Searcharxiv⌕ Search

arXiv subjects

Yiming Su

Publications and source records attributed to Yiming Su.

13 recordsLinked to original sources
arXiv2026

Specula: Scaling formal specifications for autonomous model checking of system code

Specula is a push-button agentic system that generates high-quality formal specifications for large, complex system code and uses the specifications for highly effective model checking and bug finding. Specula employs large language model (LLM) based coding agents to autonomously develop TLA+ specifications, including invariants that describe correctness properties of the target system and formal models that describe the system implementation with the right level of abstractions. Specula is fully autonomous and thus eliminates the barrier of applying formal methods to real-world system code (as in traditional human-centric approaches). Meanwhile, Specula addresses limitations of LLM-driven techniques like reward hacking and hallucinations through self-evolving loops that iteratively improve specification quality by enabling the agents to deepen their understanding of system code and its behaviors. We have used Specula to check 48 open-source system projects; Specula found 249 bugs including many deep bugs that are hard to find by existing approaches. Specula has been used by several companies and is maintained at https://github.com/specula-org/Specula.

cs.SE↗
arXiv2026

SREGym: A Live Benchmark for AI SRE Agents with High-Fidelity Failure Scenarios

AI agents are increasingly used to diagnose and mitigate failures in production systems, known as agentic Site Reliability Engineering (SRE). Current SRE benchmarks are limited to oversimplistic SRE tasks and are unfortunately hard to extend due to bespoke designs. We present SREGym, a high-fidelity benchmark for SRE agents. SREGym exposes a live system environment built atop real-world cloud-native system stacks, where high-fidelity failure scenarios are simulated through fault injectors. SREGym models the complexity of production environments by simulating (1) a wide range of faults at different layers, (2) various ambient noises, and (3) diverse failure modes such as metastable failures and correlated failures. SREGym is architected as a modular, extensible framework that orchestrates fault and noise injectors across stacks. SREGym currently includes 90 realistic, challenging SRE problems. We use SREGym to evaluate frontier agents and show that their capabilities varies significantly in addressing different kinds of failures, with up to 40% differences in end-to-end results. SREGym is actively maintained as an open-source project and has been used by researchers and practitioners.

cs.AI↗
arXiv2026

Neuro-Symbolic Verification on Instruction Following of LLMs

A fundamental problem of applying Large Language Models (LLMs) to important applications is that LLMs do not always follow instructions, and violations are often hard to observe or check. In LLM-based agentic workflows, such violations can propagate and amplify along reasoning chains, causing task failures and system incidents. This paper presents NSVIF, a neuro-symbolic framework for verifying whether an LLM's output follows the instructions used to prompt the LLM. NSVIF is a universal, general-purpose verifier; it makes no assumption about the instruction or the LLM. NSVIF formulates instruction-following verification as a constraint-satisfaction problem by modeling user instructions as constraints. NSVIF models both logical and semantic constraints; constraint solving is done by a unified solver that orchestrates logical reasoning and semantic analysis. To evaluate NSVIF, we develop VIFBENCH, a new benchmark for instruction-following verifiers with fine-grained data labels. Experiments show that NSVIF significantly outperforms LLM-based approaches and provides interpretable feedback. We also show that feedback from NSVIF helps improve LLMs' instruction-following capability without post-training.

cs.AI↗
arXiv2025

Multi-solitons to focusing mass-supercritical stochastic nonlinear Schr\"odinger equations

We consider the stochastic nonlinear Schr\"odinger equation driven by linear multiplicative noise in the mass-supercritical case. Given arbitrary $K$ solitary waves with distinct speeds, we construct stochastic multi-solitons pathwisely in the sense of controlled rough path, which behave asymptotically as the sum of the $K$ prescribed solitons as time tends to infinity. In contrast to the mass-(sub)critical case in \cite{RSZ23}, the linearized Schr\"odinger operator around the ground state has more unstable directions in the supercritical case. Our pathwise construction utilizes the rescaling approach and the modulation method in \cite{CMM11}. We derive the quantitative decay rates dictated by the noise for the unstable directions, as well as the modulation parameters and remainder in the geometrical decomposition. They are important to close the key bootstrap estimates and to implement topological arguments to control the unstable directions. As a result, the temporal convergence rate of stochastic multi-solitons, which can be of either exponential or polynomial type, is related closely to the spatial decay rate of the noise and reflects the noise impact on soliton dynamics.

math.PR↗
arXiv2025

SysMoBench: Evaluating AI on Formally Modeling Complex Real-World Systems

Formal models are essential to specifying large, complex computer systems and verifying their correctness, but are notoriously expensive to write and maintain. Recent advances in generative AI show promise in generating certain forms of specifications. However, existing work mostly targets small code, not complete systems. It is unclear whether AI can deal with realistic system artifacts, as this requires abstracting their complex behavioral properties into formal models. We present SysMoBench, a benchmark that evaluates AI's ability to formally model large, complex systems. We focus on concurrent and distributed systems, which are keystones of today's critical computing infrastructures, encompassing operating systems and cloud infrastructure. We use TLA+, the de facto specification language for concurrent and distributed systems, though the benchmark can be extended to other specification languages. We address the primary challenge of evaluating AI-generated models by automating metrics like syntactic and runtime correctness, conformance to system code, and invariant correctness. SysMoBench currently includes eleven diverse system artifacts: the Raft implementation of Etcd and Redis, the leader election of ZooKeeper, the Spinlock, Mutex, and Ringbuffer in Asterinas OS, etc., with more being added. SysMoBench enables us to understand the capabilities and limitations of today's LLMs and agents, putting tools in this area on a firm footing and opening up promising new research directions.

cs.AI↗
arXiv2025

STRATUS: A Multi-agent System for Autonomous Reliability Engineering of Modern Clouds

In cloud-scale systems, failures are the norm. A distributed computing cluster exhibits hundreds of machine failures and thousands of disk failures; software bugs and misconfigurations are reported to be more frequent. The demand for autonomous, AI-driven reliability engineering continues to grow, as existing humanin-the-loop practices can hardly keep up with the scale of modern clouds. This paper presents STRATUS, an LLM-based multi-agent system for realizing autonomous Site Reliability Engineering (SRE) of cloud services. STRATUS consists of multiple specialized agents (e.g., for failure detection, diagnosis, mitigation), organized in a state machine to assist system-level safety reasoning and enforcement. We formalize a key safety specification of agentic SRE systems like STRATUS, termed Transactional No-Regression (TNR), which enables safe exploration and iteration. We show that TNR can effectively improve autonomous failure mitigation. STRATUS significantly outperforms state-of-the-art SRE agents in terms of success rate of failure mitigation problems in AIOpsLab and ITBench (two SRE benchmark suites), by at least 1.5 times across various models. STRATUS shows a promising path toward practical deployment of agentic systems for cloud reliability.

cs.DC↗
arXiv2022

Construction of multi-bubble blow-up solutions to the $L^2$-critical half-wave equation

This paper concerns the bubbling phenomena for the $L^2$-critical half-wave equation in dimension one. Given arbitrarily finitely many distinct singularities, we construct blow-up solutions concentrating exactly at these singularities. This provides the first examples of multi-bubble solutions for the half-wave equation. In particular, the solutions exhibit the mass quantization property. Our proof strategy draws upon the modulation method in \cite{K-L-R} for the single-bubble case, and explores the localization techniques in \cite{CSZ21,RSZ21} for bubbling solutions to nonlinear Schr\"odinger equations (NLS). However, unlike the single-bubble or NLS cases, different bubbles exhibit the strongest interactions in dimension one. In order to get sharp estimates to control strong interactions, as well as nonlocal effects on localization functions, we utilize the Carlder\'on estimate and the integration representation formula of the half-wave operator, and find that there exists a narrow room between the orders $|t|^{2+}$ and $|t|^{3-}$ for the remainder in the geometrical decomposition. Based on this, a novel bootstrap scheme is introduced to address the multi-bubble non-local structure.

math.AP↗
arXiv2021

Multi solitary waves to stochastic nonlinear Schrödinger equations

In this paper, we present a pathwise construction of multi-soliton solutions for focusing stochastic nonlinear Schrödinger equations with linear multiplicative noise, in both the $L^2$-critical and subcritical cases. The constructed multi-solitons behave asymptotically as a sum of $K$ solitary waves, where $K$ is any given finite number. Moreover, the convergence rate of the remainders can be of either exponential or polynomial type, which reflects the effects of the noise in the system on the asymptotical behavior of the solutions. The major difficulty in our construction of stochastic multi-solitons is the absence of pseudo-conformal invariance. Unlike in the deterministic case [47,54], the existence of stochastic multi-solitons cannot be obtained from that of stochastic multi-bubble blow-up solutions in [54,57]. Our proof is mainly based on the rescaling approach in [39], relying on two types of Doss-Sussman transforms, and on the modulation method in [16,44], in which the crucial ingredient is the monotonicity of the Lyapunov type functional constructed by Martel, Merle and Tsai [45]. In our stochastic case, this functional depends on the Brownian paths in the noise.

math.PR↗
arXiv2021

Multi-bubble Bourgain-Wang solutions to nonlinear Schrödinger equation

We consider a general class of focusing $L^2$-critical nonlinear Schrödinger equations with lower order perturbations, for which the pseudo-conformal symmetry and the conservation law of energy are absent. In dimensions one and two, we construct Bourgain-Wang type solutions concentrating at $K$ distinct singularities, $1\leq K<\infty$, and prove that they are unique if the asymptotic behavior is within the order $(T-t)^{4+}$, for $t$ close to the blow-up time $T$. These results apply to the canonical nonlinear Schrödinger equations and, through the pseudo-conformal transform, in particular yield the existence and conditional uniqueness of non-pure multi-solitons. Furthermore, through a Doss-Sussman type transform, these results also apply to stochastic nonlinear Schrödinger equations, where the driving noise is taken in the sense of controlled rough path.

math.AP↗
arXiv2021

On uniqueness of multi-bubble blow-up solutions and multi-solitons to $L^2$-critical nonlinear Schrödinger equations

We are concerned with the focusing $L^2$-critical nonlinear Schrödinger equations in $\mathbb{R}^d$ for $d=1,2$. The uniqueness is proved for a large energy class of multi-bubble blow-up solutions, which converge to a sum of $K$ pseudo-conformal blow-up solutions particularly with low rate $(T-t)^{0+}$, as $t\to T$, $1\leq K<\infty$. Moreover, we also prove the uniqueness in the energy class of multi-solitons which converge to a sum of $K$ solitary waves with convergence rate $(1/t)^{2+}$, as $t\to \infty$. The uniqueness class is further enlarged to contain the multi-solitons with even lower convergence rate $(1/t)^{\frac 12+}$ in the pseudo-conformal space. The proof is mainly based on the pseudo-conformal invariance and the monotonicity properties of several functionals adapted to the multi-bubble case, the latter is crucial towards the upgradation of the convergence to the fast exponential decay rate.

math.AP↗
arXiv2020

On the multi-bubble blow-up solutions to rough nonlinear Schrödinger equations

We are concerned with the multi-bubble blow-up solutions to rough nonlinear Schrödinger equations in the focusing mass-critical case. In both dimensions one and two, we construct the finite time multi-bubble solutions, which concentrate at $K$ distinct points, $1\leq K<\infty$, and behave asymptotically like a sum of pseudo-conformal blow-up solutions in the pseudo-conformal space $Σ$ near the blow-up time. The upper bound of the asymptotic behavior is closely related to the flatness of noise at blow-up points. Moreover, we prove the conditional uniqueness of multi-bubble solutions in the case where the asymptotic behavior in the energy space $H^1$ is of the order $(T-t)^{3+ζ}$, $ζ>0$. These results are also obtained for nonlinear Schrödinger equations with lower order perturbations, particularly, in the absence of the classical pseudo-conformal symmetry and the conversation law of energy. The existence results are applicable to the canonical deterministic nonlinear Schrödinger equation and complement the previous work [43]. The conditional uniqueness results are new in both the stochastic and deterministic case.

math.PR↗
arXiv2020

Minimal mass blow-up solutions to rough nonlinear Schroedinger equations

We study the focusing mass-critical rough nonlinear Schroedinger equations, where the stochastic integration is taken in the sense of controlled rough path. We obtain the global well-posedness if the mass of initial data is below that of the ground state. Moreover, the existence of minimal mass blow-up solutions is also obtained in both dimensions one and two. In particular, these yield that the mass of ground state is exactly the threshold of global well-posedness and blow-up of solutions in the stochastic focusing mass-critical case. Similar results are also obtained for a class of nonlinear Schroedinger equations with lower order perturbations.

math.PR↗