arXiv subjects

Haotian Zhu

Publications and source records attributed to Haotian Zhu.

12 recordsLinked to original sources

arXiv2026

EdgeBench: Unveiling Scaling Laws of Learning from Real-World Environments

Pretraining scaling laws reveal that model capability improves predictably with data and compute. But learning from real world environments after deployment remains far less understood. Analyzing roughly 38,000 hours of agent interaction with the environment across 134 real world tasks, we find, to the best of our knowledge, the first evidence that overall performance during environment learning follows a log-sigmoid scaling law with remarkably high precision, reaching R^2 = 0.998. Across model generations, we also find that agent learning speed roughly doubles every three months. This discovery stems from EdgeBench, a suite of 134 real world tasks with ultra-long horizons, spanning scientific discovery, software engineering, combinatorial optimization, professional knowledge work, formal mathematics, and interactive games. Each task sustains at least 12 hours of continuous agent operation under rich, multilevel feedback, and is built through substantial expert effort. We publicly release 51 tasks and our full evaluation framework to accelerate the study of how agents learn from real world experience.

cs.CL↗

arXiv2026

Towards Agentic Runtime Healing

Self-healing systems have long been a focus of research, aiming to enable software to recover from unexpected runtime errors without human intervention. Traditional approaches rely on predefined heuristic rules, such as reusing error handlers or rolling back to checkpoints, but these methods struggle to adapt to the diverse range of runtime errors. The emergence of Large Language Models offers a new opportunity to address this challenge. Leveraging their ability to understand and generate code and natural language, we propose using LLMs to dynamically generate error-handling strategies in real time, tailored to specific runtime contexts such as error messages and program states. We demonstrate the feasibility of this approach by designing such a framework, Healer, and empirically showing that it can handle runtime errors with a high success rate. When an unanticipated runtime error occurs, Healer leverages its internal LLM to generate bespoke error-handling code. The generated healing code is then executed to produce a corrected program state, allowing the program to continue execution with minimal disruption. We evaluate Healer across four code datasets and three state-of-the-art LLMs (GPT-3.5, GPT-4, and CodeQwen-7B), where GPT-4 can successfully recover from 72.8% of runtime errors, underscoring the promise of LLMs in this domain. Despite these promising results, challenges remain, particularly regarding the trustworthiness of LLM-generated code and its integration into existing systems. We mention potential solutions, such as safety checks and Healer-aware programming, to mitigate risks and ensure reliable operation. This work represents the first step toward agentic runtime healing, paving the way for more adaptive, resilient, and self-healing software systems.

cs.SE↗

arXiv2026

AgentRAE: Remote Action Execution through Notification-based Visual Backdoors against Screenshots-based Mobile GUI Agents

The rapid adoption of mobile graphical user interface (GUI) agents, which autonomously control applications and operating systems (OS), exposes new system-level attack surfaces. Existing backdoors against web GUI agents and general GenAI models rely on environmental injection or deceptive pop-ups to mislead the agent operation. However, these techniques do not work on screenshots-based mobile GUI agents due to the challenges of restricted trigger design spaces, OS background interference, and conflicts in multiple trigger-action mappings. We propose AgentRAE, a novel backdoor attack capable of inducing Remote Action Execution in mobile GUI agents using visually natural triggers (e.g., benign app icons in notifications). To address the underfitting caused by natural triggers and achieve accurate multi-target action redirection, we design a novel two-stage pipeline that first enhances the agent's sensitivity to subtle iconographic differences via contrastive learning, and then associates each trigger with a specific mobile GUI agent action through a backdoor post-training. Our extensive evaluation reveals that the proposed backdoor preserves clean performance with an attack success rate of over 90% across ten mobile operations. Furthermore, it is hard to visibly detect the benign-looking triggers and circumvents eight representative state-of-the-art defenses. These results expose an overlooked backdoor vector in mobile GUI agents, underscoring the need for defenses that scrutinize notification-conditioned behaviors and internal agent representations.

cs.CR↗

arXiv2026

CondenseGraph: Communication-Efficient Distributed GNN Training via On-the-Fly Graph Condensation

Distributed Graph Neural Network (GNN) training suffers from substantial communication overhead due to the inherent neighborhood dependency in graph-structured data. This neighbor explosion problem requires workers to frequently exchange boundary node features across partitions, creating a communication bottleneck that severely limits training scalability. Existing approaches rely on static graph partitioning strategies that cannot adapt to dynamic network conditions. In this paper, we propose CondenseGraph, a novel communication-efficient framework for distributed GNN training. Our key innovation is an on-the-fly graph condensation mechanism that dynamically compresses boundary node features into compact super nodes before transmission. To compensate for the information loss introduced by compression, we develop a gradient-based error feedback mechanism that maintains convergence guarantees while reducing communication volume by 40-60%. Extensive experiments on four benchmark datasets demonstrate that CondenseGraph achieves comparable accuracy to full-precision baselines while significantly reducing communication costs and training time.

cs.DC↗

arXiv2026

E$^2$AT: Multimodal Jailbreak Defense via Dynamic Joint Optimization for Multimodal Large Language Models

Research endeavors have been made in learning robust Multimodal Large Language Models (MLLMs) against jailbreak attacks. However, existing methods for improving MLLMs' robustness still face critical challenges: \ding{172} how to efficiently tune massive weight parameters and \ding{173} how to ensure robustness against attacks across both visual and textual modalities. To this end, we propose an \textbf{E}fficient \textbf{E}nd-to-end \textbf{A}dversarial \textbf{T}raining (E$^2$AT) framework for both visual and textual adversarial attacks. Specifically, for the visual aspect, E$^2$AT incorporates an efficient projector-based AT module that aligns the attack samples at the feature level. For training objectives, we propose a Dynamic Joint Multimodal Optimization (DJMO) strategy to enhance generalization ability against jailbreak attacks by dynamically adjusting weights between normal and adversarial objectives. Extensive experiments are conducted with five major jailbreak attack methods across three mainstream MLLMs. Results demonstrate that our E$^2$AT achieves the state-of-the-art performance, outperforming existing baselines by an average margin of 34\% across text and image modalities, while maintaining clean task performance. Furthermore, evaluations of real-world embodied intelligent systems highlight the practical applicability of E$^2$AT, paving the way for the development of more secure and reliable multimodal systems. Our code is available on \href{https://anonymous.4open.science/r/E2AT_568}{\textcolor{red}{https://anonymous.4open.science/r/E2AT\_568}}.

cs.CV↗

arXiv2025

SafeSteer: Adaptive Subspace Steering for Efficient Jailbreak Defense in Vision-Language Models

As the capabilities of Vision Language Models (VLMs) continue to improve, they are increasingly targeted by jailbreak attacks. Existing defense methods face two major limitations: (1) they struggle to ensure safety without compromising the model's utility; and (2) many defense mechanisms significantly reduce the model's inference efficiency. To address these challenges, we propose SafeSteer, a lightweight, inference-time steering framework that effectively defends against diverse jailbreak attacks without modifying model weights. At the core of SafeSteer is the innovative use of Singular Value Decomposition to construct a low-dimensional "safety subspace." By projecting and reconstructing the raw steering vector into this subspace during inference, SafeSteer adaptively removes harmful generation signals while preserving the model's ability to handle benign inputs. The entire process is executed in a single inference pass, introducing negligible overhead. Extensive experiments show that SafeSteer reduces the attack success rate by over 60% and improves accuracy on normal tasks by 1-2%, without introducing significant inference latency. These results demonstrate that robust and practical jailbreak defense can be achieved through simple, efficient inference-time control.

cs.CR↗

arXiv2025

AutoCodeBench: Large Language Models are Automatic Code Benchmark Generators

Large Language Models (LLMs) have demonstrated remarkable capabilities across various domains, with code generation emerging as a key area of focus. While numerous benchmarks have been proposed to evaluate their code generation abilities, these benchmarks face several critical limitations. First, they often rely on manual annotations, which are time-consuming and difficult to scale across different programming languages and problem complexities. Second, most existing benchmarks focus primarily on Python, while the few multilingual benchmarks suffer from limited difficulty and uneven language distribution. To address these challenges, we propose AutoCodeGen, an automated method for generating high-difficulty multilingual code generation datasets without manual annotations. AutoCodeGen ensures the correctness and completeness of test cases by generating test inputs with LLMs and obtaining test outputs through a multilingual sandbox, while achieving high data quality through reverse-order problem generation and multiple filtering steps. Using this novel method, we introduce AutoCodeBench, a large-scale code generation benchmark comprising 3,920 problems evenly distributed across 20 programming languages. It is specifically designed to evaluate LLMs on challenging, diverse, and practical multilingual tasks. We evaluate over 30 leading open-source and proprietary LLMs on AutoCodeBench and its simplified version AutoCodeBench-Lite. The results show that even the most advanced LLMs struggle with the complexity, diversity, and multilingual nature of these tasks. Besides, we introduce AutoCodeBench-Complete, specifically designed for base models to assess their few-shot code generation capabilities. We hope the AutoCodeBench series will serve as a valuable resource and inspire the community to focus on more challenging and practical multilingual code generation scenarios.

cs.CL↗

arXiv2024

SciPrompt: Knowledge-augmented Prompting for Fine-grained Categorization of Scientific Topics

Prompt-based fine-tuning has become an essential method for eliciting information encoded in pre-trained language models for a variety of tasks, including text classification. For multi-class classification tasks, prompt-based fine-tuning under low-resource scenarios has resulted in performance levels comparable to those of fully fine-tuning methods. Previous studies have used crafted prompt templates and verbalizers, mapping from the label terms space to the class space, to solve the classification problem as a masked language modeling task. However, cross-domain and fine-grained prompt-based fine-tuning with an automatically enriched verbalizer remains unexplored, mainly due to the difficulty and costs of manually selecting domain label terms for the verbalizer, which requires humans with domain expertise. To address this challenge, we introduce SciPrompt, a framework designed to automatically retrieve scientific topic-related terms for low-resource text classification tasks. To this end, we select semantically correlated and domain-specific label terms within the context of scientific literature for verbalizer augmentation. Furthermore, we propose a new verbalization strategy that uses correlation scores as additional weights to enhance the prediction performance of the language model during model tuning. Our method outperforms state-of-the-art, prompt-based fine-tuning methods on scientific text classification tasks under few and zero-shot settings, especially in classifying fine-grained and emerging scientific topics.

cs.CL↗

arXiv2024

A Case Study of LLM for Automated Vulnerability Repair: Assessing Impact of Reasoning and Patch Validation Feedback

Recent work in automated program repair (APR) proposes the use of reasoning and patch validation feedback to reduce the semantic gap between the LLMs and the code under analysis. The idea has been shown to perform well for general APR, but its effectiveness in other particular contexts remains underexplored. In this work, we assess the impact of reasoning and patch validation feedback to LLMs in the context of vulnerability repair, an important and challenging task in security. To support the evaluation, we present VRpilot, an LLM-based vulnerability repair technique based on reasoning and patch validation feedback. VRpilot (1) uses a chain-of-thought prompt to reason about a vulnerability prior to generating patch candidates and (2) iteratively refines prompts according to the output of external tools (e.g., compiler, code sanitizers, test suite, etc.) on previously-generated patches. To evaluate performance, we compare VRpilot against the state-of-the-art vulnerability repair techniques for C and Java using public datasets from the literature. Our results show that VRpilot generates, on average, 14% and 7.6% more correct patches than the baseline techniques on C and Java, respectively. We show, through an ablation study, that reasoning and patch validation feedback are critical. We report several lessons from this study and potential directions for advancing LLM-empowered vulnerability repair

cs.SE↗

arXiv2021

Emergence of intrinsically isolated flat bands and their topology in fully relaxed twisted multi-layer graphene

We study the electronic structure and band topology of fully relaxed twisted multi-layer graphene (TMLG). Isolated flat bands emerge in TMLG with the number of layers [$M+N$ with $M$ the layer number of the bottom few-layer graphene (FLG)] up to 10 and with various stacking orders, and most of them are on the hole side. The touched bands of FLGs around the Fermi level are split by the moiré coupling through the electron-hole asymmetry in low-energy bands of FLGs and by the vertical hopping between next-nearest layers. The full structural relaxation leads to global gaps that completely isolate a flat band. For TMLG with given $M$ and $N$, the highest magnitude of Chern numbers ($|C|$) of the separable flat bands reaches $M+N-1$ and can be hosted by certain isolated bands. The $|C|=9$ occurs in the isolated flat valence band of several configurations with 10 layers. Such high $|C|$ originates from the lifting of the band-state degeneracy in the weak regime of moiré coupling or from the topological phase transitions induced by the strong moiré coupling. Moreover, large orbital magnetic moments arise in isolated flat bands with high $|C|$ and depend on the structural configurations of TMLG.

cond-mat.mes-hall↗

arXiv2020

NLPStatTest: A Toolkit for Comparing NLP System Performance

Statistical significance testing centered on p-values is commonly used to compare NLP system performance, but p-values alone are insufficient because statistical significance differs from practical significance. The latter can be measured by estimating effect size. In this paper, we propose a three-stage procedure for comparing NLP system performance and provide a toolkit, NLPStatTest, that automates the process. Users can upload NLP system evaluation scores and the toolkit will analyze these scores, run appropriate significance tests, estimate effect size, and conduct power analysis to estimate Type II error. The toolkit provides a convenient and systematic way to compare NLP system performance that goes beyond statistical significance testing

cs.CL↗

arXiv2020

Pressure induced gap modulation and topological transitions in twisted bilayer and double bilayer graphene

We study the electronic and topological properties of fully relaxed twisted bilayer (TBG) and double bilayer (TDBG) graphene under perpendicular pressure. An approach has been proposed to obtain the equilibrium in-plane structural deformation and out-of-plane corrugation in moiré superlattices under pressure. We find that the in-plane relaxation becomes much stronger under higher pressure, while the corrugation height in each layer is maintained. The comparison between band structures of relaxed and rigid structures demonstrates that not only the gaps on the electron and hole sides ($Δ_e$ and $Δ_h$) are significantly underestimated without relaxation but also the detailed dispersions of the middle bands of rigid structures are rather different from those of relaxed systems. $Δ_e$ and $Δ_h$ in TBG reach maximum values around critical pressures with narrowest middle bands. Topological transitions occur in TDBG under pressure with the middle valence and conduction bands in one valley touching and their Chern numbers transferred to each other. The pressure can also tune the gap at the neutrality point of TDBG, which becomes closed for a pressure range and reopened under higher pressure. The behavior of electronic structure of supertlattices under pressure is sensitive to the twist angle $θ$ with the critical pressures generally increase with $θ$.

cond-mat.mes-hall↗