While knowledge distillation has seen widespread use in pre-training and instruction tuning, its application to aligning language models with human preferences remains underexplored, particularly in the more realistic cross-tokenizer setting. The incompatibility of tokenization schemes between teacher and student models has largely prevented fine-grained, white-box distillation of preference information. To address this gap, we propose Cross-Tokenizer Preference Distillation (CTPD), the first unified framework for transferring human-aligned behavior between models with heterogeneous tokenizers. CTPD introduces three key innovations: (1) Aligned Span Projection, which maps teacher and student tokens to shared character-level spans for precise supervision transfer; (2) a cross-tokenizer adaptation of Token-level Importance Sampling (TIS-DPO) for improved credit assignment; and (3) a Teacher-Anchored Reference, allowing the student to directly leverage the teacher’s preferences in a DPO-style objective. Our theoretical analysis grounds CTPD in importance sampling, and experiments across multiple benchmarks confirm its effectiveness, with significant performance gains over existing methods. These results establish CTPD as a practical and general solution for preference distillation across diverse tokenization schemes, opening the door to more accessible and efficient alignment of language models.

Pham Khanh Chi, Quoc Phong Dao, Thuat Nguyen, Linh Ngo Van, Trung Le, Thanh Hong Nguyen. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Offline reinforcement learning (RL) has garnered significant attention for its ability to learn effective policies from pre-collected datasets without the need for further environmental interactions. While promising results have been demonstrated in single-agent settings, offline multi-agent reinforcement learning (MARL) presents additional challenges due to the large joint state-action space and the complexity of multi-agent behaviors. A key issue in offline RL is the distributional shift, which arises when the target policy being optimized deviates from the behavior policy that generated the data. This problem is exacerbated in MARL due to the interdependence between agents' local policies and the expansive joint state-action space. Prior approaches have primarily addressed this challenge by incorporating regularization in the space of either Q-functions or policies. In this work, we propose a novel type of regularizer in the space of stationary distributions to address the distributional shift more effectively. Our algorithm, ComaDICE, provides a principled framework for offline cooperative MARL to correct the stationary distribution of the global policy, which is then leveraged to derive local policies for individual agents. Through extensive experiments on the offline multi-agent MuJoCo and StarCraft II benchmarks, we demonstrate that ComaDICE achieves superior performance compared to state-of-the-art offline MARL methods across nearly all tasks.

Inferring reward functions from demonstrations is a key challenge in reinforcement learning (RL), particularly in multi-agent RL (MARL). The large joint state-action spaces and intricate inter-agent interactions in MARL make inferring the joint reward function especially challenging. While prior studies in single-agent settings have explored ways to recover reward functions and expert policies from human preference feedback, such studies in MARL remain limited. Existing methods typically combine two separate stages, supervised reward learning, and standard MARL algorithms, leading to unstable training processes. In this work, we exploit the inherent connection between reward functions and Q functions in cooperative MARL to introduce a novel end-to-end preference-based learning framework. Our framework is supported by a carefully designed multi-agent value decomposition strategy that enhances training efficiency. Extensive experiments on two state-of-the-art benchmarks, SMAC and MAMuJoCo, using preference data generated by both rule-based and large language model approaches demonstrate that our algorithm consistently outperforms existing methods across various tasks.

We study offline imitation learning (IL) in cooperative multi-agent settings, where demonstrations have unlabeled mixed quality - containing both expert and suboptimal trajectories. Our proposed solution is structured in two stages: trajectory labeling and multi-agent imitation learning, designed jointly to enable effective learning from heterogeneous, unlabeled data. In the first stage, we combine advances in large language models and preference-based reinforcement learning to construct a progressive labeling pipeline that distinguishes expert-quality trajectories. In the second stage, we introduce MisoDICE, a novel multi-agent IL algorithm that leverages these labels to learn robust policies while addressing the computational complexity of large joint state-action spaces. By extending the popular single-agent DICE framework to multi-agent settings with a new value decomposition and mixing architecture, our method yields a convex policy optimization objective and ensures consistency between global and local policies. We evaluate MisoDICE on multiple standard multi-agent RL benchmarks and demonstrate superior performance, especially when expert data is scarce.

Vision-Language Foundation Models (VLMs), trained on large-scale multimodal datasets, have driven significant advances in Artificial Intelligence (AI) by enabling rich cross-modal reasoning. Despite their success in general domains, applying these models to medical imaging remains challenging due to the limited availability of diverse imaging modalities and multilingual clinical data. Most existing medical VLMs are trained on a subset of imaging modalities and focus primarily on high-resource languages, thus limiting their generalizability and clinical utility. To address these limitations, we introduce a novel Vietnamese-language multimodal medical dataset consisting of 2,757 whole-body PET/CT volumes from independent patients and their corresponding full-length clinical reports. This dataset is designed to fill two pressing gaps in medical AI development: (1) the lack of PET/CT imaging data in existing VLMs training corpora, which hinders the development of models capable of handling functional imaging tasks; and (2) the underrepresentation of low-resource languages, particularly the Vietnamese language, in medical vision-language research. To the best of our knowledge, this is the first dataset to provide comprehensive PET/CT-report pairs in Vietnamese. We further introduce a training framework to enhance VLMs' learning, including data augmentation and expert-validated test sets. We conduct comprehensive experiments benchmarking state-of-the-art VLMs on downstream tasks, including medical report generation and visual question answering. The experimental results show that incorporating our dataset significantly improves the performance of existing VLMs. However, despite these advancements, the models still underperform on clinically critical criteria, particularly the diagnosis of lung cancer, indicating substantial room for future improvement. We believe this dataset and benchmark will serve as a pivotal step in advancing the development of more robust VLMs for medical imaging, particularly in low-resource languages, and improving their clinical relevance in Vietnamese healthcare.

This paper concerns imitation learning (IL) in cooperative multi-agent systems. The learning problem under consideration poses several challenges, characterized by high-dimensional state and action spaces and intricate inter-agent dependencies. In a single-agent setting, IL was shown to be done efficiently via an inverse soft-Q learning process. However, extending this framework to a multi-agent context introduces the need to simultaneously learn both local value functions to capture local observations and individual actions, and a joint value function for exploiting centralized learning. In this work, we introduce a new multi-agent IL algorithm designed to address these challenges. Our approach enables the centralized learning by leveraging mixing networks to aggregate decentralized Q functions. We further establish conditions for the mixing networks under which the multi-agent IL objective function exhibits convexity within the Q function space. We present extensive experiments conducted on some challenging multi-agent game environments, including an advanced version of the Star-Craft multi-agent challenge (SMACv2), which demonstrates the effectiveness of our algorithm.

Training agents in multi-agent games presents significant challenges due to their intricate nature. These challenges are exacerbated by dynamics influenced not only by the environment but also by strategies of opponents. Existing methods often struggle with slow convergence and instability. To address these challenges, we harness the potential of imitation learning (IL) to comprehend and anticipate actions of the opponents, aiming to mitigate uncertainties with respect to the game dynamics. Our key contributions include: (i) a new multi-agent IL model for predicting next moves of the opponents - our model works with hidden actions of opponents and local observations; (ii) a new multi-agent reinforcement learning (MARL) algorithm that combines our IL model and policy training into one single training process; and (iii) extensive experiments in three challenging game environments, including an advanced version of the Star-Craft multi-agent challenge (i.e., SMACv2). Experimental results show that our approach achieves superior performance compared to state-of-the-art MARL algorithms.

This paper studies the problem of multi-step manipulative attacks in Stackelberg security games, in which a clever attacker attempts to orchestrate its attacks over multiple time steps to mislead the defender's learning of the attacker's behavior. This attack manipulation eventually influences the defender's patrol strategy towards the attacker's benefit. Previous work along this line of research only focuses on one-shot games in which the defender learns the attacker's behavior and then designs a corresponding strategy only once. Our work, on the other hand, investigates the long-term impact of the attacker's manipulation in which current attack and defense choices of players determine the future learning and patrol planning of the defender. This paper has three key contributions. First, we introduce a new multi-step manipulative attack game model that captures the impact of sequential manipulative attacks carried out by the attacker over the entire time horizon. Second, we propose a new algorithm to compute an optimal manipulative attack plan for the attacker, which tackles the challenge of multiple connected optimization components involved in the computation across multiple time steps. Finally, we present extensive experimental results on the impact of such misleading attacks, showing a significant benefit for the attacker and loss for the defender.

Prior work has successfully incorporated optimization layers as the last layer in neural networks for various problems, thereby allowing joint learning and planning in one neural network forward pass. In this work, we identify a weakness in such a set-up where inputs to the optimization layer lead to undefined output of the neural network. Such undefined decision outputs can lead to possible catastrophic outcomes in critical real time applications. We show that an adversary can cause such failures by forcing rank deficiency on the matrix fed to the optimization layer which results in the optimization failing to produce a solution. We provide a defense for the failure cases by controlling the condition number of the input matrix. We study the problem in the settings of synthetic data, Jigsaw Sudoku, and in speed planning for autonomous driving. We show that our proposed defense effectively prevents the framework from failing with undefined output. Finally, we surface a number of edge cases which lead to serious bugs in popular optimization solvers which can be abused as well.

This work builds an effective AI-based message generation system for diabetes prevention in rural areas, where the diabetes rate has been increasing at an alarming rate. The messages contain information about diabetes causes and complications and the impact of nutrition and fitness on preventing diabetes. We propose to apply reinforcement learning (RL) to optimize our message selection policy over time, tailoring our messages to align with each individual participant's needs and preferences. We conduct an extensive field study in a large country in Asia which involves more than 1000 participants who are local villagers and they receive messages generated by our system, over a period of six months. Our analysis shows that with the use of AI, we can deliver significant improvements in the participants' diabetes-related knowledge, physical activity levels, and high-fat food avoidance, when compared to a static message set. Furthermore, we build a new neural network based behavior model to predict behavior changes of participants, trained on data collected during our study. By exploiting underlying characteristics of health-related behavior, we manage to significantly improve the prediction accuracy of our model compared to baselines.

This work presents CounterNet, a novel end-to-end learning framework which integrates Machine Learning (ML) model training and the generation of corresponding counterfactual (CF) explanations into a single end-to-end pipeline. Counterfactual explanations offer a contrastive case, i.e., they attempt to find the smallest modification to the feature values of an instance that changes the prediction of the ML model on that instance to a predefined output. Prior techniques for generating CF explanations suffer from two major limitations: (i) all of them are post-hoc methods designed for use with proprietary ML models --- as a result, their procedure for generating CF explanations is uninformed by the training of the ML model, which leads to misalignment between model predictions and explanations; and (ii) most of them rely on solving separate time-intensive optimization problems to find CF explanations for each input data point (which negatively impacts their runtime). This work makes a novel departure from the prevalent post-hoc paradigm (of generating CF explanations) by presenting CounterNet, an end-to-end learning framework which integrates predictive model training and the generation of counterfactual (CF) explanations into a single pipeline. Unlike post-hoc methods, CounterNet enables the optimization of the CF explanation generation only once together with the predictive model. We adopt a block-wise coordinate descent procedure which helps in effectively training CounterNet's network. Our extensive experiments on multiple real-world datasets show that CounterNet generates high-quality predictions, and consistently achieves 100% CF validity and low proximity scores (thereby achieving a well-balanced cost-invalidity trade-off) for any new input instance, and runs 3X faster than existing state-of-the-art baselines.

Many problems in Reinforcement Learning (RL) seek an optimal policy with large discrete multidimensional yet unordered action spaces; these include problems in randomized allocation of resources such as placements of multiple security resources and emergency response units, etc. A challenge in this setting is that the underlying action space is categorical (discrete and unordered) and large, for which existing RL methods do not perform well. Moreover, these problems require validity of the realized action (allocation); this validity constraint is often difficult to express compactly in a closed mathematical form. The allocation nature of the problem also prefers stochastic optimal policies, if one exists. In this work, we address these challenges by (1) applying a (state) conditional normalizing flow to compactly represent the stochastic policy — the compactness arises due to the network only producing one sampled action and the corresponding log probability of the action, which is then used by an actor-critic method; and (2) employing an invalid action rejection method (via a valid action oracle) to update the base policy. The action rejection is enabled by a modified policy gradient that we derive. Finally, we conduct extensive experiments to show the scalability of our approach compared to prior methods and the ability to enforce arbitrary state-conditional constraints on the support of the distribution of actions in any state.

Most algorithmic studies on multi-agent information design have focused on the restricted situation of optimal public signaling with no inter-agent externalities; only a few exceptions investigated special game classes such as zero-sum games and second-price auctions. This paper initiates the algorithmic information design of both public and private signaling in a fundamental class of games with negative externalities, i.e., atomic singleton congestion games, with a wide range of applications in scheduling, routing, and network design, etc.

This paper studies the problem of information design in a general security game setting in which multiple independent self-interested defenders attempt to provide protection simultaneously on the same set of important targets against an unknown attacker. A principal, who can be one of the defenders, has access to certain private information (i.e., attacker type) whereas other defenders do not. We investigate the question of how that principal, with additional private information, can influence the decisions of the defenders by partially and strategically revealing her information. We focus on the algorithmic study of information design for private signaling in this game setting. In particular, we develop a polynomial-time ellipsoid algorithm to compute an optimal private signaling scheme. Our key finding is that the separation oracle in the ellipsoid approach can be carefully reduced to bipartite matching. Furthermore, we introduce a compact representation of any ex-ante persuasive signaling schemes by exploiting intrinsic security resource allocation structures, enabling us to compute an optimal scheme significantly faster. Our experiment results show that by strategically revealing private information, the principal can significantly enhance the protection effectiveness on the targets.

Real-world security problems are generally characterized by uncertainty about attackers' preferences, behavior, or other characteristics. To handle such uncertainties, security agencies (defender) typically rely on historical attack data to build a behavior model of the attacker, and incorporate this model into generating an effective defense strategy. For example, in wildlife protection, rangers can collect poaching signs (e.g., snares) to learn the behavior of poachers. However, in the real-world, a clever attacker can manipulate its attacks to fool the learning algorithm of the defender towards its own benefit. Unfortunately, existing state-of-the-art algorithms for generating defense strategies are not equipped to handle such deceptive behavior by the attacker, and this could lead to arbitrary losses for the defender. To address these challenges, this paper investigates a basic deception strategy of the attacker, termed imitative behavior deception, in which the attacker intentionally pretends to follow a specific behavior model and consistently plays according to that model, in order to optimize its utility. We have three main contributions. First, built upon previous work on attacker-behavior modeling, we introduce new algorithms to compute an optimal imitative behavior deception strategy of the attacker. Second, we propose a novel game-theoretic counter-deception algorithm which determines effective defense strategies, taking into account the deceptive behavior of the attacker. Third, we conduct extensive experiments, which shows that under the attacker's deception, the defender accrues a significant loss whereas the attacker achieves a significant gain in utility. Our experimental results also demonstrate the impact of our counter-deception algorithm on substantially diminishing the attacker's deception.

Learning attacker behavior is an important research topic in security games as security agencies are often uncertain about attackers' decision making. Previous work has focused on developing various behavioral models of attackers based on historical attack data. However, a clever attacker can manipulate its attacks to fail such attack-driven learning, leading to ineffective defense strategies. We study attacker behavior deception with three main contributions. First, we propose a new model, named partial behavior deception model, in which there is a deceptive attacker (among multiple attackers) who controls a portion of attacks. Our model captures real-world security scenarios such as wildlife protection in which multiple poachers are present. Second, we introduce a new scalable algorithm, GAMBO, to compute an optimal deception strategy of the deceptive attacker. Our algorithm employs the projected gradient descent and uses the implicit function theorem for the computation of gradient. Third, we conduct a comprehensive set of experiments, showing a significant benefit for the attacker and loss for the defender due to attacker deception.

Allocating resources to defend targets from attack is often complicated by uncertainty about the attacker’s capabilities, objectives, or other underlying characteristics. In a repeated interaction setting, the defender can collect attack data over time to reduce this uncertainty and learn an effective defense. However, a clever attacker can manipulate the attack data to mislead the defender, influencing the learning process toward its own benefit. We investigate strategic deception on the part of an attacker with private type information, who interacts repeatedly with a defender. We present a detailed computation and analysis of both players’ optimal strategies given the attacker may play deceptively. Computational experiments illuminate conditions conducive to strategic deception, and quantify benefits to the attacker. By taking into account the attacker’s deception capacity, the defender can significantly mitigate loss from misleading attack actions.

We study the problem of allocating limited security countermeasures to protect network data from cyber-attacks, for scenarios modeled by Bayesian attack graphs. We consider multistage interactions between a network administrator and cybercriminals, formulated as a security game. This formulation is capable of representing security environments with significant dynamics and uncertainty and very large strategy spaces. We propose parameterized heuristic strategies for the attacker and defender and provide detailed analysis of their time complexity. Our heuristics exploit the topological structure of attack graphs and employ sampling methods to overcome the computational complexity in predicting opponent actions. Due to the complexity of the game, we employ a simulation-based approach and perform empirical game analysis over an enumerated set of heuristic strategies. Finally, we conduct experiments in various game settings to evaluate the performance of our heuristics in defending networks, in a manner that is robust to uncertainty about the security environment.

The conservation of key wildlife species such as tigers and elephants are threatened by poaching activities. In many conservation areas, foot patrols are conducted to prevent poaching but they may not be well-planned to make the best use of the limited patrolling resources. While prior work has introduced PAWS (Protection Assistant for Wildlife Security) as a game-theoretic decision aid to design effective foot patrol strategies to protect wildlife, the patrol routes generated by PAWS may be difficult to follow in areas with complex terrain. Subsequent research has worked on the significant evolution of PAWS, from an emerging application to a regularly deployed software. A key advance of the deployed version of PAWS is that it incorporates the complex terrain information and generates a strategy consisting of easy-to-follow routes. In this demonstration, we provide 1) a video introducing the PAWS system; 2) an interactive visualization of the patrol routes generated by PAWS in an example area with complex terrain; and 3) a machine-human competition in designing patrol strategy given complex terrain and animal distribution.

Poaching is a serious threat to the conservation of key species and whole ecosystems. While conducting foot patrols is the most commonly used approach in many countries to prevent poaching, such patrols often do not make the best use of limited patrolling resources. To remedy this situation, prior work introduced a novel emerging application called PAWS (Protection Assistant for Wildlife Security); PAWS was proposed as a game-theoretic (“security games”) decision aid to optimize the use of patrolling resources.This paper reports on PAWS’s significant evolution from a proposed decision aid to a regularly deployed application, reporting on the lessons from the first tests in Africa in Spring 2014, through its continued evolution since then, to current regular use in Southeast Asia and plans for future worldwide deployment. In this process, we have worked closely with two NGOs (Panthera and Rimba) and incorporated extensive feedback from professional patrolling teams. We outline key technical advances that lead to PAWS’s regular deployment: (i) incorporating complex topographic features, e.g., ridgelines, in generating patrol routes; (ii) handling uncertainties in species distribution (game theoretic payoffs); (iii) ensuring scalability for patrolling large-scale conservation areas with fine-grained guidance; and (iv) handling complex patrol scheduling constraints.

Stackelberg Security Games (SSG) have been widely applied for solving real-world security problems—with a significant research emphasis on modeling attackers’ behaviors to handle their bounded rationality. However, access to real-world data (used for learning an accurate behavioral model) is often limited, leading to uncertainty in attacker’s behaviors while modeling. This paper therefore focuses on addressing behavioral uncertainty in SSG with the following main contributions: 1) we present a new uncertainty game model that integrates uncertainty intervals into a behavioral model to capture behavioral uncertainty; 2) based on this game model, we propose a novel robust algorithm that approximately computes the defender’s optimal strategy in the worst-case scenario of uncertainty—with a bound guarantee on its solution quality.

Recent years have seen increasing interest in AI from outside the AI community. This is partly due to applications based on AI that have been used in real-world domains, for example, the successful deployment of game theory-based decision aids in security domains. This paper describes our teaching approach for introducing the AI concepts underlying security games to diverse audiences. We adapted a game-based research platform that served as a testbed for recent research advances in computational game theory into a set of interactive role-playing games. We guided learners in playing these games as part of our teaching strategy, which also included didactic instruction and interactive exercises on broader AI topics. We describe our experience in applying this teaching approach to diverse audiences, including students of an urban public high school, university undergraduates, and security domain experts who protect wildlife. We evaluate our approach based on results from the games and participant surveys.

Stackelberg security games (SSGs) have been deployed in a number of real-world domains. One key challenge in these applications is the assessment of attacker payoffs, which may not be perfectly known. Previous work has studied SSGs with uncertain payoffs modeled by interval uncertainty and provided maximin-based robust solutions. In contrast, in this work we propose the use of the less conservative minimax regret decision criterion for such payoff-uncertain SSGs and present the first algorithms for computing minimax regret for SSGs. We also address the challenge of preference elicitation, using minimax regret to develop the first elicitation strategies for SSGs. Experimental results validate the effectiveness of our approaches.

Many strategic actions carry a ‘contagious’ component beyond the immediate locale of the effort itself. Viral marketing and peacekeeping operations have both been observed to have a spreading effect. In this work, we use counterinsurgency as our illustrative domain. Defined as the effort to block the spread of support for an insurgency, such operations lack the manpower to defend the entire population and must focus on the opinions of a subset of local leaders. As past researchers of security resource allocation have done, we propose using game theory to develop such policies and model the interconnected network of leaders as a graph. Unlike this past work in security games, actions in these domains possess a probabilistic, non-local impact. To address this new class of security games, we combine recent research in influence-blocking maximization with a double oracle approach and create novel heuristic oracles to generate mixed strategies for a real-world leadership network from Afghanistan, synthetic leadership networks and scale-free graphs. We find that leadership networks that exhibit highly interconnected clusters can be solved equally well by our heuristic methods, but our more sophisticated heuristics outperform simpler ones in less interconnected scale-free graphs.

Recent deployments of Stackelberg security games (SSG) have led to two competing approaches to handle boundedly rational human adversaries: (1) integrating models of human (adversary) decision-making into the game-theoretic algorithms, and (2) applying robust optimization techniques that avoid adversary modeling. A recent algorithm (MATCH) based on the second approach was shown to outperform the leading modeling-based algorithm even in the presence of significant amount of data. Is there then any value in using human behavior models in solving SSGs? Through extensive experiments with 547 human subjects playing 11102 games in total, we emphatically answer the question in the affirmative, while providing the following key contributions: (i) we show that our algorithm, SU-BRQR, based on a novel integration of human behavior model with the subjective utility function, significantly outperforms both MATCH and its improvements; (ii) we are the first to present experimental results with security intelligence experts, and find that even though the experts are more rational than the Amazon Turk workers, SU-BRQR still outperforms an approach assuming perfect rationality (and to a more limited extent MATCH); (iii) we show the advantage of SU-BRQR in a new, large game setting and demonstrate that sufficient data enables it to improve its performance over MATCH.

Many strategic actions carry a ‘contagious’ component beyond the immediate locale of the effort itself. Viral marketing and peacekeeping operations have both been observed to have a spreading effect. In this work, we use counterinsurgency as our illustrative domain. Defined as the effort to block the spread of support for an insurgency, such operations lack the manpower to defend the entire population and must focus onthe opinions of a subset of local leaders. As past researchers of security resource allocation have done, we propose using game theory to develop such policies and model the interconnected network of leaders as a graph. Unlike this past work in security games, actions in these domains possess a probabilistic, non-local impact. To address this new class of security games, we combine recent research in influence blocking maximization with a double oracle approach and create novel heuristic oracles to generate mixed strategies for a real-world leadership network from Afghanistan, synthetic leadership networks, and a real social network. We find that leadership networks that exhibit highly interconnected clusters can be solved equally well by our heuristic methods, but our more sophisticated heuristics outperform simpler ones in less interconnected social networks.