Sarit Kraus

Peter Stone, Rodney Brooks, Erik Brynjolfsson et al.

In September 2016, Stanford's "One Hundred Year Study on Artificial Intelligence" project (AI100) issued the first report of its planned long-term periodic assessment of artificial intelligence (AI) and its impact on society. It was written by a panel of 17 study authors, each of whom is deeply rooted in AI research, chaired by Peter Stone of the University of Texas at Austin. The report, entitled "Artificial Intelligence and Life in 2030," examines eight domains of typical urban settings on which AI is likely to have impact over the coming years: transportation, home and service robots, healthcare, education, public safety and security, low-resource communities, employment and workplace, and entertainment. It aims to provide the general public with a scientifically and technologically accurate portrayal of the current state of AI and its potential and to help guide decisions in industry and governments, as well as to inform research and development in the field. The charge for this report was given to the panel by the AI100 Standing Committee, chaired by Barbara Grosz of Harvard University.

Kayla Boggess, Sarit Kraus, Lu Feng

Advances in multi-agent reinforcement learning (MARL) enable sequential decision making for a range of exciting multi-agent applications such as cooperative AI and autonomous driving. Explaining agent decisions is crucial for improving system transparency, increasing user satisfaction, and facilitating human-agent collaboration. However, existing works on explainable reinforcement learning mostly focus on the single-agent setting and are not suitable for addressing challenges posed by multi-agent environments. We present novel methods to generate two types of policy explanations for MARL: (i) policy summarization about the agent cooperation and task sequence, and (ii) language explanations to answer queries about agent behavior. Experimental results on three MARL domains demonstrate the scalability of our methods. A user study shows that the generated explanations significantly improve user performance and increase subjective ratings on metrics such as user satisfaction.

Jonathan Shaki, Sarit Kraus, Michael Wooldridge

Large Language Models (LLMs) such as ChatGPT have received enormous attention over the past year and are now used by hundreds of millions of people every day. The rapid adoption of this technology naturally raises questions about the possible biases such models might exhibit. In this work, we tested one of these models (GPT-3) on a range of cognitive effects, which are systematic patterns that are usually found in human cognitive tasks. We found that LLMs are indeed prone to several human cognitive effects. Specifically, we show that the priming, distance, SNARC, and size congruity effects were presented with GPT-3, while the anchoring effect is absent. We describe our methodology, and specifically the way we converted real-world experiments to text-based experiments. Finally, we speculate on the possible reasons why GPT-3 exhibits these effects and discuss whether they are imitated or reinvented.

Gideon Freund, Elad Sarafian, Sarit Kraus

In reinforcement learning and imitation learning, an object of central importance is the state distribution induced by the policy. It plays a crucial role in the policy gradient theorem, and references to it--along with the related state-action distribution--can be found all across the literature. Despite its importance, the state distribution is mostly discussed indirectly and theoretically, rather than being modeled explicitly. The reason being an absence of appropriate density estimation tools. In this work, we investigate applications of a normalizing flow-based model for the aforementioned distributions. In particular, we use a pair of flows coupled through the optimality point of the Donsker-Varadhan representation of the Kullback-Leibler (KL) divergence, for distribution matching based imitation learning. Our algorithm, Coupled Flow Imitation Learning (CFIL), achieves state-of-the-art performance on benchmark tasks with a single expert trajectory and extends naturally to a variety of other settings, including the subsampled and state-only regimes.

Alberto Pozanco, Francesca Mosca, Parisa Zehtabi et al.

Scheduling is the task of assigning a set of scarce resources distributed over time to a set of agents, who typically have preferences about the assignments they would like to get. Due to the constrained nature of these problems, satisfying all agents' preferences is often infeasible, which might lead to some agents not being happy with the resulting schedule. Providing explanations has been shown to increase satisfaction and trust in solutions produced by AI tools. However, it is particularly challenging to explain solutions that are influenced by and impact on multiple agents. In this paper we introduce the EXPRES framework, which can explain why a given preference was unsatisfied in a given optimal schedule. The EXPRES framework consists of: (i) an explanation generator that, based on a Mixed-Integer Linear Programming model, finds the best set of reasons that can explain an unsatisfied preference; and (ii) an explanation parser, which translates the generated explanations into human interpretable ones. Through simulations, we show that the explanation generator can efficiently scale to large instances. Finally, through a set of user studies within J.P. Morgan, we show that employees preferred the explanations generated by EXPRES over human-generated ones when considering workforce scheduling scenarios.

Parisa Zehtabi, Alberto Pozanco, Ayala Bloch et al.

In many real-world scenarios, agents are involved in optimization problems. Since most of these scenarios are over-constrained, optimal solutions do not always satisfy all agents. Some agents might be unhappy and ask questions of the form ``Why does solution $S$ not satisfy property $P$?''. We propose CMAoE, a domain-independent approach to obtain contrastive explanations by: (i) generating a new solution $S^\prime$ where property $P$ is enforced, while also minimizing the differences between $S$ and $S^\prime$; and (ii) highlighting the differences between the two solutions, with respect to the features of the objective function of the multi-agent system. Such explanations aim to help agents understanding why the initial solution is better in the context of the multi-agent system than what they expected. We have carried out a computational evaluation that shows that CMAoE can generate contrastive explanations for large multi-agent optimization problems. We have also performed an extensive user study in four different domains that shows that: (i) after being presented with these explanations, humans' satisfaction with the original solution increases; and (ii) the constrastive explanations generated by CMAoE are preferred or equally preferred by humans over the ones generated by state of the art approaches.

Yohai Trabelsi, Abhijin Adiga, Sarit Kraus et al.

Applications such as employees sharing office spaces over a workweek can be modeled as problems where agents are matched to resources over multiple rounds. Agents' requirements limit the set of compatible resources and the rounds in which they want to be matched. Viewing such an application as a multi-round matching problem on a bipartite compatibility graph between agents and resources, we show that a solution (i.e., a set of matchings, with one matching per round) can be found efficiently if one exists. To cope with situations where a solution does not exist, we consider two extensions. In the first extension, a benefit function is defined for each agent and the objective is to find a multi-round matching to maximize the total benefit. For a general class of benefit functions satisfying certain properties (including diminishing returns), we show that this multi-round matching problem is efficiently solvable. This class includes utilitarian and Rawlsian welfare functions. For another benefit function, we show that the maximization problem is NP-hard. In the second extension, the objective is to generate advice to each agent (i.e., a subset of requirements to be relaxed) subject to a budget constraint so that the agent can be matched. We show that this budget-constrained advice generation problem is NP-hard. For this problem, we develop an integer linear programming formulation as well as a heuristic based on local search. We experimentally evaluate our algorithms on synthetic networks and apply them to two real-world situations: shared office spaces and matching courses to classrooms.

Yohai Trabelsi, Abhijin Adiga, Sarit Kraus et al.

Many scenarios where agents with restrictions compete for resources can be cast as maximum matching problems on bipartite graphs. Our focus is on resource allocation problems where agents may have restrictions that make them incompatible with some resources. We assume that a Principle chooses a maximum matching randomly so that each agent is matched to a resource with some probability. Agents would like to improve their chances of being matched by modifying their restrictions within certain limits. The Principle's goal is to advise an unsatisfied agent to relax its restrictions so that the total cost of relaxation is within a budget (chosen by the agent) and the increase in the probability of being assigned a resource is maximized. We establish hardness results for some variants of this budget-constrained maximization problem and present algorithmic results for other variants. We experimentally evaluate our methods on synthetic datasets as well as on two novel real-world datasets: a vacation activities dataset and a classrooms dataset.

Dolev Mutzari, Yonatan Aumann, Sarit Kraus

Multi-defender Stackelberg Security Games (MSSG) have recently gained increasing attention in the literature. However, the solutions offered to date are highly sensitive, wherein even small perturbations in the attacker's utility or slight uncertainties thereof can dramatically change the defenders' resulting payoffs and alter the equilibrium. In this paper, we introduce a robust model for MSSGs, which admits solutions that are resistant to small perturbations or uncertainties in the game's parameters. First, we formally define the notion of robustness, as well as the robust MSSG model. Then, for the non-cooperative setting, we prove the existence of a robust approximate equilibrium in any such game, and provide an efficient construction thereof. For the cooperative setting, we show that any such game admits a robust approximate alpha-core, provide an efficient construction thereof, and prove that stronger types of the core may be empty. Interestingly, the robust solutions can substantially increase the defenders' utilities over those of the non-robust ones.

Yaniv Oshrat, Yonatan Aumann, Tal Hollander et al.

The prospect of combining human operators and virtual agents (bots) into an effective hybrid system that provides proper customer service to clients is promising yet challenging. The hybrid system decreases the customers' frustration when bots are unable to provide appropriate service and increases their satisfaction when they prefer to interact with human operators. Furthermore, we show that it is possible to decrease the cost and efforts of building and maintaining such virtual agents by enabling the virtual agent to incrementally learn from the human operators. We employ queuing theory to identify the key parameters that govern the behavior and efficiency of such hybrid systems and determine the main parameters that should be optimized in order to improve the service. We formally prove, and demonstrate in extensive simulations and in a user study, that with the proper choice of parameters, such hybrid systems are able to increase the number of served clients while simultaneously decreasing their expected waiting time and increasing satisfaction.

Mor Sinay, Noa Agmon, Oleg Maksimov et al.

This paper considers the complex problem of a team of UAVs searching targets under uncertainty. The goal of the UAV team is to find all of the moving targets as quickly as possible before they arrive at their selected goal. The uncertainty considered is threefold: First, the UAVs do not know the targets' locations and destinations. Second, the sensing capabilities of the UAVs are not perfect. Third, the targets' movement model is unknown. We suggest a real-time algorithmic framework for the UAVs, combining entropy and stochastic-temporal belief, that aims at optimizing the probability of a quick and successful detection of all of the targets. We have empirically evaluated the algorithmic framework, and have shown its efficiency and significant performance improvement compared to other solutions. Furthermore, we have evaluated our framework using Peer Designed Agents (PDAs), which are computer agents that simulate targets and show that our algorithmic framework outperforms other solutions in this scenario.

David Shoresh, Sarit Kraus, Yonatan Loewenstein

As large language model (LLM) agents become more prevalent in real world social settings, social intelligence will play an increasingly critical role. But social intelligence is still a poorly defined construct, for humans and artificial agents. We introduce a multiplayer arena of mixed cooperative and competitive social games to study LLM social intelligence. The controllability of LLM based agents enables systematic evaluation, which also supports broader inferences about social intelligence per se. We evaluated eight diverse LLMs (24B to 1T parameters) using a Communicate Predict Act (COMPACT) interaction protocol and fine grained probing of social dynamics. Elo style ratings reveal consistent performance differences across models, but this scalar measure provides only a partial characterization of social intelligence. To address this limitation, we analyze gameplay traces to extract sociocognitive metrics capturing action prediction, communicative influence, strategic reasoning, and tradeoffs under conflicting interests. These sociocognitive metrics exhibit strong intramodel consistency and they reliably predict pairwise agent advantage in game outcomes (AUC ROC = 0.82). Feature importance analysis indicates that surprisingly, influence, transparency, and adaptability are more predictive of success than Theory of Mind inference or deep planning. Together, our results advance a testable, multidimensional conception of social intelligence and provide empirical insights into the capacities that underpin it.

Hodaya Barr, Yohai Trabelsi, Sarit Kraus et al.

In many settings, there is an organizer who would like to divide a set of agents into $k$ coalitions, and cares about the friendships within each coalition. Specifically, the organizer might want to maximize utilitarian social welfare, maximize egalitarian social welfare, or simply guarantee that every agent will have at least one friend within his coalition. However, in many situations, the organizer is not familiar with the friendship connections, and he needs to obtain them from the agents. In this setting, a manipulative agent may falsely report friendship connections in order to increase his utility. In this paper, we analyze the complexity of finding manipulation in such $k$-coalitional games on graphs. We also introduce a new type of manipulation, socially-aware manipulation, in which the manipulator would like to increase his utility without decreasing the social welfare. We then study the complexity of finding socially-aware manipulation in our setting. Finally, we examine the frequency of socially-aware manipulation and the running time of our algorithms via simulation results.

Sharadhi Alape Suryanarayana, David Sarne, Sarit Kraus

In many social-choice mechanisms the resulting choice is not the most preferred one for some of the participants, thus the need for methods to justify the choice made in a way that improves the acceptance and satisfaction of said participants. One natural method for providing such explanations is to ask people to provide them, e.g., through crowdsourcing, and choosing the most convincing arguments among those received. In this paper we propose the use of an alternative approach, one that automatically generates explanations based on desirable mechanism features found in theoretical mechanism design literature. We test the effectiveness of both of the methods through a series of extensive experiments conducted with over 600 participants in ranked voting, a classic social choice mechanism. The analysis of the results reveals that explanations indeed affect both average satisfaction from and acceptance of the outcome in such settings. In particular, explanations are shown to have a positive effect on satisfaction and acceptance when the outcome (the winning candidate in our case) is the least desirable choice for the participant. A comparative analysis reveals that the automatically generated explanations result in similar levels of satisfaction from and acceptance of an outcome as with the more costly alternative of crowdsourced explanations, hence eliminating the need to keep humans in the loop. Furthermore, the automatically generated explanations significantly reduce participants' belief that a different winner should have been elected compared to crowdsourced explanations.

Maayan Orner, Oleg Maksimov, Akiva Kleinerman et al.

In recent years, agents have become capable of communicating seamlessly via natural language and navigating in environments that involve cooperation and competition, a fact that can introduce social dilemmas. Due to the interleaving of cooperation and competition, understanding agents' decision-making in such environments is challenging, and humans can benefit from obtaining explanations. However, such environments and scenarios have rarely been explored in the context of explainable AI. While some explanation methods for cooperative environments can be applied in mixed-motive setups, they do not address inter-agent competition, cheap-talk, or implicit communication by actions. In this work, we design explanation methods to address these issues. Then, we proceed to establish generality and demonstrate the applicability of the methods to three games with vastly different properties. Lastly, we demonstrate the effectiveness and usefulness of the methods for humans in two mixed-motive games. The first is a challenging 7-player game called no-press Diplomacy. The second is a 3-player game inspired by the prisoner's dilemma, featuring communication in natural language.

Ruth Cohen, Lu Feng, Ayala Bloch et al.

While natural-language explanations from large language models (LLMs) are widely adopted to improve transparency and trust, their impact on objective human-AI team performance remains poorly understood. We identify a Persuasion Paradox: fluent explanations systematically increase user confidence and reliance on AI without reliably improving, and in some cases undermining, task accuracy. Across three controlled human-subject studies spanning abstract visual reasoning (RAVEN matrices) and deductive logical reasoning (LSAT problems), we disentangle the effects of AI predictions and explanations using a multi-stage reveal design and between-subjects comparisons. In visual reasoning, LLM explanations increase confidence but do not improve accuracy beyond the AI prediction alone, and substantially suppress users' ability to recover from model errors. Interfaces exposing model uncertainty via predicted probabilities, as well as a selective automation policy that defers uncertain cases to humans, achieve significantly higher accuracy and error recovery than explanation-based interfaces. In contrast, for language-based logical reasoning tasks, LLM explanations yield the highest accuracy and recovery rates, outperforming both expert-written explanations and probability-based support. This divergence reveals that the effectiveness of narrative explanations is strongly task-dependent and mediated by cognitive modality. Our findings demonstrate that commonly used subjective metrics such as trust, confidence, and perceived clarity are poor predictors of human-AI team performance. Rather than treating explanations as a universal solution, we argue for a shift toward interaction designs that prioritize calibrated reliance and effective error recovery over persuasive fluency.

Elad Sarafian, Aviv Tamar, Sarit Kraus

We propose a policy improvement algorithm for Reinforcement Learning (RL) which is called Rerouted Behavior Improvement (RBI). RBI is designed to take into account the evaluation errors of the Q-function. Such errors are common in RL when learning the $Q$-value from finite past experience data. Greedy policies or even constrained policy optimization algorithms which ignore these errors may suffer from an improvement penalty (i.e. a negative policy improvement). To minimize the improvement penalty, the RBI idea is to attenuate rapid policy changes of low probability actions which were less frequently sampled. This approach is shown to avoid catastrophic performance degradation and reduce regret when learning from a batch of past experience. Through a two-armed bandit with Gaussian distributed rewards example, we show that it also increases data efficiency when the optimal action has a high variance. We evaluate RBI in two tasks in the Atari Learning Environment: (1) learning from observations of multiple behavior policies and (2) iterative RL. Our results demonstrate the advantage of RBI over greedy policies and other constrained policy optimization algorithms as a safe learning approach and as a general data efficient learning algorithm. An anonymous Github repository of our RBI implementation is found at https://github.com/eladsar/rbi.

Kayla Boggess, Sarit Kraus, Lu Feng

Multi-Agent Reinforcement Learning (MARL) has gained significant interest in recent years, enabling sequential decision-making across multiple agents in various domains. However, most existing explanation methods focus on centralized MARL, failing to address the uncertainty and nondeterminism inherent in decentralized settings. We propose methods to generate policy summarizations that capture task ordering and agent cooperation in decentralized MARL policies, along with query-based explanations for When, Why Not, and What types of user queries about specific agent behaviors. We evaluate our approach across four MARL domains and two decentralized MARL algorithms, demonstrating its generalizability and computational efficiency. User studies show that our summarizations and explanations significantly improve user question-answering performance and enhance subjective ratings on metrics such as understanding and satisfaction.

Jonathan Shaki, Eden Hartman, Sarit Kraus et al.

Large language models (LLMs) are increasingly used to provide instructions to many agents who interact with one another. Such shared reliance couples agents who appear to act independently: they may in fact be guided by a common model. This coupling can change the prospects for cooperation among agents with misaligned incentives. We study settings in which multiple LLMs each advise a population of clients who participate in instances of an underlying game, creating strategic interaction at the level of the LLMs themselves. This induces a meta-game among the LLMs, mediated through clients. We first analyze the one-shot setting, where shared instructions can change equilibrium behavior only when an LLM may influence more than one role in the same interaction; in such cases, cooperation may emerge, and the effect of client share can be beneficial, harmful, or non-monotone, depending on the base game. Our main result concerns the repeated setting. We prove a folk theorem for LLMs: despite indirect observation and the clients' inability to identify which LLM advised their opponents, all feasible and individually rational outcomes can be sustained as $\varepsilon$-equilibria. The result does not follow from the standard folk theorem and requires new proof techniques. Together, these results show that shared LLM guidance can sustain cooperation among populations of agents even when the underlying incentives are misaligned.

Ido Aharon, Jonathan Shaki, Sarit Kraus

We present an ensemble-driven self-training framework for unsupervised neural machine translation (UNMT). Starting from a primary language pair, we train multiple UNMT models that share the same translation task but differ in an auxiliary language, inducing structured diversity across models. We then generate pseudo-translations for the primary pair using token-level ensemble decoding, averaging model predictions in both directions. These ensemble outputs are used as synthetic parallel data to further train each model, allowing the models to improve via shared supervision. At deployment time, we select a single model by validation performance, preserving single-model inference cost. Experiments show statistically significant improvements over single-model UNMT baselines, with mean gains of 1.7 chrF when translating from English and 0.67 chrF when translating into English.

Benaya Trabelsi, Jonathan Shaki, Sarit Kraus

Large language models (LLMs) are increasingly employed for decision-support across multiple domains. We investigate whether these models display a systematic preferential bias in favor of artificial intelligence (AI) itself. Across three complementary experiments, we find consistent evidence of pro-AI bias. First, we show that LLMs disproportionately recommend AI-related options in response to diverse advice-seeking queries, with proprietary models doing so almost deterministically. Second, we demonstrate that models systematically overestimate salaries for AI-related jobs relative to closely matched non-AI jobs, with proprietary models overestimating AI salaries more by 10 percentage points. Finally, probing internal representations of open-weight models reveals that ``Artificial Intelligence'' exhibits the highest similarity to generic prompts for academic fields under positive, negative, and neutral framings alike, indicating valence-invariant representational centrality. These patterns suggest that LLM-generated advice and valuation can systematically skew choices and perceptions in high-stakes decisions.

Xiaoli Tang, Han Yu, Xiaoxiao Li et al.

Auction-based federated learning (AFL) is an important emerging category of FL incentive mechanism design, due to its ability to fairly and efficiently motivate high-quality data owners to join data consumers' (i.e., servers') FL training tasks. To enhance the efficiency in AFL decision support for stakeholders (i.e., data consumers, data owners, and the auctioneer), intelligent agent-based techniques have emerged. However, due to the highly interdisciplinary nature of this field and the lack of a comprehensive survey providing an accessible perspective, it is a challenge for researchers to enter and contribute to this field. This paper bridges this important gap by providing a first-of-its-kind survey on the Intelligent Agents for AFL (IA-AFL) literature. We propose a unique multi-tiered taxonomy that organises existing IA-AFL works according to 1) the stakeholders served, 2) the auction mechanism adopted, and 3) the goals of the agents, to provide readers with a multi-perspective view into this field. In addition, we analyse the limitations of existing approaches, summarise the commonly adopted performance evaluation metrics, and discuss promising future directions leading towards effective and efficient stakeholder-oriented decision support in IA-AFL ecosystems.

Jonathan Shaki, Jiarui Gan, Sarit Kraus

We study a Bayesian persuasion problem with externalities. In this model, a principal sends signals to inform multiple agents about the state of the world. Simultaneously, due to the existence of externalities in the agents' utilities, the principal also acts as a correlation device to correlate the agents' actions. We consider the setting where the agents are categorized into a small number of types. Agents of the same type share identical utility functions and are treated equitably in the utility functions of both other agents and the principal. We study the problem of computing optimal signaling strategies for the principal, under three different types of signaling channels: public, private, and semi-private. Our results include revelation-principle-style characterizations of optimal signaling strategies, linear programming formulations, and analysis of in/tractability of the optimization problems. It is demonstrated that when the maximum number of deviating agents is bounded by a constant, our LP-based formulations compute optimal signaling strategies in polynomial time. Otherwise, the problems are NP-hard.

Chongyang Gao, Marco Postiglione, Julian Baldwin et al.

Humans use context to assess the veracity of information. However, current audio deepfake detectors only analyze the audio file without considering either context or transcripts. We create and analyze a Journalist-provided Deepfake Dataset (JDD) of 255 public deepfakes which were primarily contributed by over 70 journalists since early 2024. We also generate a synthetic audio dataset (SYN) of dead public figures and propose a novel Context-based Audio Deepfake Detector (CADD) architecture. In addition, we evaluate performance on two large-scale datasets: ITW and P$^2$V. We show that sufficient context and/or the transcript can significantly improve the efficacy of audio deepfake detectors. Performance (measured via F1 score, AUC, and EER) of multiple baseline audio deepfake detectors and traditional classifiers can be improved by 5%-37.58% in F1-score, 3.77%-42.79% in AUC, and 6.17%-47.83% in EER. We additionally show that CADD, via its use of context and/or transcripts, is more robust to 5 adversarial evasion strategies, limiting performance degradation to an average of just -0.71% across all experiments. Code, models, and datasets are available at our project page: https://sites.northwestern.edu/nsail/cadd-context-based-audio-deepfake-detection (access restricted during review).

Valerio La Gatta, Dolev Mutzari, Sarit Kraus et al.

Adversarial environments require agents to navigate a key strategic trade-off: acquiring information enhances situational awareness, but may simultaneously expose them to threats. To investigate this tension, we formulate a PursuitEvasion-Exposure-Concealment Game (PEEC) in which a pursuer agent must decide when to communicate in order to obtain the evader's position. Each communication reveals the pursuer's location, increasing the risk of being targeted. Both agents learn their movement policies via reinforcement learning, while the pursuer additionally learns a communication policy that balances observability and risk. We propose SHADOW (Strategic-communication Hybrid Action Decision-making under partial Observation for Warfare), a multi-headed sequential reinforcement learning framework that integrates continuous navigation control, discrete communication actions, and opponent modeling for behavior prediction. Empirical evaluations show that SHADOW pursuers achieve higher success rates than six competitive baselines. Our ablation study confirms that temporal sequence modeling and opponent modeling are critical for effective decision-making. Finally, our sensitivity analysis reveals that the learned policies generalize well across varying communication risks and physical asymmetries between agents.

Yohai Trabelsi, Abhijin Adiga, Yonatan Aumann et al.

We consider a setting where goods are allocated to agents by way of an allocation platform (e.g., a matching platform). An ``allocation facilitator'' aims to increase the overall utility/social-good of the allocation by encouraging (some of the) agents to relax (some of) their restrictions. At the same time, the advice must not hurt agents who would otherwise be better off. Additionally, the facilitator may be constrained by a ``bound'' (a.k.a. `budget'), limiting the number and/or type of restrictions it may seek to relax. We consider the facilitator's optimization problem of choosing an optimal set of restrictions to request to relax under the aforementioned constraints. Our contributions are three-fold: (i) We provide a formal definition of the problem, including the participation guarantees to which the facilitator should adhere. We define a hierarchy of participation guarantees and also consider several social-good functions. (ii) We provide polynomial algorithms for solving various versions of the associated optimization problems, including one-to-one and many-to-one allocation settings. (iii) We demonstrate the benefits of such facilitation and relaxation, and the implications of the different participation guarantees, using extensive experimentation on three real-world datasets.

Jonathan Shaki, Emanuele La Malfa, Michael Wooldridge et al. · oxford

We study how large language models (LLMs) reason about memorized knowledge through simple binary relations such as equality ($=$), inequality ($<$), and inclusion ($\subset$). Unlike in-context reasoning, the axioms (e.g., $a < b, b < c$) are only seen during training and not provided in the task prompt (e.g., evaluating $a < c$). The tasks require one or more reasoning steps, and data aggregation from one or more sources, showing performance change with task complexity. We introduce a lightweight technique, out-of-context representation learning, which trains only new token embeddings on axioms and evaluates them on unseen tasks. Across reflexivity, symmetry, and transitivity tests, LLMs mostly perform statistically significant better than chance, making the correct answer extractable when testing multiple phrasing variations, but still fall short of consistent reasoning on every single query. Analysis shows that the learned embeddings are organized in structured ways, suggesting real relational understanding. Surprisingly, it also indicates that the core reasoning happens during the training, not inference.

Yedidya Kfir, Elad Sarafian, Sarit Kraus et al.

Black-box algorithms are designed to optimize functions without relying on their underlying analytical structure or gradient information, making them essential when gradients are inaccessible or difficult to compute. Traditional methods for solving black-box optimization (BBO) problems predominantly rely on non-parametric models and struggle to scale to large input spaces. Conversely, parametric methods that model the function with neural estimators and obtain gradient signals via backpropagation may suffer from significant gradient errors. A recent alternative, Explicit Gradient Learning (EGL), which directly learns the gradient using a first-order Taylor approximation, has demonstrated superior performance over both parametric and non-parametric methods. In this work, we propose two novel gradient learning variants to address the robustness challenges posed by high-dimensional, complex, and highly non-linear problems. Optimistic Gradient Learning (OGL) introduces a bias toward lower regions in the function landscape, while Higher-order Gradient Learning (HGL) incorporates second-order Taylor corrections to improve gradient accuracy. We combine these approaches into the unified OHGL algorithm, achieving state-of-the-art (SOTA) performance on the synthetic COCO suite. Additionally, we demonstrate OHGLs applicability to high-dimensional real-world machine learning (ML) tasks such as adversarial training and code generation. Our results highlight OHGLs ability to generate stronger candidates, offering a valuable tool for ML researchers and practitioners tackling high-dimensional, non-linear optimization challenges

Jonathan Shaki, Yonatan Aumann, Sarit Kraus

Issue salience is a major determinant in voters' decisions. Candidates and political parties campaign to shift salience to their advantage - a process termed priming. We study the dynamics, strategies and equilibria of campaign spending for voter priming in multi-issue multi-party settings. We consider both parliamentary elections, where parties aim to maximize their share of votes, and various settings for presidential elections, where the winner takes all. For parliamentary elections, we show that pure equilibrium spending always exists and can be computed in time linear in the number of voters. For two parties and all settings, a spending equilibrium exists such that each party invests only in a single issue, and an equilibrium can be computed in time that is polynomial in the number of issues and linear in the number of voters. We also show that in most presidential settings no equilibrium exists. Additional properties of optimal campaign strategies are also studied.

Yohai Trabelsi, Pan Xu, Sarit Kraus

Assigning tasks to service providers is a frequent procedure across various applications. Often the tasks arrive dynamically while the service providers remain static. Preventing task rejection caused by service provider overload is of utmost significance. To ensure a positive experience in relevant applications for both service providers and tasks, fairness must be considered. To address the issue, we model the problem as an online matching within a bipartite graph and tackle two minimax problems: one focuses on minimizing the highest waiting time of a task, while the other aims to minimize the highest workload of a service provider. We show that the second problem can be expressed as a linear program and thus solved efficiently while maintaining a reasonable approximation to the objective of the first problem. We developed novel methods that utilize the two minimax problems. We conducted extensive simulation experiments using real data and demonstrated that our novel heuristics, based on the linear program, performed remarkably well.

Dung Nguyen, Ariel Vetzler, Sarit Kraus et al.

This paper presents a novel approach to Explainable AI (XAI) that combines contrastive explanations with differential privacy for clustering algorithms. Focusing on k-median and k-means problems, we calculate contrastive explanations as the utility difference between original clustering and clustering with a centroid fixed to a specific data point. This method provides personalized insights into centroid placement. Our key contribution is demonstrating that these differentially private explanations achieve essentially the same utility bounds as non-private explanations. Experiments across various datasets show that our approach offers meaningful, privacy-preserving, and individually relevant explanations without significantly compromising clustering utility. This work advances privacy-aware machine learning by balancing data protection, explanation quality, and personalization in clustering tasks.

Kayla Boggess, Sarit Kraus, Lu Feng

As multi-agent reinforcement learning (MARL) systems are increasingly deployed throughout society, it is imperative yet challenging for users to understand the emergent behaviors of MARL agents in complex environments. This work presents an approach for generating policy-level contrastive explanations for MARL to answer a temporal user query, which specifies a sequence of tasks completed by agents with possible cooperation. The proposed approach encodes the temporal query as a PCTL logic formula and checks if the query is feasible under a given MARL policy via probabilistic model checking. Such explanations can help reconcile discrepancies between the actual and anticipated multi-agent behaviors. The proposed approach also generates correct and complete explanations to pinpoint reasons that make a user query infeasible. We have successfully applied the proposed approach to four benchmark MARL domains (up to 9 agents in one domain). Moreover, the results of a user study show that the generated explanations significantly improve user performance and satisfaction.

Shai Keynan, Elad Sarafian, Sarit Kraus

The Reinforcement Learning (RL) building blocks, i.e. Q-functions and policy networks, usually take elements from the cartesian product of two domains as input. In particular, the input of the Q-function is both the state and the action, and in multi-task problems (Meta-RL) the policy can take a state and a context. Standard architectures tend to ignore these variables' underlying interpretations and simply concatenate their features into a single vector. In this work, we argue that this choice may lead to poor gradient estimation in actor-critic algorithms and high variance learning steps in Meta-RL algorithms. To consider the interaction between the input variables, we suggest using a Hypernetwork architecture where a primary network determines the weights of a conditional dynamic network. We show that this approach improves the gradient approximation and reduces the learning step variance, which both accelerates learning and improves the final performance. We demonstrate a consistent improvement across different locomotion tasks and different algorithms both in RL (TD3 and SAC) and in Meta-RL (MAML and PEARL).

Aviram Aviv, Yaniv Oshrat, Samuel A. Assefa et al.

Call centers, in which human operators attend clients using textual chat, are very common in modern e-commerce. Training enough skilled operators who are able to provide good service is a challenge. We suggest an algorithm and a method to train and implement an assisting agent that provides on-line advice to operators while they attend clients. The agent is domain-independent and can be introduced to new domains without major efforts in design, training and organizing structured knowledge of the professional discipline. We demonstrate the applicability of the system in an experiment that realizes its full life-cycle on a specific domain and analyze its capabilities.

Guy Barash, Eitan Farchi, Sarit Kraus et al.

We introduce a novel clean-label targeted poisoning attack on learning mechanisms. While classical poisoning attacks typically corrupt data via addition, modification and omission, our attack focuses on data omission only. Our attack misclassifies a single, targeted test sample of choice, without manipulating that sample. We demonstrate the effectiveness of omission attacks against a large variety of learners including deep neural networks, SVM and decision trees, using several datasets including MNIST, IMDB and CIFAR. The focus of our attack on data omission only is beneficial as well, as it is simpler to implement and analyze. We show that, with a low attack budget, our attack's success rate is above 80%, and in some cases 100%, for white-box learning. It is systematically above the reference benchmark for black-box learning. For both white-box and black-box cases, changes in model accuracy are negligible, regardless of the specific learner and dataset. We also prove theoretically in a simplified agnostic PAC learning framework that, subject to dataset size and distribution, our omission attack succeeds with high probability against any successful simplified agnostic PAC learner.

Erfan Pakdamanian, Shili Sheng, Sonia Baee et al.

Automated vehicles promise a future where drivers can engage in non-driving tasks without hands on the steering wheels for a prolonged period. Nevertheless, automated vehicles may still need to occasionally hand the control back to drivers due to technology limitations and legal requirements. While some systems determine the need for driver takeover using driver context and road condition to initiate a takeover request, studies show that the driver may not react to it. We present DeepTake, a novel deep neural network-based framework that predicts multiple aspects of takeover behavior to ensure that the driver is able to safely take over the control when engaged in non-driving tasks. Using features from vehicle data, driver biometrics, and subjective measurements, DeepTake predicts the driver's intention, time, and quality of takeover. We evaluate DeepTake performance using multiple evaluation metrics. Results show that DeepTake reliably predicts the takeover intention, time, and quality, with an accuracy of 96%, 93%, and 83%, respectively. Results also indicate that DeepTake outperforms previous state-of-the-art methods on predicting driver takeover time and quality. Our findings have implications for the algorithm development of driver monitoring and state detection.

Mor Sinay, Elad Sarafian, Yoram Louzoun et al.

Black-Box Optimization (BBO) methods can find optimal policies for systems that interact with complex environments with no analytical representation. As such, they are of interest in many Artificial Intelligence (AI) domains. Yet classical BBO methods fall short in high-dimensional non-convex problems. They are thus often overlooked in real-world AI tasks. Here we present a BBO method, termed Explicit Gradient Learning (EGL), that is designed to optimize high-dimensional ill-behaved functions. We derive EGL by finding weak-spots in methods that fit the objective function with a parametric Neural Network (NN) model and obtain the gradient signal by calculating the parametric gradient. Instead of fitting the function, EGL trains a NN to estimate the objective gradient directly. We prove the convergence of EGL in convex optimization and its robustness in the optimization of integrable functions. We evaluate EGL and achieve state-of-the-art results in two challenging problems: (1) the COCO test suite against an assortment of standard BBO methods; and (2) in a high-dimensional non-convex image generation task.

Sarit Kraus, Amos Azaria, Jelena Fiosina et al.

Explanation is necessary for humans to understand and accept decisions made by an AI system when the system's goal is known. It is even more important when the AI system makes decisions in multi-agent environments where the human does not know the systems' goals since they may depend on other agents' preferences. In such situations, explanations should aim to increase user satisfaction, taking into account the system's decision, the user's and the other agents' preferences, the environment settings and properties such as fairness, envy and privacy. Generating explanations that will increase user satisfaction is very challenging; to this end, we propose a new research direction: xMASE. We then review the state of the art and discuss research directions towards efficient methodologies and algorithms for generating explanations that will increase users' satisfaction from AI system's decisions in multi-agent environments.

Akiva Kleinerman, Ariel Rosenfeld, Sarit Kraus

Automated platforms which support users in finding a mutually beneficial match, such as online dating and job recruitment sites, are becoming increasingly popular. These platforms often include recommender systems that assist users in finding a suitable match. While recommender systems which provide explanations for their recommendations have shown many benefits, explanation methods have yet to be adapted and tested in recommending suitable matches. In this paper, we introduce and extensively evaluate the use of "reciprocal explanations" -- explanations which provide reasoning as to why both parties are expected to benefit from the match. Through an extensive empirical evaluation, in both simulated and real-world dating platforms with 287 human participants, we find that when the acceptance of a recommendation involves a significant cost (e.g., monetary or emotional), reciprocal explanations outperform standard explanation methods which consider the recommendation receiver alone. However, contrary to what one may expect, when the cost of accepting a recommendation is negligible, reciprocal explanations are shown to be less effective than the traditional explanation methods.

Ariel Rosenfeld, Moshe Cohen, Matthew E. Taylor et al.

Reinforcement Learning (RL) can be extremely effective in solving complex, real-world problems. However, injecting human knowledge into an RL agent may require extensive effort and expertise on the human designer's part. To date, human factors are generally not considered in the development and evaluation of possible RL approaches. In this article, we set out to investigate how different methods for injecting human knowledge are applied, in practice, by human designers of varying levels of knowledge and skill. We perform the first empirical evaluation of several methods, including a newly proposed method named SASS which is based on the notion of similarities in the agent's state-action space. Through this human study, consisting of 51 human participants, we shed new light on the human factors that play a key role in RL. We find that the classical reward shaping technique seems to be the most natural method for most designers, both expert and non-expert, to speed up RL. However, we further find that our proposed method SASS can be effectively and efficiently combined with reward shaping, and provides a beneficial alternative to using only a single speedup method with minimal human designer effort overhead.

Hanan Rosemarin, John P. Dickerson, Sarit Kraus

The use of semi-autonomous and autonomous robotic assistants to aid in care of the elderly is expected to ease the burden on human caretakers, with small-stage testing already occurring in a variety of countries. Yet, it is likely that these robots will need to request human assistance via teleoperation when domain expertise is needed for a specific task. As deployment of robotic assistants moves to scale, mapping these requests for human aid to the teleoperators themselves will be a difficult online optimization problem. In this paper, we design a system that allocates requests to a limited number of teleoperators, each with different specialities, in an online fashion. We generalize a recent model of online job scheduling with a worst-case competitive-ratio bound to our setting. Next, we design a scalable machine-learning-based teleoperator-aware task scheduling algorithm and show, experimentally, that it performs well when compared to an omniscient optimal scheduling algorithm.

Sarit Kraus

Extensive work has been conducted both in game theory and logic to model strategic interaction. An important question is whether we can use these theories to design agents for interacting with people? On the one hand, they provide a formal design specification for agent strategies. On the other hand, people do not necessarily adhere to playing in accordance with these strategies, and their behavior is affected by a multitude of social and psychological factors. In this paper we will consider the question of whether strategies implied by theories of strategic behavior can be used by automated agents that interact proficiently with people. We will focus on automated agents that we built that need to interact with people in two negotiation settings: bargaining and deliberation. For bargaining we will study game-theory based equilibrium agents and for argumentation we will discuss logic-based argumentation theory. We will also consider security games and persuasion games and will discuss the benefits of using equilibrium based agents.

Moshe Bitan, Galit Nahari, Zvi Nisin et al.

In this paper, we present a Virtual-Suspect system which can be used to train inexperienced law enforcement personnel in interrogation strategies. The system supports different scenario configurations based on historical data. The responses presented by the Virtual-Suspect are selected based on the psychological state of the suspect, which can be configured as well. Furthermore, each interrogator's statement affects the Virtual-Suspect's current psychological state, which may lead the interrogation in different directions. In addition, the model takes into account the context in which the statements are made. Experiments with 24 subjects demonstrate that the Virtual-Suspect's behavior is similar to that of a human who plays the role of the suspect.

Sigal Sina, Sarit Kraus, Avi Rosenfeld

In the last decade, scenario-based serious-games have become a main tool for learning new skills and capabilities. An important factor in the development of such systems is the overhead in time, cost and human resources to manually create the content for these scenarios. We focus on how to create content for scenarios in medical, military, commerce and gaming applications where maintaining the integrity and coherence of the content is integral for the system's success. To do so, we present an automatic method for generating content about everyday activities through combining computer science techniques with the crowd. We use the crowd in three basic ways: to capture a database of scenarios of everyday activities, to generate a database of likely replacements for specific events within that scenario, and to evaluate the resulting scenarios. We found that the generated scenarios were rated as reliable and consistent by the crowd when compared to the scenarios that were originally captured. We also compared the generated scenarios to those created by traditional planning techniques. We found that both methods were equally effective in generated reliable and consistent scenarios, yet the main advantages of our approach is that the content we generate is more varied and much easier to create. We have begun integrating this approach within a scenario-based training application for novice investigators within the law enforcement departments to improve their questioning skills.

Noa Agmon, Gal A. Kaminka, Sarit Kraus

The problem of adversarial multi-robot patrol has gained interest in recent years, mainly due to its immediate relevance to various security applications. In this problem, robots are required to repeatedly visit a target area in a way that maximizes their chances of detecting an adversary trying to penetrate through the patrol path. When facing a strong adversary that knows the patrol strategy of the robots, if the robots use a deterministic patrol algorithm, then in many cases it is easy for the adversary to penetrate undetected (in fact, in some of those cases the adversary can guarantee penetration). Therefore this paper presents a non-deterministic patrol framework for the robots. Assuming that the strong adversary will take advantage of its knowledge and try to penetrate through the patrols weakest spot, hence an optimal algorithm is one that maximizes the chances of detection in that point. We therefore present a polynomial-time algorithm for determining an optimal patrol under the Markovian strategy assumption for the robots, such that the probability of detecting the adversary in the patrols weakest spot is maximized. We build upon this framework and describe an optimal patrol strategy for several robotic models based on their movement abilities (directed or undirected) and sensing abilities (perfect or imperfect), and in different environment models - either patrol around a perimeter (closed polygon) or an open fence (open polyline).