@inproceedings{20668, abstract = {The Message Layer Security (MLS) protocol has recently been standardized by the IETF. MLS is a scalable secure group messaging protocol expected to run more efficiently compared to the Signal protocol at scale, while offering a similar level of strong security. Even though MLS has undergone extensive examination by researchers, the majority of the works have focused on confidentiality. In this work, we focus on the authenticity of the application messages exchanged in MLS. Currently, MLS authenticates every application message with an EdDSA signature and while manageable, the overhead is greatly amplified in the post-quantum setting as the NIST-recommended Dilithium signature results in a 40x increase in size. We view this as an invitation to explore new authentication modes that can be used instead. We start by taking a systematic view on how application messages are authenticated in MLS and categorize authenticity into four different security notions. We then propose several authentication modes, offering a range of different efficiency and security profiles. For instance, in one of our modes, COSMOS++, we replace signatures with one-time tokens and a MAC tag, offering roughly a 75x savings in the post-quantum communication overhead. While this comes at the cost of weakening security compared to the authentication mode used by MLS, the lower communication overhead seems to make it a worthwhile trade-off with security.}, author = {Hashimoto, Keitaro and Katsumata, Shuichi and Pascual Perez, Guillermo}, booktitle = {34th Usenix Security Symposium}, isbn = {9781939133526}, location = {Seattle, WA, USA}, pages = {6699--6716}, publisher = {Usenix Association}, title = {{Exploring how to authenticate application messages in MLS: More efficient, post-quantum, and anonymous blocklistable}}, year = {2025}, } @article{22774, abstract = {We consider the problem of certifying the feasibility of a (possibly degenerate) semidefinite programming problem represented by rational input data together with a numerical solution that is guaranteed to be sufficiently close to an exact solution of maximal rank. Our method constructs a polynomial system whose set of real solutions has an isolated point that satisfies the input program. An experimental comparison between the corresponding symbolic-numerical algorithm and the purely symbolic algorithm from [3, 5] shows that the hybrid algorithm is able to certify instances where the exact method fails.}, author = {Kolmogorov, Vladimir and Naldi, Simone and Zapata, Jeferson}, issn = {1932-2240}, journal = {ACM Communications in Computer Algebra}, number = {4}, pages = {147--152}, publisher = {ACM}, title = {{Certifying solutions of degenerate semidefinite programs}}, doi = {10.1145/3832046.3832063}, volume = {59}, year = {2025}, } @inproceedings{22826, abstract = {Large Language Models (LLMs) have demonstrated the ability to tackle increasingly complex tasks through advanced reasoning, long-form content generation, and tool use. Solving these tasks often involves long inference-time computations. In human problem solving, a common strategy to expedite work is collaboration: by dividing the problem into sub-tasks, exploring different strategies concurrently, etc. Recent research has shown that LLMs can also operate in parallel by implementing explicit cooperation frameworks, such as voting mechanisms or the explicit creation of independent sub-tasks that can be executed in parallel. However, each of these frameworks may not be suitable for all types of tasks, which can hinder their applicability. In this work, we propose a different design approach: we run LLM “workers” in parallel , allowing them to synchronize via a concurrently-updated attention cache and prompt these workers to decide how best to collaborate. Our approach allows the LLM instances to come up with their own collaboration strategy for the problem at hand, all the while “seeing” each other’s memory in the concurrent KV cache. We implement this approach via Hogwild! Inference: a parallel LLM inference engine where multiple instances of the same LLM run in parallel with the same attention cache, with “instant” access to each other’s memory.1 Hogwild! Inference takes advantage of Rotary Position Embeddings (RoPE) to avoid recomputation while improving parallel hardware utilization. We find that modern reasoning-capable LLMs can perform inference with shared Key-Value cache out of the box, without additional fine-tuning.}, author = {Rodionov, Gleb and Garipov, Roman and Shutova, Alina and Yakushev, George and Schultheis, Erik and Egiazarian, Vage and Sinitsin, Anton and Kuznedelev, Denis and Alistarh, Dan-Adrian}, booktitle = {39th Conference on Neural Information Processing Systems}, isbn = {9798331338275}, issn = {1049-5258}, location = {San Diego, CA, United States}, pages = {52014--52055}, publisher = {Neural Information Processing Systems Foundation}, title = {{Hogwild! Inference: Parallel LLM generation via concurrent attention}}, doi = {10.52202/085713-1551}, volume = {38}, year = {2025}, } @inproceedings{22825, abstract = {The empirical emergence of neural collapse—a surprising symmetry in the feature representations of the training data in the penultimate layer of deep neural networks—has spurred a line of theoretical research aimed at its understanding. However, existing work focuses on data-agnostic models or, when data structure is taken into account, it remains limited to multi-layer perceptrons. Our paper fills both these gaps by analyzing modern architectures in a data-aware regime: we prove that global optima of deep regularized transformers and residual networks (ResNets) with LayerNorm trained with cross entropy or mean squared error loss are approximately collapsed, and the approximation gets tighter as the depth grows. More generally, we formally reduce any end-to-end large-depth ResNet or transformer training into an equivalent unconstrained features model, thus justifying its wide use in the literature even beyond data-agnostic settings. Our theoretical results are supported by experiments on computer vision and language datasets showing that, as the depth grows, neural collapse indeed becomes more prominent.}, author = {Súkeník, Peter and Lampert, Christoph and Mondelli, Marco}, booktitle = {39th Conference on Neural Information Processing Systems}, isbn = {9798331338275}, issn = {1049-5258}, location = {San Diego, CA, United States}, pages = {48646--48677}, publisher = {Neural Information Processing Systems Foundation}, title = {{Neural collapse is globally optimal in deep regularized ResNets and transformers}}, doi = {10.52202/085713-1450}, volume = {38}, year = {2025}, } @inproceedings{22824, abstract = {We present new fast-rate PAC-Bayesian generalization bounds for multi-task and meta-learning in the unbalanced setting, i.e. when the tasks have training sets of different sizes, as is typically the case in real-world scenarios. Previously, only standard-rate bounds were known for this situation, while fast-rate bounds were limited to the setting where all training sets are of equal size. Our new bounds are numerically computable as well as interpretable, and we demonstrate their flexibility in handling a number of cases where they give stronger guarantees than previous bounds. Besides the bounds themselves, we also make conceptual contributions: we demonstrate that the unbalanced multi-task setting has different statistical properties than the balanced situation, specifically that proofs from the balanced situation do not carry over to the unbalanced setting. Additionally, we shed light on the fact that the unbalanced situation allows two meaningful definitions of multi-task risk, depending on whether all tasks should be considered equally important or if sample-rich tasks should receive more weight than samplepoor ones.}, author = {Zakerinia, Hossein and Lampert, Christoph}, booktitle = {39th Conference on Neural Information Processing Systems}, isbn = {9798331338275}, issn = {1049-5258}, location = {San Diego, CA, United States}, pages = {9062--9093}, publisher = {Neural Information Processing Systems Foundation}, title = {{Fast rate bounds for multi-task and meta-learning with different sample sizes}}, doi = {10.52202/085713-0278}, volume = {38}, year = {2025}, } @inproceedings{22830, abstract = {We introduce AutoJudge, a method that accelerates large language model (LLM) inference with task-specific lossy speculative decoding. Instead of matching the original model output distribution token-by-token, we identify the generated tokens that affect the downstream quality of the response, relaxing the distribution match guarantee so that the "unimportant" tokens can be generated faster. Our approach relies on a semi‑greedy search algorithm to test which of the mismatches between target and draft models should be corrected to preserve quality and which ones may be skipped. We then train a lightweight classifier based on existing LLM embeddings to predict, at inference time, which mismatching tokens can be safely accepted without compromising the final answer quality. We evaluate AutoJudge with multiple draft/target model pairs on mathematical reasoning and programming benchmarks, achieving significant speedups at the cost of a minor accuracy reduction. Notably, on GSM8K with the Llama 3.1 70B target model, our approach achieves up to ≈ 2 × speedup \textit{over speculative decoding} at the cost of a ≤ 1 % drop in accuracy. When applied to the LiveCodeBench benchmark, AutoJudge automatically detects programming-specific important tokens, accepting ≥ 25 tokens per speculation cycle at a 2 % drop in Pass@1. Our approach requires no human annotation and is easy to integrate with modern LLM inference frameworks.}, author = {Garipov, Roman and Velikonivtsev, Fedor and Ermakov, Ivan and Svirschevski, Ruslan and Egiazarian, Vage and Ryabinin, Max}, booktitle = {39th Annual Conference on Neural Information Processing Systems}, isbn = {9798331338275}, issn = {1049-5258}, location = {San Diego, CA, United States}, pages = {104904--104941}, publisher = {Neural Information Processing Systems Foundation}, title = {{AutoJudge: Judge decoding without manual annotation}}, doi = {10.52202/085713-3163}, volume = {38}, year = {2025}, } @article{19595, abstract = {We investigate the locality of magnetic response in polycyclic aromatic molecules using a novel deep-learning approach. Our method employs graph neural networks (GNNs) with a graph-of-rings representation to predict nucleus independent chemical shifts (NICS) in the space around the molecule. We train a series of models, each time reducing the size of the largest molecules used in training. The accuracy of prediction remains high (MAE < 0.5 ppm), even when training the model only on molecules with up to four rings, thus providing strong evidence for the locality of magnetic response. To overcome the known problem of generalization of GNNs, we implement a k-hop expansion strategy and succeed in achieving accurate predictions for molecules with up to 15 rings (almost 4 times the size of the largest training example). Our findings have implications for understanding the magnetic response in complex molecules and demonstrate a promising approach to overcoming GNN scalability limitations. Furthermore, the trained models enable rapid characterization, without the need for more expensive DFT calculations.}, author = {Davidson, Yair and Philipp, Aviad and Chakraborty, Sabyasachi and Bronstein, Alexander and Gershoni-Poranne, Renana}, issn = {1089-7690}, journal = {Journal of Chemical Physics}, number = {14}, publisher = {AIP Publishing}, title = {{How local is “local”? Deep learning reveals locality of the induced magnetic field of polycyclic aromatic hydrocarbons}}, doi = {10.1063/5.0257558}, volume = {162}, year = {2025}, } @article{17468, abstract = {Oxygen redox chemistry is central to life1 and many human-made technologies, such as in energy storage2,3,4. The large energy gain from oxygen redox reactions is often connected with the occurrence of harmful reactive oxygen species3,5,6. Key species are superoxide and the highly reactive singlet oxygen3,4,5,6,7, which may evolve from superoxide. However, the factors determining the formation of singlet oxygen, rather than the relatively unreactive triplet oxygen, are unknown. Here we report that the release of triplet or singlet oxygen is governed by individual Marcus normal and inverted region behaviour. We found that as the driving force for the reaction increases, the initially dominant evolution of triplet oxygen slows down, and singlet oxygen evolution becomes predominant with higher maximum kinetics. This behaviour also applies to the widely observed superoxide disproportionation, in which one superoxide is oxidized by another, in both non-aqueous and aqueous systems, with Lewis and Brønsted acidity controlling the driving forces. Singlet oxygen yields governed by these conditions are relevant, for example, in batteries or cellular organelles in which superoxide forms. Our findings suggest ways to understand and control spin states and kinetics in oxygen redox chemistry, with implications for fields, including life sciences, pure chemistry and energy storage.}, author = {Mondal, Soumyadip and Nguyen, Huyen T.K. and Hauschild, Robert and Freunberger, Stefan Alexander}, issn = {1476-4687}, journal = {Nature}, number = {8085}, pages = {601–605}, publisher = {Springer Nature}, title = {{Marcus kinetics control singlet and triplet oxygen evolving from superoxide}}, doi = {10.1038/s41586-025-09587-7}, volume = {646}, year = {2025}, } @inproceedings{19743, abstract = {The possibility of errors in human-engineered formal verification software, such as model checkers, poses a serious threat to the purpose of these tools. An established approach to mitigate this problem are certificates—lightweight, easy-to-check proofs of the verification results. In this paper, we develop novel certificates for model checking of Markov decision processes (MDPs) with quantitative reachability and expected reward properties. Our approach is conceptually simple and relies almost exclusively on elementary fixed point theory. Our certificates work for arbitrary finite MDPs and can be readily computed with little overhead using standard algorithms. We formalize the soundness of our certificates in Isabelle/HOL and provide a formally verified certificate checker. Moreover, we augment existing algorithms in the probabilistic model checker Storm with the ability to produce certificates and demonstrate practical applicability by conducting the first formal certification of the reference results in the Quantitative Verification Benchmark Set.}, author = {Chatterjee, Krishnendu and Quatmann, Tim and Schäffeler, Maximilian and Weininger, Maximilian and Winkler, Tobias and Zilken, Daniel}, booktitle = {31st International Conference on Tools and Algorithms for the Construction and Analysis of Systems}, isbn = {9783031906527}, issn = {1611-3349}, location = {Hamilton, ON, Canada}, pages = {130--151}, publisher = {Springer Nature}, title = {{Fixed point certificates for reachability and expected rewards in MDPs}}, doi = {10.1007/978-3-031-90653-4_7}, volume = {15697}, year = {2025}, } @inproceedings{19740, abstract = {Two standard models for probabilistic systems are Markov chains (MCs) and Markov decision processes (MDPs). Classic objectives for such probabilistic models for control and planning problems are reachability and stochastic shortest path. The widely studied algorithmic approach for these problems is the Value Iteration (VI) algorithm which iteratively applies local updates called Bellman updates. There are many practical approaches for VI in the literature but they all require exponentially many Bellman updates for MCs in the worst case. A preprocessing step is an algorithm that is discrete, graph-theoretical, and requires linear space. An important open question is whether, after a polynomial-time preprocessing, VI can be achieved with sub-exponentially many Bellman updates. In this work, we present a new approach for VI based on guessing values. Our theoretical contributions are twofold. First, for MCs, we present an almost-linear-time preprocessing algorithm after which, along with guessing values, VI requires only subexponentially many Bellman updates. Second, we present an improved analysis of the speed of convergence of VI for MDPs. Finally, we present a practical algorithm for MDPs based on our new approach. Experimental results show that our approach provides a considerable improvement over existing VI-based approaches on several benchmark examples from the literature.}, author = {Chatterjee, Krishnendu and Jafariraviz, Mahdi and Saona Urmeneta, Raimundo J and Svoboda, Jakub}, booktitle = {31st International Conference on Tools and Algorithms for the Construction and Analysis of Systems}, isbn = {9783031906527}, issn = {1611-3349}, location = {Hamilton, ON, Canada}, pages = {217--236}, publisher = {Springer Nature}, title = {{Value iteration with guessing for Markov chains and Markov decision processes}}, doi = {10.1007/978-3-031-90653-4_11}, volume = {15697}, year = {2025}, } @inproceedings{19744, abstract = {We consider the problem of refuting equivalence of probabilistic programs, i.e., the problem of proving that two probabilistic programs induce different output distributions. We study this problem in the context of programs with conditioning (i.e., with observe and score statements), where the output distribution is conditioned by the event that all the observe statements along a run evaluate to true, and where the probability densities of different runs may be updated via the score statements. Building on a recent work on programs without conditioning, we present a new equivalence refutation method for programs with conditioning. Our method is based on weighted restarting, a novel transformation of probabilistic programs with conditioning to the output equivalent probabilistic programs without conditioning that we introduce in this work. Our method is the first to be both a) fully automated, and b) providing provably correct answers. We demonstrate the applicability of our method on a set of programs from the probabilistic inference literature.}, author = {Chatterjee, Krishnendu and Kafshdar Goharshadi, Ehsan and Novotný, Petr and Zikelic, Dorde}, booktitle = {31st International Conference on Tools and Algorithms for the Construction and Analysis of Systems}, isbn = {9783031906527}, issn = {1611-3349}, location = {Hamilton, ON, Canada}, pages = {279--300}, publisher = {Springer Nature}, title = {{Refuting equivalence in probabilistic programs with conditioning}}, doi = {10.1007/978-3-031-90653-4_14}, volume = {15697}, year = {2025}, } @inproceedings{19667, abstract = {The problem of checking satisfiability of linear real arithmetic (LRA) and non-linear real arithmetic (NRA) formulas has broad applications, in particular, they are at the heart of logic-related applications such as logic for artificial intelligence, program analysis, etc. While there has been much work on checking satisfiability of unquantified LRA and NRA formulas, the problem of checking satisfiability of quantified LRA and NRA formulas remains a significant challenge. The main bottleneck in the existing methods is a computationally expensive quantifier elimination step. In this work, we propose a novel method for efficient quantifier elimination in quantified LRA and NRA formulas. We propose a template-based Skolemization approach, where we automatically synthesize linear/polynomial Skolem functions in order to eliminate quantifiers in the formula. The key technical ingredient in our approach are Positivstellensätze theorems from algebraic geometry, which allow for an efficient manipulation of polynomial inequalities. Our method offers a range of appealing theoretical properties combined with a strong practical performance. On the theory side, our method is sound, semi-complete, and runs in subexponential time and polynomial space, as opposed to existing sound and complete quantifier elimination methods that run in doubly-exponential time and at least exponential space. On the practical side, our experiments show superior performance compared to state of the art SMT solvers in terms of the number of solved instances and runtime, both on LRA and on NRA benchmarks.}, author = {Chatterjee, Krishnendu and Kafshdar Goharshadi, Ehsan and Karrabi, Mehrdad and Motwani, Harshit J. and Seeliger, Maximilian and Zikelic, Dorde}, booktitle = {Proceedings of the 39th AAAI Conference on Artificial Intelligence}, issn = {2374-3468}, location = {Philadelphia, PA, United States}, number = {11}, pages = {11158--11166}, publisher = {Association for the Advancement of Artificial Intelligence}, title = {{Quantified linear and polynomial arithmetic satisfiability via template-based skolemization}}, doi = {10.1609/aaai.v39i11.33213}, volume = {39}, year = {2025}, } @inproceedings{19669, abstract = {We consider a class of optimization problems defined by a system of linear equations with min and max operators. This class of optimization problems has been studied under restrictive conditions, such as, (C1) the halting or stability condition; (C2) the non-negative coefficients condition; (C3) the sum upto 1 condition; and (C4) the only min or only max operator condition. Several seminal results in the literature focus on special cases. For example, turn-based stochastic games correspond to conditions C2 and C3; and Markov decision process to conditions C2, C3, and C4. However, the systematic computational complexity study of all the cases has not been explored, which we address in this work. Some highlights of our results are: with conditions C2 and C4, and with conditions C3 and C4, the problem is NP-complete, whereas with condition C1 only, the problem is in UP intersects coUP. Finally, we establish the computational complexity of the decision problem of checking the respective conditions.}, author = {Chatterjee, Krishnendu and Luo, Ruichen and Saona Urmeneta, Raimundo J and Svoboda, Jakub}, booktitle = {Proceedings of the 39th AAAI Conference on Artificial Intelligence}, issn = {2374-3468}, location = {Philadelphia, PA, United States}, number = {11}, pages = {11150--11157}, publisher = {Association for the Advancement of Artificial Intelligence}, title = {{Linear equations with min and max operators: Computational complexity}}, doi = {10.1609/aaai.v39i11.33212}, volume = {39}, year = {2025}, } @misc{19771, abstract = {This artifact allows to review and reproduce the Isabelle proofs and practical experiments from the paper *Fixed Point Certificates for Reachability and Expected Rewards in MDPs*. The contents are two-fold: First, the artifact contains a formally verified certificate checker for the certificates presented in the paper. The formal Isabelle/HOL proofs of the background theory can be inspected, checked by Isabelle and the code extraction can be retraced. Second, the artifact contains a modified version of the model checking tool `Storm` with support for certificate generation. Together with the provided scripts and benchmark files, this allows to reproduce the experiments from the paper. An appropriate subset of the experiments is given to allow a review in a timely manner. In addition, original logfiles from our experiments are provided, allowing a detailed inspection. The package includes convenient installation scripts for [the TACAS 2023 VM](https://doi.org/10.5281/zenodo.7113223) (based on Ubuntu 22.04). A native installation on Linux or macOS systems (including the newer ARM-based machines) is also possible.}, author = {Chatterjee, Krishnendu and Quatmann, Tim and Schäffeler, Maximilian and Weininger, Maximilian and Winkler, Tobias and Zilken, Daniel}, publisher = {Zenodo}, title = {{Artifact: Fixed point certificates for reachability and expected rewards in MDPs}}, doi = {10.5281/ZENODO.14626585}, year = {2025}, } @article{19843, abstract = {Social dilemmas are collective-action problems where individual interests are at odds with group interests. Such dilemmas occur frequently at all scales of human interactions. When dealing with collective-action problems, people often act reciprocally. They adjust their behavior to match the previous behavior of the recipient. The literature distinguishes two kinds of reciprocity. According to direct reciprocity, individuals react to their immediate experiences with the recipient. They are more likely to cooperate if the recipient previously cooperated with them. According to indirect reciprocity, individuals react to the recipient’s general behavior, irrespectively of whether or not they benefited directly. In practice, the two kinds of reciprocity are often intertwined; people typically base their decisions on both direct experiences and indirect observations. Yet only recently have researchers begun to explore how the two kinds of reciprocity interact. So far, this research only addresses a single type of social dilemma, the donation game, where the effects of individual behaviors are independent. Instead, here we allow for all pairwise social dilemmas. By applying novel techniques to generalize the theory of zero-determinant strategies, we establish an important proof of principle: In all social dilemmas, socially optimal outcomes can be sustained as an equilibrium, using either direct or indirect reciprocity, or arbitrary mixtures thereof. These results neither require games to be repeated infinitely often, nor that individual opinions are synchronized. In this way, we considerably generalize the scope of models of reciprocity, and we build further bridges between the literatures on direct and indirect reciprocity.}, author = {Hübner, Valentin and Schmid, Laura and Hilbe, Christian and Chatterjee, Krishnendu}, issn = {2752-6542}, journal = {PNAS Nexus}, number = {5}, publisher = {Oxford University Press}, title = {{Stable strategies of direct and indirect reciprocity across all social dilemmas}}, doi = {10.1093/pnasnexus/pgaf154}, volume = {4}, year = {2025}, } @article{19074, abstract = {The public goods game is among the most studied metaphors of cooperation in groups. In this game, individuals can use their endowments to make contributions towards a good that benefits everyone. Each individual, however, is tempted to free-ride on the contributions of others. Herein, we study repeated public goods games among asymmetric players. Previous work has explored to which extent asymmetry allows for full cooperation, such that players contribute their full endowment each round. However, by design that work focusses on equilibria where individuals make the same contribution each round. Instead, here we consider players whose contributions along the equilibrium path can change from one round to the next. We do so for three different models – one without any budget constraints, one with endowment constraints, and one in which individuals can save their current endowment to be used in subsequent rounds. In each case, we explore two key quantities: the welfare and the resource efficiency that can be achieved in equilibrium. Welfare corresponds to the sum of all players’ payoffs. Resource efficiency relates this welfare to the total contributions made by the players. Compared to constant contribution sequences, we find that time-dependent contributions can improve resource efficiency across all three models. Moreover, they can improve the players’ welfare in the model with savings.}, author = {Hübner, Valentin and Hilbe, Christian and Staab, Manuel and Kleshnina, Maria and Chatterjee, Krishnendu}, issn = {2153-0793}, journal = {Dynamic Games and Applications}, pages = {1617--1645}, publisher = {Springer Nature}, title = {{Time-dependent strategies in repeated asymmetric public goods games}}, doi = {10.1007/s13235-025-00627-5}, volume = {15}, year = {2025}, } @inproceedings{20299, abstract = {Deterministic Markov Decision Processes (DMDPs) are a mathematical framework for decision-making where the outcomes and future possible actions are deterministically determined by the current action taken. DMDPs can be viewed as a finite directed weighted graph, where in each step, the controller chooses an outgoing edge. An objective is a measurable function on runs (or infinite trajectories) of the DMDP, and the value for an objective is the maximal cumulative reward (or weight) that the controller can guarantee. We consider the classical mean-payoff (aka limit-average) objective, which is a basic and fundamental objective. Howard's policy iteration algorithm is a popular method for solving DMDPs with mean-payoff objectives. Although Howard's algorithm performs well in practice, as experimental studies suggested, the best known upper bound is exponential and the current known lower bound is as follows: For the input size I, the algorithm requires (math formular) iterations, where (math formular) hides the poly-logarithmic factors, i.e., the current lower bound on iterations is sub-linear with respect to the input size. Our main result is an improved lower bound for this fundamental algorithm where we show that for the input size I, the algorithm requires (math formular) iterations.}, author = {Asadi, Ali and Chatterjee, Krishnendu and De Raaij, Jakob}, booktitle = {The 41st Conference on Uncertainty in Artificial Intelligence}, issn = {2640-3498}, location = {Rio de Janeiro, Brazil}, pages = {223--232}, publisher = {ML Research Press}, title = {{Lower bound on Howard policy iteration for deterministic Markov Decision Processes}}, volume = {286}, year = {2025}, } @inproceedings{20297, abstract = {A standard model that arises in several applications in sequential decision-making is partially observable Markov decision processes (POMDPs) where a decision-making agent interacts with an uncertain environment. A basic objective in POMDPs is the reachability objective, where given a target set of states, the goal is to eventually arrive at one of them. The limit-sure problem asks whether reachability can be ensured with probability arbitrarily close to 1. In general, the limit-sure reachability problem for POMDPs is undecidable. However, in many practical cases, the most relevant question is the existence of policies with a small amount of memory. In this work, we study the limit-sure reachability problem for POMDPs with a fixed amount of memory. We establish that the computational complexity of the problem is NP-complete.}, author = {Asadi, Ali and Chatterjee, Krishnendu and Saona Urmeneta, Raimundo J and Shafiee, Ali}, booktitle = {The 41st Conference on Uncertainty in Artificial Intelligence}, issn = {2640-3498}, location = {Rio de Janeiro, Brazil}, pages = {238--247}, publisher = {ML Research Press}, title = {{Limit-sure reachability for small memory policies in POMDPs is NP-complete}}, volume = {286}, year = {2025}, } @inproceedings{20296, abstract = {Learning-based systems are increasingly deployed across various domains, yet the complexity of traditional neural networks poses significant challenges for formal verification. Unlike conventional neural networks, learned Logic Gate Networks (LGNs) replace multiplications with Boolean logic gates, yielding a sparse, netlist-like architecture that is inherently more amenable to symbolic verification, while still delivering promising performance. In this paper, we introduce a SAT encoding for verifying global robustness and fairness in LGNs. We evaluate our method on five benchmark datasets, including a newly constructed 5-class variant, and find that LGNs are both verification-friendly and maintain strong predictive performance.}, author = {Kresse, Fabian and Yu, Zhengqi and Lampert, Christoph and Henzinger, Thomas A}, booktitle = {2nd International Conferenceon Neuro-Symbolic Systems}, issn = {2640-3498}, location = {Philadephia, PA, United States}, publisher = {ML Research Press}, title = {{Logic gate neural networks are good for verification}}, volume = {288}, year = {2025}, } @article{19508, abstract = {We consider random two-player zero-sum dynamic games with perfect information on a class of infinite directed graphs. Starting from a fixed vertex, the players take turns to move a token along the edges of the graph. Every vertex is assigned a payoff known in advance by both players. Every time the token visits a vertex, Player 2 pays Player 1 the corresponding payoff. We consider a distribution over such games by assigning i.i.d. payoffs to the vertices. On the one hand, for acyclic directed graphs of bounded degree and sub-exponential expansion, we show that, when the duration of the game tends to infinity, the value converges almost surely to a constant at an exponential rate dominated in terms of the expansion. On the other hand, for the infinite d-ary tree (that does not fall into the previous class of graphs), we show convergence at a double-exponential rate.}, author = {Attia, Luc and Lichev, Lyuben and Mitsche, Dieter and Saona Urmeneta, Raimundo J and Ziliotto, Bruno}, issn = {2153-0793}, journal = {Dynamic Games and Applications}, pages = {1517--1535}, publisher = {Springer Nature}, title = {{Random zero-sum dynamic games on infinite directed graphs}}, doi = {10.1007/s13235-025-00636-4}, volume = {15}, year = {2025}, }