arXiv subjects

Kai Cai

Publications and source records attributed to Kai Cai.

At least 19 recordsLinked to original sources

arXiv2026

SWE-Bench ProMax: Benchmarking Agents on Large-Scale Multilingual Code Refactoring

As AI coding agents take on increasingly complex, long-horizon software engineering tasks, existing benchmarks are rapidly saturating and their evaluation quality has come under serious scrutiny: a recent audit found that nearly 60% of unsolved SWE-bench Verified instances contain flawed tests -- either overly narrow tests that reject correct solutions or overly broad tests that check unstated requirements -- and that frontier models can verbatim reproduce gold patches from training data. Code refactoring, which requires coordinated, behavior-preserving changes across many files, offers a substantially harder and more realistic test of agent capability, yet remains underserved by current benchmarks. We introduce SWE-Bench ProMax, an expert-curated, multilingual code refactoring benchmark of 170 instances drawn from real commits across seven programming languages (Python, Java, TypeScript, Go, C, C++, and Rust). Every instance undergoes rigorous, multi-stage curation that directly addresses the quality problems identified in prior benchmarks: issue descriptions are rewritten from scratch to provide precise, unambiguous specifications, and test suites are manually reviewed to remove overly narrow and overly broad tests. Tasks with insufficient complexity or limited cross-file scope are filtered out, yielding a benchmark of challenging, large-scale refactoring tasks that average 11.4 modified files and 261.6 lines of code per instance, substantially exceeding the scale of existing benchmarks. Experiments with frontier models under two agent scaffolds show that the best model achieves only 41.2% resolve rate, confirming that SWE-Bench ProMax presents a meaningful and unsaturated challenge for current AI coding agents. Our benchmark is available at https://huggingface.co/datasets/swe-bench-promax/SWE-Bench-ProMax.

cs.CL↗

arXiv2026

Supervisory Control with Event Forcing Under Partial Observation

In this paper, we address supervisory control in discrete-event systems with event forcing. In particular, partial observation of the supervisor is considered. Unlike traditional supervisory control, which relies on event enablement and disablement, the forcing mechanism allows a supervisor to preempt unwanted transitions by actively triggering forcible events. Under partial observation, we face a new challenge where the supervisor's forcing decisions must be consistent across all indistinguishable strings meanwhile guarantee all violating transitions are excluded through preemption. We introduce the concept of forcing consistency as a necessary and sufficient condition for the existence of a supervisor that achieves a given specification. We prove that forcing consistency is strictly stronger than the notion of forcibility but, unlike the latter, is not closed under union. These results lay the theoretical foundation for supervisor synthesis in the presence of both forcing mechanisms and observation limitations.

eess.SY↗

arXiv2026

SWE-Pruner Pro: The Coder LLM Already Knows What to Prune

Pruning long context for coding agents has been a vital technology for efficient context management. While existing context pruning methods such as SWE-Pruner realize this by attaching a separate code classifier, we find the agent itself encodes internal representations indicating the relevance of code context when reading tool output. Based on this finding, we propose SWE-Pruner Pro, which prunes tool outputs directly inside the agent. Concretely, a small head turns the agent's own internal representations into a keep-or-prune label for each line, with a length-aware embedding keyed to each tool output's line count. Across two open-weight backbones and four multi-turn benchmarks, SWE-Pruner Pro saves up to 39% of prompt and completion tokens while preserving task quality, with bounded inference overhead. Notably, on MiMo-V2-Flash SWE-Pruner Pro additionally raises the SWE-Bench Verified resolve rate by +3.8% and the long-context Oolong accuracy by +2.2 points.

cs.CL↗

arXiv2026

AutoTrainess: Teaching Language Models to Improve Language Models Autonomously

Training language models (LMs) remains a highly human-intensive process, even as frontier language model agents become increasingly capable at software engineering and other long-horizon tasks. A central challenge is that autonomous post-training is not just a coding problem: it requires the agent to repeatedly plan iterations, construct benchmark-aligned data, run stable training jobs, evaluate checkpoints, and preserve experiment state across many hours of interaction. We present AutoTrainess, a LM agent that exposes these operations as a repository of agent-computer interfaces for planning, data preparation, training, evaluation, and logging. Rather than leaving the agent to operate in a raw CLI environment with an underspecified action space, AutoTrainess externalizes prior human experience as explicit workflows, rules, and execution constraints that guide the agent toward effective and reliable training behavior. On PostTrainBench, AutoTrainess consistently outperforms CLI-only baselines, achieving 26.94 average score with GPT-5.4 (Codex) versus 23.21 for CLI-only. It also generalizes across models and harnesses, improving DeepSeek-V4-Flash (OpenCode) from 12.13 to 19.58.

cs.CL↗

arXiv2026

SWE-MeM: Learning Adaptive Memory Management for Long-Horizon Coding Agents

Long-horizon software engineering agents often need to manage lengthy and noisy interaction histories under limited context budgets. Existing memory management methods typically rely on static compression workflows or impose rigid constraints on compression timing and granularity. Moreover, these approaches fail to jointly optimize memory management and issue resolution capabilities to improve performance while reducing token usage. We present SWE-MeM, a training framework for proactive and on-demand memory management in software engineering agents. SWE-MeM provides a flexible memory tool that lets agents decide when, what, and how to compress based on trajectory state, task progress, and remaining context budget. We train agents with synthesized proactive memory-management trajectories and Memory-aware GRPO, which jointly optimizes memory management and issue resolution through memory-aware trajectory splitting and step-level credit assignment. On SWE-Bench Verified, SWE-MeM achieves 43.4% and 60.2% resolve rate with 4B and 30B models, respectively, outperforming existing memory management baselines in both performance and efficiency.

cs.SE↗

arXiv2026

Dockerless: Environment-Free Program Verifier for Coding Agents

Program verifiers play a central role in training coding agents, including selecting trajectories for supervised fine-tuning (SFT) and providing rewards for reinforcement learning (RL). Standard execution-based verification requires running unit tests inside per-repository environments such as Docker images, incurring substantial environment setup costs. We propose Dockerless, an environment-free agentic patch verifier that evaluates generated code patches without executing them. Rather than simply matching candidate patches to references, Dockerless judges patch correctness using evidence gathered through agentic repository exploration. On a verifier evaluation benchmark, Dockerless outperforms the strongest open-source verifier by 14.3 AUC points. Using Dockerless as both the SFT trajectory filter and the RL reward enables a fully environment-free post-training pipeline. The resulting model reaches 62.0%, 50.0%, and 35.2% resolve rate on SWE-bench Verified, Multilingual, and Pro, respectively. It surpasses the Qwen3.5-9B baseline by 2.4, 8.7, and 2.9 points, matching environment-based post-training.

cs.SE↗

arXiv2026

SWE-Explore: Benchmarking How Coding Agents Explore Repositories

Repository-level coding benchmarks such as SWE-bench have driven a rapid surge in the capabilities of coding agents. Yet they usually treat coding tasks as a holistic, binary prediction problem (e.g., resolved or unresolved), neglecting fine-grained agent capabilities such as repository understanding, context retrieval, code localization, and bug diagnosis. In this paper, we introduce SWE-Explore, a benchmark that isolates the evaluation of repository exploration, a critical capability of coding agents. Given a repository and an issue, SWE-Explore asks an explorer to return a ranked list of relevant code regions under a fixed line budget. SWE-Explore covers 848 issues across 10 programming languages and 203 open-source repositories. For each instance, we derive line-level ground truth from independent agent trajectories that successfully solved the same issue, distilling the specific code regions their solution paths actually consulted. We evaluate exploration along coverage, ranking, and context-efficiency dimensions, showing that these metrics strongly track downstream repair behavior. Across a broad set of retrieval methods, general coding agents, and specialized localizers, we find that agentic explorers form a clear tier above classical retrieval. While file-level localization is already strong for modern methods, line-level coverage and efficient ranking remain the key axes differentiating state-of-the-art explorers.

cs.SE↗

arXiv2026

Data-Driven Synthesis of Probabilistic Controlled Invariant Sets for Linear MDPs

We study data-driven computation of probabilistic controlled invariant sets (PCIS) for safety-critical reinforcement learning under unknown dynamics. Assuming a linear MDP model, we use regularized least squares and self-normalized confidence bounds to construct a conservative estimate of the states from which the system can be kept inside a prescribed safe region over an \(N\)-step horizon, together with the corresponding set-valued safe action map. This construction is obtained through a backward recursion and can be interpreted as a conservative approximation of the \(N\)-step safety predecessor operator. When the associated conservative-inclusion event holds, a conservative fixed point of the approximate recursion can be certified as an \((N,\epsilon)\)-PCIS with confidence at least \(\eta\). For continuous state spaces, we introduce a lattice abstraction and a Lipschitz-based discretization error bound to obtain a tractable approximation scheme. Finally, we use the resulting conservative fixed-point approximation as a runtime candidate PCIS in a practical shielding architecture with iterative updates, and illustrate the approach on a numerical experiment.

eess.SY↗

arXiv2026

SWE-Pruner: Self-Adaptive Context Pruning for Coding Agents

LLM agents have demonstrated remarkable capabilities in software development, but their performance is hampered by long interaction contexts, which incur high API costs and latency. While various context compression approaches such as LongLLMLingua have emerged to tackle this challenge, they typically rely on fixed metrics such as PPL, ignoring the task-specific nature of code understanding. As a result, they frequently disrupt syntactic and logical structure and fail to retain critical implementation details. In this paper, we propose SWE-Pruner, a self-adaptive context pruning framework tailored for coding agents. Drawing inspiration from how human programmers "selectively skim" source code during development and debugging, SWE-Pruner performs task-aware adaptive pruning for long contexts. Given the current task, the agent formulates an explicit goal (e.g., "focus on error handling") as a hint to guide the pruning targets. A lightweight neural skimmer (0.6B parameters) is trained to dynamically select relevant lines from the surrounding context given the goal. Evaluations across four benchmarks and multiple models validate SWE-Pruner's effectiveness in various scenarios, achieving 23-54% token reduction on agent tasks like SWE-Bench Verified while even improving success rates, and up to 14.84x compression on single-turn tasks like LongCodeQA with minimal performance impact.

cs.SE↗

arXiv2026

Marking Data-Informativity and Data-Driven Supervisory Control of Discrete-Event Systems

In this paper we develop a data-driven approach for marking nonblocking supervisory control of discrete-event systems (DES). We consider a setup in which models of DES to be controlled are unknown, but a set of data concerning the behaviors of DES is available. We ask the question: Under what conditions of the available data set can a valid marking noblocking supervisor be designed for the unknown DES to satisfy a given specification? Answering this question, we identify and formalize a novel concept called marking data-informativity. Moreover, we design an algorithm for the verification of this concept. Next, if the data set fails to be marking informative, we propose two related new concepts of restricted marking data-informativity and marking informatizability. Finally, we develop an algorithm to compute the largest subset of control specification for which the data set is least restricted marking informative.

cs.FL↗

arXiv2025

UI-TARS-2 Technical Report: Advancing GUI Agent with Multi-Turn Reinforcement Learning

The development of autonomous agents for graphical user interfaces (GUIs) presents major challenges in artificial intelligence. While recent advances in native agent models have shown promise by unifying perception, reasoning, action, and memory through end-to-end learning, open problems remain in data scalability, multi-turn reinforcement learning (RL), the limitations of GUI-only operation, and environment stability. In this technical report, we present UI-TARS-2, a native GUI-centered agent model that addresses these challenges through a systematic training methodology: a data flywheel for scalable data generation, a stabilized multi-turn RL framework, a hybrid GUI environment that integrates file systems and terminals, and a unified sandbox platform for large-scale rollouts. Empirical evaluation demonstrates that UI-TARS-2 achieves significant improvements over its predecessor UI-TARS-1.5. On GUI benchmarks, it reaches 88.2 on Online-Mind2Web, 47.5 on OSWorld, 50.6 on WindowsAgentArena, and 73.3 on AndroidWorld, outperforming strong baselines such as Claude and OpenAI agents. In game environments, it attains a mean normalized score of 59.8 across a 15-game suite-roughly 60% of human-level performance-and remains competitive with frontier proprietary models (e.g., OpenAI o3) on LMGame-Bench. Additionally, the model can generalize to long-horizon information-seeking tasks and software engineering benchmarks, highlighting its robustness across diverse agent tasks. Detailed analyses of training dynamics further provide insights into achieving stability and efficiency in large-scale agent RL. These results underscore UI-TARS-2's potential to advance the state of GUI agents and exhibit strong generalization to real-world interactive scenarios.

cs.AI↗

arXiv2025

Markov Clustering based Fully Automated Nonblocking Hierarchical Supervisory Control of Large-Scale Discrete-Event Systems

In this paper we revisit the abstraction-based approach to synthesize a hierarchy of decentralized supervisors and coordinators for nonblocking control of large-scale discrete-event systems (DES), and augment it with a new clustering method for automatic and flexible grouping of relevant components during the hierarchical synthesis process. This method is known as Markov clustering, which not only automatically performs grouping but also allows flexible tuning the sizes of the resulting clusters using a single parameter. Compared to the existing abstraction-based approach that lacks effective grouping method for general cases, our proposed approach based on Markov clustering provides a fully automated and effective hierarchical synthesis procedure applicable to general large-scale DES. Moreover, it is proved that the resulting hierarchy of supervisors and coordinators collectively achieves global nonblocking (and maximally permissive) controlled behavior under the same conditions as those in the existing abstraction-based approach. Finally, a benchmark case study is conducted to empirically demonstrate the effectiveness of our approach.

eess.SY↗

arXiv2025

UI-TARS: Pioneering Automated GUI Interaction with Native Agents

This paper introduces UI-TARS, a native GUI agent model that solely perceives the screenshots as input and performs human-like interactions (e.g., keyboard and mouse operations). Unlike prevailing agent frameworks that depend on heavily wrapped commercial models (e.g., GPT-4o) with expert-crafted prompts and workflows, UI-TARS is an end-to-end model that outperforms these sophisticated frameworks. Experiments demonstrate its superior performance: UI-TARS achieves SOTA performance in 10+ GUI agent benchmarks evaluating perception, grounding, and GUI task execution. Notably, in the OSWorld benchmark, UI-TARS achieves scores of 24.6 with 50 steps and 22.7 with 15 steps, outperforming Claude (22.0 and 14.9 respectively). In AndroidWorld, UI-TARS achieves 46.6, surpassing GPT-4o (34.5). UI-TARS incorporates several key innovations: (1) Enhanced Perception: leveraging a large-scale dataset of GUI screenshots for context-aware understanding of UI elements and precise captioning; (2) Unified Action Modeling, which standardizes actions into a unified space across platforms and achieves precise grounding and interaction through large-scale action traces; (3) System-2 Reasoning, which incorporates deliberate reasoning into multi-step decision making, involving multiple reasoning patterns such as task decomposition, reflection thinking, milestone recognition, etc. (4) Iterative Training with Reflective Online Traces, which addresses the data bottleneck by automatically collecting, filtering, and reflectively refining new interaction traces on hundreds of virtual machines. Through iterative training and reflection tuning, UI-TARS continuously learns from its mistakes and adapts to unforeseen situations with minimal human intervention. We also analyze the evolution path of GUI agents to guide the further development of this domain.

cs.AI↗

arXiv2024

Supervisory Control Theory with Event Forcing

In the Ramadge-Wonham supervisory control theory the only interaction mechanism between supervisor and plant is that the supervisor may enable/disable events from the plant and the plant makes a final decision about which of the enabled events is actually taking place. In this paper, the interaction between supervisor and plant is enriched by allowing the supervisor to force specific events (called forcible events) that are allowed to preempt uncontrollable events. A notion of forcible-controllability is defined that captures the interplay between controllability of a supervisor w.r.t. the uncontrollable events provided by a plant in the setting with event forcing. Existence of a maximally permissive, forcibly-controllable, nonblocking supervisor is shown and an algorithm is provided that computes such a supervisor. The approach is illustrated by two small case studies.

cs.FL↗

arXiv2024

Distributed Secret Securing in Discrete-Event Systems

In this paper, we study a security problem of protecting secrets in distributed systems. Specifically, we employ discrete-event systems to describe the structure and behaviour of distributed systems, in which global secret information is separated into pieces and stored in local component agents. The goal is to prevent such secrets from being exposed to intruders by imposing appropriate protection measures. This problem is formulated as to ensure that at least one piece of every distributed global secret is secured by a required number of protections, while the overall cost to apply protections is minimum. We first characterize the solvability of this security problem by providing a necessary and sufficient condition, and then develop an algorithm to compute a solution based on the supervisory control theory of discrete-event systems. Finally, we illustrate the effectiveness of our solution with an example system comprising distributed databases.

eess.SY↗

arXiv2024

Bounded-Time Nonblocking Supervisory Control of Timed Discrete-Event Systems

Recently an automaton property of quantitative nonblockingness was proposed in supervisory control of untimed discrete-event systems (DES), which quantifies the standard nonblocking property by capturing the practical requirement that all tasks be completed within a bounded number of steps. However, in practice tasks may be further required to be completed in specific time. To meet this new requirement, in this paper we introduce the concept of bounded-time nonblockingness, which extends the concept of quantitative nonblockingness from untimed DES to timed DES. This property requires that each task must be completed within a bounded time counted by the number of ticks, rather than bounded number of transition steps in quantitative nonblockingness. Accordingly, we formulate a new bounded-time nonblocking supervisory control problem (BTNSCP) of timed DES, and characterize its solvability in terms of a new concept of bounded-time language completability. Then we present an approach to compute the maximally permissive solution to the new BTNSCP.

eess.SY↗

arXiv2023

Quantitatively Nonblocking Supervisory Control of Discrete-Event Systems

In this paper, we propose two new nonblocking properties of automata as quantitative measures of maximal distances to marker states. The first property, called {\em quantitative nonblockingness}, captures the practical requirement that at least one of the marker states (representing e.g. task completion) be reachable within a prescribed number of steps from any non-marker state and following any trajectory of the system. The second property, called {\em heterogeneously quantitative nonblockingness}, distinguishes individual marker states and requires that each marker state be reached within a given bounded number of steps from any other state and following any trajectory of the system. Accordingly, we formulate two new problems of quantitatively nonblocking supervisory control and heterogeneously quantitatively nonblocking supervisory control, and characterize their solvabilities in terms of new concepts of quantitative language completability and heterogeneously quantitative language completability, respectively. It is proved that there exists the unique supremal (heterogeneously) quantitatively completable sublanguage of a given language, and we develop effective algorithms to compute the supremal sublanguages. Finally, combining with the algorithm of computing the supremal controllable sublanguage, we design algorithms to compute the maximally permissive solutions to the formulated (heterogeneously) quantitatively nonblocking supervisory control problems.

eess.SY↗

arXiv2023

Transformational Supervisor Localization

Supervisor localization can be applied to distribute a monolithic supervisor into local supervisors. Performing supervisor localization can be computationally costly. In this work, we consider systems that evolve over time. We study how to reuse the results from a previous supervisor localization, to more efficiently compute local supervisors when the system is adapted. We call this approach transformational supervisor localization, and present algorithms for the procedure. The efficiency of the procedure is experimentally evaluated.

cs.FL↗