Zheqing Zhu

Zheqing Zhu, Benjamin Van Roy · stanford

High-quality recommender systems ought to deliver both innovative and relevant content through effective and exploratory interactions with users. Yet, supervised learning-based neural networks, which form the backbone of many existing recommender systems, only leverage recognized user interests, falling short when it comes to efficiently uncovering unknown user preferences. While there has been some progress with neural contextual bandit algorithms towards enabling online exploration through neural networks, their onerous computational demands hinder widespread adoption in real-world recommender systems. In this work, we propose a scalable sample-efficient neural contextual bandit algorithm for recommender systems. To do this, we design an epistemic neural network architecture, Epistemic Neural Recommendation (ENR), that enables Thompson sampling at a large scale. In two distinct large-scale experiments with real-world tasks, ENR significantly boosts click-through rates and user ratings by at least 9% and 6% respectively compared to state-of-the-art neural contextual bandit algorithms. Furthermore, it achieves equivalent performance with at least 29% fewer user interactions compared to the best-performing baseline algorithm. Remarkably, while accomplishing these improvements, ENR demands orders of magnitude fewer computational resources than neural contextual bandit baseline algorithms.

Rohan Chitnis, Yingchen Xu, Bobak Hashemi et al.

Model-based reinforcement learning (RL) has shown great promise due to its sample efficiency, but still struggles with long-horizon sparse-reward tasks, especially in offline settings where the agent learns from a fixed dataset. We hypothesize that model-based RL agents struggle in these environments due to a lack of long-term planning capabilities, and that planning in a temporally abstract model of the environment can alleviate this issue. In this paper, we make two key contributions: 1) we introduce an offline model-based RL algorithm, IQL-TD-MPC, that extends the state-of-the-art Temporal Difference Learning for Model Predictive Control (TD-MPC) with Implicit Q-Learning (IQL); 2) we propose to use IQL-TD-MPC as a Manager in a hierarchical setting with any off-the-shelf offline RL algorithm as a Worker. More specifically, we pre-train a temporally abstract IQL-TD-MPC Manager to predict "intent embeddings", which roughly correspond to subgoals, via planning. We empirically show that augmenting state representations with intent embeddings generated by an IQL-TD-MPC manager significantly improves off-the-shelf offline RL agents' performance on some of the most challenging D4RL benchmark tasks. For instance, the offline RL algorithms AWAC, TD3-BC, DT, and CQL all get zero or near-zero normalized evaluation scores on the medium and large antmaze tasks, while our modification gives an average score over 40.

Dmytro Korenkevych, Frank Cheng, Artsiom Balakir et al.

The online advertising market, with its thousands of auctions run per second, presents a daunting challenge for advertisers who wish to optimize their spend under a budget constraint. Thus, advertising platforms typically provide automated agents to their customers, which act on their behalf to bid for impression opportunities in real time at scale. Because these proxy agents are owned by the platform but use advertiser funds to operate, there is a strong practical need to balance reliability and explainability of the agent with optimizing power. We propose a generalizable approach to optimizing bidding policies in production environments by learning from real data using offline reinforcement learning. This approach can be used to optimize any differentiable base policy (practically, a heuristic policy based on principles which the advertiser can easily understand), and only requires data generated by the base policy itself. We use a hybrid agent architecture that combines arbitrary base policies with deep neural networks, where only the optimized base policy parameters are eventually deployed, and the neural network part is discarded after training. We demonstrate that such an architecture achieves statistically significant performance gains in both simulated and at-scale production bidding environments. Our approach does not incur additional infrastructure, safety, or explainability costs, as it directly optimizes parameters of existing production routines without replacing them with black box-style models like neural networks.

Hongbo Guo, Ruben Naeff, Alex Nikulkov et al.

Bandit learning has been an increasingly popular design choice for recommender system. Despite the strong interest in bandit learning from the community, there remains multiple bottlenecks that prevent many bandit learning approaches from productionalization. One major bottleneck is how to test the effectiveness of bandit algorithm with fairness and without data leakage. Different from supervised learning algorithms, bandit learning algorithms emphasize greatly on the data collection process through their explorative nature. Such explorative behavior may induce unfair evaluation in a classic A/B test setting. In this work, we apply upper confidence bound (UCB) to our large scale short video recommender system and present a test framework for the production bandit learning life-cycle with a new set of metrics. Extensive experiment results show that our experiment design is able to fairly evaluate the performance of bandit learning in the recommender system.

Zheqing Zhu, Rodrigo de Salvo Braz, Jalaj Bhandari et al.

Reinforcement learning (RL) is a versatile framework for optimizing long-term goals. Although many real-world problems can be formalized with RL, learning and deploying a performant RL policy requires a system designed to address several important challenges, including the exploration-exploitation dilemma, partial observability, dynamic action spaces, and safety concerns. While the importance of these challenges has been well recognized, existing open-source RL libraries do not explicitly address them. This paper introduces Pearl, a Production-Ready RL software package designed to embrace these challenges in a modular way. In addition to presenting benchmarking results, we also highlight examples of Pearl's ongoing industry adoption to demonstrate its advantages for production use cases. Pearl is open sourced on GitHub at github.com/facebookresearch/pearl and its official website is pearlagent.github.io.

Zheqing Zhu, Yueyang Liu, Xu Kuang et al.

Real-world applications of contextual bandits often exhibit non-stationarity due to seasonality, serendipity, and evolving social trends. While a number of non-stationary contextual bandit learning algorithms have been proposed in the literature, they excessively explore due to a lack of prioritization for information of enduring value, or are designed in ways that do not scale in modern applications with high-dimensional user-specific features and large action set, or both. In this paper, we introduce a novel non-stationary contextual bandit algorithm that addresses these concerns. It combines a scalable, deep-neural-network-based architecture with a carefully designed exploration mechanism that strategically prioritizes collecting information with the most lasting value in a non-stationary environment. Through empirical evaluations on two real-world recommendation datasets, which exhibit pronounced non-stationarity, we demonstrate that our approach significantly outperforms the state-of-the-art baselines.

Yi Wan, Jiuqi Wang, Liam Li et al.

Tool-augmented large language models (LLMs) are emerging as deep research agents, systems that decompose complex queries, retrieve external evidence, and synthesize grounded responses. Yet current agents remain limited by shallow retrieval, weak alignment metrics, and brittle tool-use behavior. We introduce PokeeResearch-7B, a 7B-parameter deep research agent built under a unified reinforcement learning framework for robustness, alignment, and scalability. PokeeResearch-7B is trained by an annotation-free Reinforcement Learning from AI Feedback (RLAIF) framework to optimize policies using LLM-based reward signals that capture factual accuracy, citation faithfulness, and instruction adherence. A chain-of-thought-driven multi-call reasoning scaffold further enhances robustness through self-verification and adaptive recovery from tool failures. Among 10 popular deep research benchmarks, PokeeResearch-7B achieves state-of-the-art performance among 7B-scale deep research agents. This highlights that careful reinforcement learning and reasoning design can produce efficient, resilient, and research-grade AI agents. The model and inference code is open-sourced under Apache 2.0 license at https://github.com/Pokee-AI/PokeeResearchOSS.

Daniel R. Jiang, Alex Nikulkov, Yu-Chia Chen et al.

Generative artificial intelligence (AI), in particular large language models (LLMs), is poised to drive transformative economic change. LLMs are pre-trained on vast text data to learn general language patterns, but a subsequent post-training phase is critical to align them for specific real-world tasks. Reinforcement learning (RL) is the leading post-training technique, yet its economic impact remains largely underexplored and unquantified. We examine this question through the lens of the first deployment of an RL-trained LLM for generative advertising on Facebook. Integrated into Meta's Text Generation feature, our model, "AdLlama," powers an AI tool that helps advertisers create new variations of human-written ad text. To train this model, we introduce reinforcement learning with performance feedback (RLPF), a post-training method that uses historical ad performance data as a reward signal. In a large-scale 10-week A/B test on Facebook spanning nearly 35,000 advertisers and 640,000 ad variations, we find that AdLlama improves click-through rates by 6.7% (p=0.0296) compared to a supervised imitation model trained on curated ads. This represents a substantial improvement in advertiser return on investment on Facebook. We also find that advertisers who used AdLlama generated more ad variations, indicating higher satisfaction with the model's outputs. To our knowledge, this is the largest study to date on the use of generative AI in an ecologically valid setting, offering an important data point quantifying the tangible impact of RL post-training. Furthermore, the results show that RLPF is a promising and generalizable approach for metric-driven post-training that bridges the gap between highly capable language models and tangible outcomes.

Xiaokun Liu, Sayedmohammadreza Rastegari, Yijun Huang et al.

In cancer therapeutics, protein-metal binding mechanisms critically govern the pharmacokinetics and targeting efficacy of drugs, thereby fundamentally shaping the rational design of anticancer metallodrugs. While conventional laboratory methods used to study such mechanisms are often costly, low throughput, and limited in capturing dynamic biological processes, machine learning (ML) has emerged as a promising alternative. Despite increasing efforts to develop protein-metal binding datasets and ML algorithms, the application of ML in tumor protein-metal binding remains limited. Key challenges include a shortage of high-quality, tumor-specific datasets, insufficient consideration of multiple data modalities, and the complexity of interpreting results due to the ''black box'' nature of complex ML models. This paper summarizes recent progress and ongoing challenges in using ML to predict tumor protein-metal binding, focusing on data, modeling, and interpretability. We present multimodal protein-metal binding datasets and outline strategies for acquiring, curating, and preprocessing them for training ML models. Moreover, we explore the complementary value provided by different data modalities and examine methods for their integration. We also review approaches for improving model interpretability to support more trustworthy decisions in cancer research. Finally, we offer our perspective on research opportunities and propose strategies to address the scarcity of tumor protein data and the limited number of predictive models for tumor protein-metal binding. We also highlight two promising directions for effective metal-based drug design: integrating protein-protein interaction data to provide structural insights into metal-binding events and predicting structural changes in tumor proteins after metal binding.

Yi Wan, Dmytro Korenkevych, Zheqing Zhu

In reinforcement learning (RL), continuing tasks refer to tasks where the agent-environment interaction is ongoing and can not be broken down into episodes. These tasks are suitable when environment resets are unavailable, agent-controlled, or predefined but where all rewards-including those beyond resets-are critical. These scenarios frequently occur in real-world applications and can not be modeled by episodic tasks. While modern deep RL algorithms have been extensively studied and well understood in episodic tasks, their behavior in continuing tasks remains underexplored. To address this gap, we provide an empirical study of several well-known deep RL algorithms using a suite of continuing task testbeds based on Mujoco and Atari environments, highlighting several key insights concerning continuing tasks. Using these testbeds, we also investigate the effectiveness of a method for improving temporal-difference-based RL algorithms in continuing tasks by centering rewards, as introduced by Naik et al. (2024). While their work primarily focused on this method in conjunction with Q-learning, our results extend their findings by demonstrating that this method is effective across a broader range of algorithms, scales to larger tasks, and outperforms two other reward-centering approaches.

Hong Jun Jeon, Songbin Liu, Yuantong Li et al.

The exploding popularity of online content and its user base poses an evermore challenging matching problem for modern recommendation systems. Unlike other frontiers of machine learning such as natural language, recommendation systems are responsible for collecting their own data. Simply exploiting current knowledge can lead to pernicious feedback loops but naive exploration can detract from user experience and lead to reduced engagement. This exploration-exploitation trade-off is exemplified in the classic multi-armed bandit problem for which algorithms such as upper confidence bounds (UCB) and Thompson sampling (TS) demonstrate effective performance. However, there have been many challenges to scaling these approaches to settings which do not exhibit a conjugate prior structure. Recent scalable approaches to uncertainty quantification via epinets have enabled efficient approximations of Thompson sampling even when the learning model is a complex neural network. In this paper, we demonstrate the first application of epinets to an online recommendation system. Our experiments demonstrate improvements in both user traffic and engagement efficiency on the Facebook Reels online video platform.

Hongbo Guo, Zheqing Zhu

Contextual bandit learning is increasingly favored in modern large-scale recommendation systems. To better utlize the contextual information and available user or item features, the integration of neural networks have been introduced to enhance contextual bandit learning and has triggered significant interest from both academia and industry. However, a major challenge arises when implementing a disjoint neural contextual bandit solution in large-scale recommendation systems, where each item or user may correspond to a separate bandit arm. The huge number of items to recommend poses a significant hurdle for real world production deployment. This paper focuses on a joint neural contextual bandit solution which serves all recommending items in one single model. The output consists of a predicted reward $μ$, an uncertainty $σ$ and a hyper-parameter $α$ which balances exploitation and exploration, e.g., $μ+ ασ$. The tuning of the parameter $α$ is typically heuristic and complex in practice due to its stochastic nature. To address this challenge, we provide both theoretical analysis and experimental findings regarding the uncertainty $σ$ of the joint neural contextual bandit model. Our analysis reveals that $α$ demonstrates an approximate square root relationship with the size of the last hidden layer $F$ and inverse square root relationship with the amount of training data $N$, i.e., $σ\propto \sqrt{\frac{F}{N}}$. The experiments, conducted with real industrial data, align with the theoretical analysis, help understanding model behaviors and assist the hyper-parameter tuning during both offline training and online deployment.

Ruiyang Xu, Jalaj Bhandari, Dmytro Korenkevych et al.

Auction-based recommender systems are prevalent in online advertising platforms, but they are typically optimized to allocate recommendation slots based on immediate expected return metrics, neglecting the downstream effects of recommendations on user behavior. In this study, we employ reinforcement learning to optimize for long-term return metrics in an auction-based recommender system. Utilizing temporal difference learning, a fundamental reinforcement learning algorithm, we implement an one-step policy improvement approach that biases the system towards recommendations with higher long-term user engagement metrics. This optimizes value over long horizons while maintaining compatibility with the auction framework. Our approach is grounded in dynamic programming ideas which show that our method provably improves upon the existing auction-based base policy. Through an online A/B test conducted on an auction-based recommender system which handles billions of impressions and users daily, we empirically establish that our proposed method outperforms the current production system in terms of long-term user engagement metrics.

Zheqing Zhu, Benjamin Van Roy

Modern recommendation systems ought to benefit by probing for and learning from delayed feedback. Research has tended to focus on learning from a user's response to a single recommendation. Such work, which leverages methods of supervised and bandit learning, forgoes learning from the user's subsequent behavior. Where past work has aimed to learn from subsequent behavior, there has been a lack of effective methods for probing to elicit informative delayed feedback. Effective exploration through probing for delayed feedback becomes particularly challenging when rewards are sparse. To address this, we develop deep exploration methods for recommendation systems. In particular, we formulate recommendation as a sequential decision problem and demonstrate benefits of deep exploration over single-step exploration. Our experiments are carried out with high-fidelity industrial-grade simulators and establish large improvements over existing algorithms.

Zheqing Zhu, Erdem Bıyık, Dorsa Sadigh

Multi-agent safe systems have become an increasingly important area of study as we can now easily have multiple AI-powered systems operating together. In such settings, we need to ensure the safety of not only each individual agent, but also the overall system. In this paper, we introduce a novel multi-agent safe learning algorithm that enables decentralized safe navigation when there are multiple different agents in the environment. This algorithm makes mild assumptions about other agents and is trained in a decentralized fashion, i.e. with very little prior knowledge about other agents' policies. Experiments show our algorithm performs well with the robots running other algorithms when optimizing various objectives.