Ad hoc teamwork (AHT) requires agents to collaborate with previously unseen teammates, which is crucial for many real-world applications. The core challenge of AHT is to develop an ego agent that can predict and adapt to unknown teammates on the fly. Conventional RL-based approaches optimize a single expected return, which often causes policies to collapse into a single dominant behavior, thus failing to capture the multimodal cooperation patterns inherent in AHT. In this work, we introduce PADiff, a diffusion-based approach that captures agent's multimodal behaviors, unlocking its diverse cooperation modes with teammates. However, standard diffusion models lack the ability to predict and adapt in non-stationary AHT scenarios. To address this limitation, we propose a novel diffusion-based policy that integrates critical predictive information about teammates into the denoising process. Extensive experiments across three environments demonstrate that PADiff outperforms existing AHT methods significantly.

ZhenChun Xu, Yi Cai, Dajun Zheng, li Yuan, Mengchen Zhao, Qixiang Wang, Jiexin Wang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Ronghui Yang, Jie Liu, Jiajie Zeng, Jiexin Wang, Jiuchuan Jiang, Bo An, Yi Cai, Mengchen Zhao. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Decision models based on sequence modeling have become prevalent in the field of offline reinforcement learning. However, existing approaches such as Decision Transformer and Trajectory Transformer suffer from poor data efficiency. One important reason is that they fail to extract useful information from potentially high-dimensional and noisy states. To resolve this issue, we propose a data-efficient decision sequence modeling method called Data Efficient Decision Sequence Model (DEDS), which dynamically identifies and filters out task-irrelevant state features for more compact state representations. Specifically, DEDS employs a state mask model guided by the bisimulation metric to ensure that only the most task-relevant state features are preserved for decision-making. Extensive experiments on various environments demonstrate that DEDS outperforms existing offline RL methods and achieves a significantly high data efficiency, especially in tasks with high-dimensional and complex state spaces.

Training autonomous agents to collaborate with unknown teammates in cooperative multi-agent environments remains a fundamental challenge in ad hoc teamwork research. Conventional approaches rely heavily on online interactions with arbitrary teammates under the assumption of full observability. However, in real-world scenarios, teammate policies are often inaccessible, making historical trajectory rollouts a more practical alternative. We propose DiTAC, a method that learns discrete teamwork abstractions for ad hoc collaboration by automatically extracting latent cooperation patterns from short trajectory segments and adapting effectively to diverse teammate behaviors. To mitigate the out-of-distribution challenge, we constrain learned representations within a discrete code-book. Furthermore, we employ a masked bidirectional transformer architecture to infer teammate behaviors from local observations, thereby relaxing the full observability assumption. Empirical results demonstrate that DiTAC significantly outperforms existing baselines and its variants across widely-used ad hoc teamwork tasks.

The ability of agents to collaborate with previously unknown teammates on the fly, known as ad hoc teamwork (AHT), is crucial in many real-world applications. Existing approaches to AHT require online interactions with the environment and some carefully designed teammates. However, these prerequisites can be infeasible in practice. In this work, we extend the AHT problem to the offline setting, where the policy of the ego agent is directly learned from a multi-agent interaction dataset. We propose a hierarchical sequence modeling framework called TAGET that addresses critical challenges in the offline setting, including limited data, partial observability and online adaptation. The core idea of TAGET is to dynamically predict teammate-aware rewards-to-go and sub-goals, so that the ego agent can adapt to the changes of teammates’ behaviors in real time. Extensive experimental results show that TAGET significantly outperforms existing solutions to AHT in the offline setting.

Recommendation models are predominantly trained using implicit user feedback, since explicit feedback is often costly to obtain. However, implicit feedback, such as clicks, does not always reflect users' real preferences. For example, a user might click on a news article because of its attractive headline, but end up feeling uncomfortable after reading the content. In the absence of explicit feedback, such erroneous implicit signals may severely mislead recommender systems. In this paper, we propose MTRec, a novel sequential recommendation framework designed to align with real user preferences by uncovering their internal satisfaction on recommended items. Specifically, we introduce a mental reward model to quantify user satisfaction and propose a distributional inverse reinforcement learning approach to learn it. The learned mental reward model is then used to guide recommendation models to better align with users’ real preferences. Our experiments show that MTRec brings significant improvements to a variety of recommendation models. We also deploy MTRec on an industrial short video platform and observe a 7\% increase in average user viewing time.

Multi-Domain Click-Through Rate (MDCTR) prediction is crucial for online recommendation platforms, which involves providing personalized recommendation services to users in different domains. However, current MDCTR models are confronted with the following limitations. Firstly, due to varying data sparsity in different domains, models can easily be dominated by some specific domains, which leads to significant performance degradation in other domains (i.e., the “seesaw phenomenon”). Secondly, when new domain emerges, the scalability of existing methods is limited, making it difficult to adapt to the dynamic growth of the domain. Traditional MDCTR models usually use one-hot encoding for semantic information such as product titles, thus losing rich semantic information and leading to insufficient generalization of the model. In this article, we propose a novel solution Uni-CTR to address these challenges. Uni-CTR leverages Large Language Model (LLM) to extract layer-wise semantic representations that capture domain commonalities, mitigating the seesaw phenomenon and enhancing generalization. Besides, it incorporates a pluggable domain-specific network to capture domain characteristics, ensuring scalability to dynamic domain growth. Experimental results on public datasets and industrial scenarios show that Uni-CTR significantly outperforms state-of-the-art (SOTA) models. In addition, Uni-CTR shows significant results in zero shot prediction. Code is available at Applied Machine Learning Lab (Pytorch), GitHub (Pytorch) and Gitee (MindSpore).

Learning representations to capture the very fundamental understanding of the world is a key challenge in machine learning. The hierarchical structure of explanatory factors hidden in data is such a general representation and could be potentially achieved with a hierarchical VAE. However, training a hierarchical VAE always suffers from the “posterior collapse”, where the data information is hard to propagate to the higher-level latent variables, hence resulting in a bad hierarchical representation. To address this issue, we first analyze the shortcomings of existing methods for mitigating the posterior collapse from an information theory perspective, then highlight the necessity of regularization for explicitly propagating data information to higher-level latent variables while maintaining the dependency between different levels. This naturally leads to formulating the inference of the hierarchical latent representation as a sequential decision process, which could benefit from applying reinforcement learning (RL). Aligning RL's objective with the regularization, we first introduce a skip-generative path to acquire a reward for evaluating the information content of an inferred latent representation, and then the developed Q-value function based on it could have a consistent optimization direction of the regularization. Finally, policy gradient, one of the typical RL methods, is employed to train a hierarchical VAE without introducing a gradient estimator. Experimental results firmly support our analysis and demonstrate that our proposed method effectively mitigates the posterior collapse issue, learns an informative hierarchy, acquires explainable latent representations, and significantly outperforms other hierarchical VAE-based methods in downstream tasks.

There is a growing demand of automatically synthesizing co-speech gestures for virtual characters. However, it remains a challenge due to the complex relationship between input speeches and target gestures. Most existing works focus on predicting the next gesture that fits the data best, however, such methods are myopic and lack the ability to plan for future gestures. In this paper, we propose a novel reinforcement learning (RL) framework called RACER to generate sequences of gestures that maximize the overall satisfactory. RACER employs a vector quantized variational autoencoder to learn compact representations of gestures and a GPT-based policy architecture to generate coherent sequence of gestures autoregressively. In particular, we propose a contrastive pre-training approach to calculate the rewards, which integrates contextual information into action evaluation and successfully captures the complex relationships between multi-modal speech-gesture data. Experimental results show that our method significantly outperforms existing baselines in terms of both objective metrics and subjective human judgements. Demos can be found at https://github.com/RLracer/RACER.git.

Offline Reinforcement Learning (RL) has demonstrated promising results in various applications by learning policies from previously collected datasets, reducing the need for online exploration and interactions. However, real-world scenarios usually involve partial observability, which brings crucial challenges of the deployment of offline RL methods: i) the policy trained on data with full observability is not robust against the masked observations during execution, and ii) the information of which parts of observations are masked is usually unknown during training. In order to address these challenges, we present Offline RL with DiscrEte pRoxy representations (ORDER), a probabilistic framework which leverages novel state representations to improve the robustness against diverse masked observabilities. Specifically, we propose a discrete representation of the states and use a proxy representation to recover the states from masked partial observable trajectories. The training of ORDER can be compactly described as the following three steps. i) Learning the discrete state representations on data with full observations, ii) Training the decision module based on the discrete representations, and iii) Training the proxy discrete representations on the data with various partial observations, aligning with the discrete representations. We conduct extensive experiments to evaluate ORDER, showcasing its effectiveness in offline RL for diverse partially observable scenarios and highlighting the significance of discrete proxy representations in generalization performance. ORDER is a flexible framework to employ any offline RL algorithms and we hope that ORDER can pave the way for the deployment of RL policy against various partial observabilities in the real world.

In single-agent Markov decision processes, an agent can optimize its policy based on the interaction with the environment. In multiplayer Markov games (MGs), however, the interaction is nonstationary due to the behaviors of other players, so the agent has no fixed optimization objective. The challenge becomes finding equilibrium policies for all players. In this research, we treat the evolution of player policies as a dynamical process and propose a novel learning scheme for Nash equilibrium. The core is to evolve one's policy according to not just its current in-game performance, but an aggregation of its performance over history. We show that for a variety of MGs, players in our learning scheme will provably converge to a point that is an approximation to Nash equilibrium. Combined with neural networks, we develop an empirical policy optimization algorithm, which is implemented in a reinforcement-learning framework and runs in a distributed way, with each player optimizing its policy based on own observations. We use two numerical examples to validate the convergence property on small-scale MGs, and a pong example to show the potential on large games.

Communicating agents with each other in a distributed manner and behaving as a group are essential in multi-agent reinforcement learning. However, real-world multi-agent systems suffer from restrictions on limited bandwidth communication. If the bandwidth is fully occupied, some agents are not able to send messages promptly to others, causing decision delay and impairing cooperative effects. Recent related work has started to address the problem but still fails in maximally reducing the consumption of communication resources. In this article, we propose an event-triggered communication network (ETCNet) to enhance communication efficiency in multi-agent systems by communicating only when necessary. For different task requirements, two paradigms of the ETCNet framework, event-triggered sending network (ETSNet) and event-triggered receiving network (ETRNet), are proposed for learning efficient sending and receiving protocols, respectively. Leveraging the information theory, the limited bandwidth is translated to the penalty threshold of an event-triggered strategy, which determines whether an agent at each step participates in communication or not. Then, the design of the event-triggered strategy is formulated as a constrained Markov decision problem and reinforcement learning finds the feasible and optimal communication protocol that satisfies the limited bandwidth constraint. Experiments on typical multi-agent tasks demonstrate that ETCNet outperforms other methods in reducing bandwidth occupancy and still preserves the cooperative performance of multi-agent systems at the most.

This paper describes the ReprGesture entry to the Generation and Evaluation of Non-verbal Behaviour for Embodied Agents (GENEA) challenge 2022. The GENEA challenge provides the processed datasets and performs crowdsourced evaluations to compare the performance of different gesture generation systems. In this paper, we explore an automatic gesture generation system based on multimodal representation learning. We use WavLM features for audio, FastText features for text and position and rotation matrix features for gesture. Each modality is projected to two distinct subspaces: modality-invariant and modality-specific. To learn inter-modality-invariant commonalities and capture the characters of modality-specific representations, gradient reversal layer based adversarial classifier and modality reconstruction decoders are used during training. The gesture decoder generates proper gestures using all representations and features related to the rhythm in the audio. Our code, pre-trained models and demo are available at https://github.com/YoungSeng/ReprGesture.

Offline reinforcement learning is a promising approach for practical applications since it does not require interactions with real-world environments. However, existing offline RL methods only work well in environments with continuous or small discrete action spaces. In environments with large and discrete action spaces, such as recommender systems and dialogue systems, the performance of existing methods decreases drastically because they suffer from inaccurate value estimation for a large proportion of out-of-distribution (o.o.d.) actions. While recent works have demonstrated that online RL benefits from incorporating semantic information in action representations, unfortunately, they fail to learn reasonable relative distances between action representations, which is key to offline RL to reduce the influence of o.o.d. actions. This paper proposes an action representation learning framework for offline RL based on a pseudometric, which measures both the behavioral relation and the data-distributional relation between actions. We provide theoretical analysis on the continuity of the expected Q-values and the offline policy improvement using the learned action representations. Experimental results show that our methods significantly improve the performance of two typical offline RL methods in environments with large and discrete action spaces.

Aiming to alleviate data sparsity and cold-start problems of tradi- tional recommender systems, incorporating knowledge graphs (KGs) to supplement auxiliary information has recently gained considerable attention. Via unifying the KG with user-item interactions into a tripartite graph, recent works explore the graph topologies to learn the low-dimensional representations of users and items with rich semantics. These real-world tripartite graphs are usually scale-free, however, the intrinsic hierarchical graph structures of which are underemphasized in existing works, consequently, leading to suboptimal recommendation performance. To address this issue and provide more accurate recommendation, we propose a knowledge-aware recommendation method with Lorentz model of the hyperbolic geometry, namely Lorentzian Knowledge-enhanced Graph convolutional networks for Recommendation (LKGR). LKGR facilitates better modeling of scale-free tripartite graphs after the data unification. Specifically, we employ different information propagation strategies in the hyperbolic space to explicitly encode heterogeneous information from historical interactions and KGs. Additionally, our proposed knowledge-aware attention mechanism enables the model to automatically measure the information contribution, producing the coherent information aggregation in the hyperbolic space. Extensive experiments on three real-world benchmarks demonstrate that LKGR outperforms state-of-the-art methods by 3.6-15.3% of [email protected] on Top-K recommendation.

Practical recommender systems experience a cold-start problem when observed user-item interactions in the history are insufficient. Meta learning, especially gradient based one, can be adopted to tackle this problem by learning initial parameters of the model and thus allowing fast adaptation to a specific task from limited data examples. Though with significant performance improvement, it commonly suffers from two critical issues: the non-compatibility with mainstream industrial deployment and the heavy computational burdens, both due to the inner-loop gradient operation. These two issues make them hard to be applied in practical recommender systems. To enjoy the benefits of meta learning framework and mitigate these problems, we propose a recommendation framework called Contextual Modulation Meta Learning (CMML). CMML is composed of fully feed-forward operations so it is computationally efficient and completely compatible with the mainstream industrial deployment. CMML consists of three components, including a context encoder that can generate context embedding to represent a specific task, a hybrid context generator that aggregates specific user-item features with task-level context, and a contextual modulation network, which can modulate the recommendation model to adapt effectively. We validate our approach on both scenario-specific and user-specific cold-start setting on various real-world datasets, showing CMML can achieve comparable or even better performance with gradient based methods yet with higher computational efficiency and better interpretability.

With the advancement of large-scale coordinated attacks, the adversary is shifting away from traditional distributed denial of service (DDoS) attacks against servers to sophisticated DDoS attacks against Internet infrastructures. Link flooding attacks (LFAs) are such powerful attacks against Internet links. Employing network measurement techniques, the defender could detect the link under attack. However, given the large number of Internet links, the defender can only monitor a subset of the links simultaneously, whereas any link might be attacked. Therefore, it remains challenging to practically deploy detection methods. This paper addresses this challenge from a game-theoretic perspective, and proposes a randomized approach (like security patrolling) to optimize LFA detection strategies. Specifically, we formulate the LFA detection problem as a Stackelberg security game, and design randomized detection strategies in consideration of the adversary's behavior, where best and quantal response models are leveraged to characterize the adversary's behavior. We employ a series of techniques to solve the nonlinear and nonconvex NP-hard optimization problems for finding the equilibrium. The experimental results demonstrate the necessity of handling LFAs from a game-theoretic perspective and the effectiveness of our solutions. We believe our study is a significant step forward in formally understanding LFA detection strategies.

Fraud transactions are one of the major threats faced by online e-commerce platforms. Recently, deep learning based classifiers have been deployed to detect fraud transactions. Inspired by findings on adversarial examples, this paper is the first to analyze the vulnerability of deep fraud detector to slight perturbations on input transactions, which is very challenging since the sparsity and discretization of transaction data result in a non-convex discrete optimization. Inspired by the iterative Fast Gradient Sign Method (FGSM) for the L8 attack, we first propose the Iterative Fast Coordinate Method (IFCM) for discrete L1 and L2 attacks which is efficient to generate large amounts of instances with satisfactory effectiveness. We then provide two novel attack algorithms to solve the discrete optimization. The first one is the Augmented Iterative Search (AIS) algorithm, which repeatedly searches for effective “simple” perturbation. The second one is called the Rounded Relaxation with Reparameterization (R3), which rounds the solution obtained by solving a relaxed and unconstrained optimization problem with reparameterization tricks. Finally, we conduct extensive experimental evaluation on the deployed fraud detector in TaoBao, one of the largest e-commerce platforms in the world, with millions of real-world transactions. Results show that (i) The deployed detector is highly vulnerable to attacks as the average precision is decreased from nearly 90% to as low as 20% with little perturbations; (ii) Our proposed attacks significantly outperform the adaptions of the state-of-the-art attacks. (iii) The model trained with an adversarial training process is significantly robust against attacks and performs well on the unperturbed data.

It has been an open challenge for self-interested agents to make optimal sequential decisions in complex multiagent systems, where agents might achieve higher utility via collaboration. The Microsoft Malmo Collaborative AI Challenge (MCAC), which is designed to encourage research relating to various problems in Collaborative AI, takes the form of a Minecraft mini-game where players might work together to catch a pig or deviate from cooperation, for pursuing high scores to win the challenge. Various characteristics, such as complex interactions among agents, uncertainties, sequential decision making and limited learning trials all make it extremely challenging to find effective strategies. We present HogRider---the champion agent of MCAC in 2017 out of 81 teams from 26 countries. One key innovation of HogRider is a generalized agent type hypothesis framework to identify the behavior model of the other agents, which is demonstrated to be robust to observation uncertainty. On top of that, a second key innovation is a novel Q-learning approach to learn effective policies against each type of the collaborating agents. Various ideas are proposed to adapt traditional Q-learning to handle complexities in the challenge, including state-action abstraction to reduce problem scale, a warm start approach using human reasoning for addressing limited learning trials, and an active greedy strategy to balance exploitation-exploration. Challenge results show that HogRider outperforms all the other teams by a significant edge, in terms of both optimality and stability.

Multi-task learning (MTL) is a machine learning paradigm that improves the performance of each task by exploiting useful information contained in multiple related tasks. However, the relatedness of tasks can be exploited by attackers to launch data poisoning attacks, which has been demonstrated a big threat to single-task learning. In this paper, we provide the first study on the vulnerability of MTL. Specifically, we focus on multi-task relationship learning (MTRL) models, a popular subclass of MTL models where task relationships are quantized and are learned directly from training data. We formulate the problem of computing optimal poisoning attacks on MTRL as a bilevel program that is adaptive to arbitrary choice of target tasks and attacking tasks. We propose an efficient algorithm called PATOM for computing optimal attack strategies. PATOM leverages the optimality conditions of the subproblem of MTRL to compute the implicit gradients of the upper level objective function. Experimental results on real-world datasets show that MTRL models are very sensitive to poisoning attacks and the attacker can significantly degrade the performance of target tasks, by either directly poisoning the target tasks or indirectly poisoning the related tasks exploiting the task relatedness. We also found that the tasks being attacked are always strongly correlated, which provides a clue for defending against such attacks.

Conducting fraud transactions has become popular among e-commerce sellers to make their products favorable to the platform and buyers, which decreases the utilization efficiency of buyer impressions and jeopardizes the business environment. Fraud detection techniques are necessary but not enough for the platform since it is impossible to recognize all the fraud transactions. In this paper, we focus on improving the platform's impression allocation mechanism to maximize its profit and reduce the sellers' fraudulent behaviors simultaneously. First, we learn a seller behavior model to predict the sellers' fraudulent behaviors from the real-world data provided by one of the largest e-commerce company in the world. Then, we formulate the platform's impression allocation problem as a continuous Markov Decision Process (MDP) with unbounded action space. In order to make the action executable in practice and facilitate learning, we propose a novel deep reinforcement learning algorithm DDPG-ANP that introduces an action norm penalty to the reward function. Experimental results show that our algorithm significantly outperforms existing baselines in terms of scalability and solution quality.

Label contamination attack (LCA) is an important type of data poisoning attack where an attacker manipulates the labels of training data to make the learned model beneficial to him. Existing work on LCA assumes that the attacker has full knowledge of the victim learning model, whereas the victim model is usually a black-box to the attacker. In this paper, we develop a Projected Gradient Ascent (PGA) algorithm to compute LCAs on a family of empirical risk minimizations and show that an attack on one victim model can also be effective on other victim models. This makes it possible that the attacker designs an attack against a substitute model and transfers it to a black-box victim model. Based on the observation of the transferability, we develop a defense algorithm to identify the data points that are most likely to be attacked. Empirical studies show that PGA significantly outperforms existing baselines and linear learning models are better substitute models than nonlinear ones.

Highly targeted spear phishing attacks are increasingly common, and have been implicated in many major security breeches. Email filtering systems are the first line of defense against such attacks. These filters are typically configured with uniform thresholds for deciding whether or not to allow a message to be delivered to a user. However, users have very significant differences in both their susceptibility to phishing attacks as well as their access to critical information and credentials that can cause damage. Recent work has considered setting personalized thresholds for individual users based on a Stackelberg game model. We consider two important extensions of the previous model. First, in our model user values can be substitutable, modeling cases where multiple users provide access to the same information or credential. Second, we consider attackers who make sequential attack plans based on the outcome of previous attacks. Our analysis starts from scenarios where there is only one credential and then extends to more general scenarios with multiple credentials. For single-credential scenarios, we demonstrate that the optimal defense strategy can be found by solving a binary combinatorial optimization problem called PEDS. For multiple-credential scenarios, we formulate it as a bilevel optimization problem for finding the optimal defense strategy and then reduce it to a single level optimization problem called PEMS using complementary slackness conditions. Experimental results show that both PEDS and PEMS lead to significant higher defender utilities than two existing benchmarks in different parameter settings. Also, both PEDS and PEMS are more robust than the existing benchmarks considering uncertainties.