Specifically, we develop a dynamic hypergraph to model user preferences and introduce an intrinsic motivation to train from weak supervision over the director. Finally, to alleviate the bad effect of model bias on the mutual influence between the director and actor, we model the director's option by sampling from a categorical distribution. Extensive experiments demonstrate that DAHCR outperforms state-of-the-art methods.

cs.IR↗

arXiv2022

On-device Training: A First Overview on Existing Systems

The recent breakthroughs in machine learning (ML) and deep learning (DL) have catalyzed the design and development of various intelligent systems over wide application domains. While most existing machine learning models require large memory and computing power, efforts have been made to deploy some models on resource-constrained devices as well. A majority of the early application systems focused on exploiting the inference capabilities of ML and DL models, where data captured from different mobile and embedded sensing components are processed through these models for application goals such as classification and segmentation. More recently, the concept of exploiting the mobile and embedded computing resources for ML/DL model training has gained attention, as such capabilities allow (i) the training of models via local data without the need to share data over wireless links, thus enabling privacy-preserving computation by design, (ii) model personalization and environment adaptation, and (ii) deployment of accurate models in remote and hardly accessible locations without stable internet connectivity. This work targets to summarize and analyze state-of-the-art systems research that allows such on-device model training capabilities and provide a survey of on-device training from a systems perspective.

cs.LG↗

arXiv2022

AdaCoach: A Virtual Coach for Training Customer Service Agents

With the development of online business, customer service agents gradually play a crucial role as an interface between the companies and their customers. Most companies spend a lot of time and effort on hiring and training customer service agents. To this end, we propose AdaCoach: A Virtual Coach for Training Customer Service Agents, to promote the ability of newly hired service agents before they get to work. AdaCoach is designed to simulate real customers who seek help and actively initiate the dialogue with the customer service agents. Besides, AdaCoach uses an automated dialogue evaluation model to score the performance of the customer agent in the training process, which can provide necessary assistance when the newly hired customer service agent encounters problems. We apply recent NLP technologies to ensure efficient run-time performance in the deployed system. To the best of our knowledge, this is the first system that trains the customer service agent through human-computer interaction. Until now, the system has already supported more than 500,000 simulation training and cultivated over 1000 qualified customer service agents.

cs.CL↗