Mahesh Viswanathan

Chuchu Fan, Bolun Qi, Sayan Mitra et al.

We present the DRYVR framework for verifying hybrid control systems that are described by a combination of a black-box simulator for trajectories and a white-box transition graph specifying mode switches. The framework includes (a) a probabilistic algorithm for learning sensitivity of the continuous trajectories from simulation data, (b) a bounded reachability analysis algorithm that uses the learned sensitivity, and (c) reasoning techniques based on simulation relations and sequential composition, that enable verification of complex systems under long switching sequences, from the reachability analysis of a simpler system under shorter sequences. We demonstrate the utility of the framework by verifying a suite of automotive benchmarks that include powertrain control, automatic transmission, and several autonomous and ADAS features like automatic emergency braking, lane-merge, and auto-passing controllers.

Andrea Morandi, Mahesh Viswanathan

[Abridged] Production LLM deployments receive feedback from a non-random fraction of users: thumbs sit mostly in the tails of the satisfaction distribution, and a naive average over them can land 40-50 percentage points away from true system quality. We treat this as a topic- and sentiment- stratified selection-bias problem and propose a three-agent hierarchical Bayesian pipeline that does not require ground-truth labels on individual interactions. A Topic Clustering Agent partitions the stream via UMAP + HDBSCAN over text embeddings; a Bias Modeling Agent fits a two-stage hierarchical Beta-Binomial under NUTS, inferring per-topic selection rates $s_c$ and quality $q_c$ with partial pooling; a Synthesis Agent reweights $q_c$ by true topic prevalence $\hatπ_c = n_c/N$ to report a bias-corrected aggregate posterior $\bar Q = \sum_c \hatπ_c q_c$ with credible interval, plus drift signals for online recalibration. Validation uses UltraFeedback (N=10,232 retained interactions, $C=18$ clusters, $Q^\star=0.6249$) with simulated topic- and sentiment-dependent selection biases. We compare five Bayesian variants against Naive and IPW baselines. A mild prior on the feedback channel (typical positive-feedback rate and negative-to-positive ratio, both readable from any production dashboard without labels) keeps Hierarchical-Informed within 4-13 pp of $Q^\star$ as the bias ratio sweeps from 1:1 to 30:1, with 95% credible intervals covering $Q^\star$ in 50/50 random-seed replicates at $κ_{\max}=10$. Without channel-side priors, every weak-prior variant misses $Q^\star$ by 22-33 pp: the per-cluster sufficient statistics admit a one-parameter family of equally good fits, and the prior on the bias channel (not on latent quality) is what breaks the degeneracy.

Kristina Miller, Christopher K. Zeitler, William Shen et al.

A runtime assurance system (RTA) for a given plant enables the exercise of an untrusted or experimental controller while assuring safety with a backup (or safety) controller. The relevant computational design problem is to create a logic that assures safety by switching to the safety controller as needed, while maximizing some performance criteria, such as the utilization of the untrusted controller. Existing RTA design strategies are well-known to be overly conservative and, in principle, can lead to safety violations. In this paper, we formulate the optimal RTA design problem and present a new approach for solving it. Our approach relies on reward shaping and reinforcement learning. It can guarantee safety and leverage machine learning technologies for scalability. We have implemented this algorithm and present experimental results comparing our approach with state-of-the-art reachability and simulation-based RTA approaches in a number of scenarios using aircraft models in 3D space with complex safety requirements. Our approach can guarantee safety while increasing utilization of the experimental controller over existing approaches.

Umang Mathur, Andreas Pavlogiannis, Hünkar Can Tunç et al.

Dynamic techniques are a scalable and effective way to analyze concurrent programs. Instead of analyzing all behaviors of a program, these techniques detect errors by focusing on a single program execution. Often a crucial step in these techniques is to define a causal ordering between events in the execution, which is then computed using vector clocks, a simple data structure that stores logical times of threads. The two basic operations of vector clocks, namely join and copy, require $Θ(k)$ time, where $k$ is the number of threads. Thus they are a computational bottleneck when $k$ is large. In this work, we introduce tree clocks, a new data structure that replaces vector clocks for computing causal orderings in program executions. Joining and copying tree clocks takes time that is roughly proportional to the number of entries being modified, and hence the two operations do not suffer the a-priori $Θ(k)$ cost per application. We show that when used to compute the classic happens-before (HB) partial order, tree clocks are optimal, in the sense that no other data structure can lead to smaller asymptotic running time. Moreover, we demonstrate that tree clocks can be used to compute other partial orders, such as schedulable-happens-before (SHB) and the standard Mazurkiewicz (MAZ) partial order, and thus are a versatile data structure. Our experiments show that just by replacing vector clocks with tree clocks, the computation becomes from $2.02 \times$ faster (MAZ) to $2.66 \times$ (SHB) and $2.97 \times$ (HB) on average per benchmark. These results illustrate that tree clocks have the potential to become a standard data structure with wide applications in concurrent analyses.

Seth Poulsen, Mahesh Viswanathan, Geoffrey L. Herman et al.

Proof Blocks is a software tool which enables students to write proofs by dragging and dropping prewritten proof lines into the correct order. These proofs can be graded completely automatically, enabling students to receive rapid feedback on how they are doing with their proofs. When constructing a problem, the instructor specifies the dependency graph of the lines of the proof, so that any correct arrangement of the lines can receive full credit. This innovation can improve assessment tools by increasing the types of questions we can ask students about proofs, and can give greater access to proof knowledge by increasing the amount that students can learn on their own with the help of a computer.

Rohit Chadha, A. Prasad Sistla, Mahesh Viswanathan

We introduce an automata model for describing interesting classes of differential privacy mechanisms/algorithms that include known mechanisms from the literature. These automata can model algorithms whose inputs can be an unbounded sequence of real-valued query answers. We consider the problem of checking whether there exists a constant $d$ such that the algorithm described by these automata are $dε$-differentially private for all positive values of the privacy budget parameter $ε$. We show that this problem can be decided in time linear in the automaton's size by identifying a necessary and sufficient condition on the underlying graph of the automaton. This paper's results are the first decidability results known for algorithms with an unbounded number of query answers taking values from the set of reals.

Yu Wang, Nima Roohi, Matthew West et al.

Probabilistic Computation Tree Logic (PCTL) is frequently used to formally specify control objectives such as probabilistic reachability and safety. In this work, we focus on model checking PCTL specifications statistically on Markov Decision Processes (MDPs) by sampling, e.g., checking whether there exists a feasible policy such that the probability of reaching certain goal states is greater than a threshold. We use reinforcement learning to search for such a feasible policy for PCTL specifications, and then develop a statistical model checking (SMC) method with provable guarantees on its error. Specifically, we first use upper-confidence-bound (UCB) based Q-learning to design an SMC algorithm for bounded-time PCTL specifications, and then extend this algorithm to unbounded-time specifications by identifying a proper truncation time by checking the PCTL specification and its negation at the same time. Finally, we evaluate the proposed method on case studies.

Gilles Barthe, Rohit Chadha, Vishal Jagannath et al.

Differential privacy is a de facto standard for statistical computations over databases that contain private data. The strength of differential privacy lies in a rigorous mathematical definition that guarantees individual privacy and yet allows for accurate statistical results. Thanks to its mathematical definition, differential privacy is also a natural target for formal analysis. A broad line of work uses logical methods for proving privacy. However, these methods are not complete, and only partially automated. A recent and complementary line of work uses statistical methods for finding privacy violations. However, the methods only provide statistical guarantees (but no proofs). We propose the first decision procedure for checking the differential privacy of a non-trivial class of probabilistic computations. Our procedure takes as input a program P parametrized by a privacy budget $ε$, and either proves differential privacy for all possible values of $ε$ or generates a counterexample. In addition, our procedure applies both to $ε$-differential privacy and $(ε,δ)$-differential privacy. Technically, the decision procedure is based on a novel and judicious encoding of the semantics of programs in our class into a decidable fragment of the first-order theory of the reals with exponentiation. We implement our procedure and use it for (dis)proving privacy bounds for many well-known examples, including randomized response, histogram, report noisy max and sparse vector.

Umang Mathur, Dileep Kini, Mahesh Viswanathan

Dynamic race detection is the problem of determining if an observed program execution reveals the presence of a data race in a program. The classical approach to solving this problem is to detect if there is a pair of conflicting memory accesses that are unordered by Lamport's happens-before (HB) relation. HB based race detection is known to not report false positives, i.e., it is sound. However, the soundness guarantee of HB only promises that the first pair of unordered, conflicting events is a schedulable data race. That is, there can be pairs of HB-unordered conflicting data accesses that are not schedulable races because there is no reordering of the events of the execution, where the events in race can be executed immediately after each other. We introduce a new partial order, called schedulable happens-before (SHB) that exactly characterizes the pairs of schedulable data races --- every pair of conflicting data accesses that are identified by SHB can be scheduled, and every HB-race that can be scheduled is identified by SHB. Thus, the SHB partial order is truly sound. We present a linear time, vector clock algorithm to detect schedulable races using SHB. Our experiments demonstrate the value of our algorithm for dynamic race detection --- SHB incurs only little performance overhead and can scale to executions from real-world software applications without compromising soundness.

Dileep Kini, Umang Mathur, Mahesh Viswanathan

We consider the problem of detecting data races in program traces that have been compressed using straight line programs (SLP), which are special context-free grammars that generate exactly one string, namely the trace that they represent. We consider two classical approaches to race detection --- using the happens-before relation and the lockset discipline. We present algorithms for both these methods that run in time that is linear in the size of the compressed, SLP representation. Typical program executions almost always exhibit patterns that lead to significant compression. Thus, our algorithms are expected to result in large speedups when compared with analyzing the uncompressed trace. Our experimental evaluation of these new algorithms on standard benchmarks confirms this observation.

Dileep Kini, Umang Mathur, Mahesh Viswanathan

Writing reliable concurrent software remains a huge challenge for today's programmers. Programmers rarely reason about their code by explicitly considering different possible inter-leavings of its execution. We consider the problem of detecting data races from individual executions in a sound manner. The classical approach to solving this problem has been to use Lamport's happens-before (HB) relation. Until now HB remains the only approach that runs in linear time. Previous efforts in improving over HB such as causally-precedes (CP) and maximal causal models fall short due to the fact that they are not implementable efficiently and hence have to compromise on their race detecting ability by limiting their techniques to bounded sized fragments of the execution. We present a new relation weak-causally-precedes (WCP) that is provably better than CP in terms of being able to detect more races, while still remaining sound. Moreover it admits a linear time algorithm which works on the entire execution without having to fragment it.