arXiv subjects

Hao Lin

Publications and source records attributed to Hao Lin.

At least 55 records · Page 3Linked to original sources

arXiv2025

RLinf: Flexible and Efficient Large-scale Reinforcement Learning via Macro-to-Micro Flow Transformation

Reinforcement learning (RL) has demonstrated immense potential in advancing artificial general intelligence, agentic intelligence, and embodied intelligence. However, the inherent heterogeneity and dynamicity of RL workflows often lead to low hardware utilization and slow training on existing systems. In this paper, we present RLinf, a high-performance RL training system based on our key observation that the major roadblock to efficient RL training lies in system flexibility. To maximize flexibility and efficiency, RLinf is built atop a novel RL system design paradigm called macro-to-micro flow transformation (M2Flow), which automatically breaks down high-level, easy-to-compose RL workflows at both the temporal and spatial dimensions, and recomposes them into optimized execution flows. Supported by RLinf worker's adaptive communication capability, we devise context switching and elastic pipelining to realize M2Flow transformation, and a profiling-guided scheduling policy to generate optimal execution plans. Extensive evaluations on both reasoning RL and embodied RL tasks demonstrate that RLinf consistently outperforms state-of-the-art systems, achieving $1.07\times-2.43\times$ speedup in end-to-end training throughput.

cs.LG↗

arXiv2025

FUSCO: High-Performance Distributed Data Shuffling via Transformation-Communication Fusion

Large-scale Mixture-of-Experts (MoE) models rely on \emph{expert parallelism} for efficient training and inference, which splits experts across devices and necessitates distributed data shuffling to route each token to its assigned experts. However, existing communication libraries handle this shuffling poorly; its overhead can account for over half of end-to-end runtime. We present FUSCO, an MoE-friendly communication library that achieves efficient and lightweight data shuffling through fused data transformation and communication, based on the key observation that MoE's expert-major data layout conflicts with the device-major layout expected by communication operations. FUSCO captures the fine-grained data layout, which is then interpreted by a pipelined communication engine that performs the required shuffling efficiently along the communication path. Lightweight planning and load-balancing mechanisms complement the engine by eliminating redundant communication and dispersing traffic. Evaluations on representative benchmarks illustrate that FUSCO achieves up to 3.84$\times$ and 2.01$\times$ speedups over NCCL and DeepEP (the state-of-the-art MoE communication library), respectively. In end-to-end MoE tasks, compared to NCCL and DeepEP, FUSCO reduces the training latency by 1.17-1.39$\times$ and 1.10-1.19$\times$, and lowers the first-token generation latency in inference by 1.09-1.25$\times$ and 1.06-1.16$\times$.

cs.DC↗

arXiv2025

Stabilizing Reinforcement Learning with LLMs: Formulation and Practices

This paper proposes a novel formulation for reinforcement learning (RL) with large language models, explaining why and under what conditions the true sequence-level reward can be optimized via a surrogate token-level objective in policy gradient methods such as REINFORCE. Specifically, through a first-order approximation, we show that this surrogate becomes increasingly valid only when both the training-inference discrepancy and policy staleness are minimized. This insight provides a principled explanation for the crucial role of several widely adopted techniques in stabilizing RL training, including importance sampling correction, clipping, and particularly Routing Replay for Mixture-of-Experts (MoE) models. Through extensive experiments with a 30B MoE model totaling hundreds of thousands of GPU hours, we show that for on-policy training, the basic policy gradient algorithm with importance sampling correction achieves the highest training stability. When off-policy updates are introduced to accelerate convergence, combining clipping and Routing Replay becomes essential to mitigate the instability caused by policy staleness. Notably, once training is stabilized, prolonged optimization consistently yields comparable final performance regardless of cold-start initialization. We hope that the shared insights and the developed recipes for stable RL training will facilitate future research.

cs.LG↗

arXiv2025

STAlloc: Enhancing Memory Efficiency in Large-Scale Model Training with Spatio-Temporal Planning

The rapid scaling of large language models (LLMs) has significantly increased GPU memory pressure, which is further aggravated by training optimization techniques such as virtual pipeline and recomputation that disrupt tensor lifespans and introduce considerable memory fragmentation. Such fragmentation stems from the use of online GPU memory allocators in popular deep learning frameworks like PyTorch, which disregard tensor lifespans. As a result, this inefficiency can waste as much as 43% of memory and trigger out-of-memory errors, undermining the effectiveness of optimization methods. To address this, we introduce STAlloc, a GPU memory allocator for deep learning frameworks that reduces fragmentation by exploiting the spatial and temporal regularity in memory allocation behaviors of training workloads. STAlloc introduces a novel paradigm that combines offline planning with online allocation. The offline planning leverages spatio-temporal regularities to generate a near-optimal allocation plan, while the online allocation handles complex and dynamic models such as Mixture-of-Experts (MoE). Built as a pluggable PyTorch memory allocator, STAlloc reduces fragmentation ratio on average by 85.1% (up to 100%) across both dense and MoE models, with negligible overhead. This enables more efficient, high-throughput training configurations and improves throughput performance by up to 32.5%.

cs.LG↗

arXiv2025

Unraveling Intertwined Orders in the Strongly Correlated Kagome Metal CsCr3Sb5

While correlated phenomena of flat bands have been extensively studied in twisted systems, the ordered states that emerge from interactions in the intrinsic flat bands of kagome lattice materials remain largely unexplored. The newly discovered kagome metal CsCr3Sb5 offers a unique and rich platform for this research, as its multi-orbital flat bands at the Fermi surface result in a complex interplay of pressurized superconductivity, antiferromagnetism, a structural phase transition, and density wave orders. Here, using ultrafast optical techniques, we provide strong spectroscopic evidence for a charge density wave transition in CsCr3Sb5, resolving previous ambiguities. Crucially, we identify rotational symmetry breaking that manifests as a three-state Potts-type nematicity. Our elastoresistance measurements directly demonstrate the electronic origin of this order, as the rotational-symmetry-breaking E2g component of the elastoresistance shows a divergent behaviour around the transition temperature. This exotic nematicity results from the lifting of degeneracy of the multi-orbital flat bands, akin to phenomena seen in certain iron-based superconductors. Our study pioneers the investigation of ultrafast dynamics in flat-band systems at the Fermi surface, offering new insights into the interactions between multiple elementary excitations in strongly correlated systems.

cond-mat.str-el↗

arXiv2025

TIT: A Tree-Structured Instruction Tuning Approach for LLM-Based Code Translation

Large Language Models (LLMs) have shown strong performance in automated source-to-target code translation through pretraining on extensive code corpora. However, mainstream LLM-based code translation methods suffer from two critical limitations. First, they are highly sensitive to language-specific features, which often introduce source-language syntax or lexicon into the output, leading to syntactic confusion. Second, they lack fine-grained semantic alignment due to an over-reliance on function-level parallel datasets, resulting in semantic misalignment between the translated code and the original source. To overcome these limitations, we propose TIT, a Tree-structured Instruction Tuning paradigm for LLM-based code translation. Specifically, TIT consists of three modules. First, to mitigate syntactic confusion, the syntactic information representation module integrates language-agnostic syntactic features via structured parsing. Then, to generate high-quality fine-grained parallel data, the fine-grained parallel dataset augmentation module aligns nodes with code segments through statement-level segmentation and contrastive matching. Finally, we leverage the dual-stage tree instruction tuning module to alleviate the contextual processing burden on the LLM caused by the introduction of syntactic information. The first stage employs syntax-aware fine-tuning to enable the LLM to autonomously comprehend structured syntactic information, while the second stage utilizes code generation fine-tuning to guide the model in generating accurate target code based on function-level syntactic dependencies. The experimental results demonstrate that the proposed method significantly outperforms existing approaches in multiple LLMs, achieving a success rate 1.22x-1.75x higher in code translation while markedly reducing syntactic confusion.

cs.SE↗

arXiv2025

Domain-Aware RAG: MoL-Enhanced RL for Efficient Training and Scalable Retrieval

Retrieval-Augmented Generation (RAG) systems rely heavily on the retrieval stage, particularly the coarse-ranking process. Existing coarse-ranking optimization approaches often struggle to balance domain-specific knowledge learning with query enhencement, resulting in suboptimal retrieval performance. To address this challenge, we propose MoLER, a domain-aware RAG method that uses MoL-Enhanced Reinforcement Learning to optimize retrieval. MoLER has a two-stage pipeline: a continual pre-training (CPT) phase using a Mixture of Losses (MoL) to balance domain-specific knowledge with general language capabilities, and a reinforcement learning (RL) phase leveraging Group Relative Policy Optimization (GRPO) to optimize query and passage generation for maximizing document recall. A key innovation is our Multi-query Single-passage Late Fusion (MSLF) strategy, which reduces computational overhead during RL training while maintaining scalable inference via Multi-query Multi-passage Late Fusion (MMLF). Extensive experiments on benchmark datasets show that MoLER achieves state-of-the-art performance, significantly outperforming baseline methods. MoLER bridges the knowledge gap in RAG systems, enabling robust and scalable retrieval in specialized domains.

cs.CL↗

arXiv2025

MEGA: Second-Order Gradient Alignment for Catastrophic Forgetting Mitigation in GFSCIL

Graph Few-Shot Class-Incremental Learning (GFSCIL) enables models to continually learn from limited samples of novel tasks after initial training on a large base dataset. Existing GFSCIL approaches typically utilize Prototypical Networks (PNs) for metric-based class representations and fine-tune the model during the incremental learning stage. However, these PN-based methods oversimplify learning via novel query set fine-tuning and fail to integrate Graph Continual Learning (GCL) techniques due to architectural constraints. To address these challenges, we propose a more rigorous and practical setting for GFSCIL that excludes query sets during the incremental training phase. Building on this foundation, we introduce Model-Agnostic Meta Graph Continual Learning (MEGA), aimed at effectively alleviating catastrophic forgetting for GFSCIL. Specifically, by calculating the incremental second-order gradient during the meta-training stage, we endow the model to learn high-quality priors that enhance incremental learning by aligning its behaviors across both the meta-training and incremental learning stages. Extensive experiments on four mainstream graph datasets demonstrate that MEGA achieves state-of-the-art results and enhances the effectiveness of various GCL methods in GFSCIL. We believe that our proposed MEGA serves as a model-agnostic GFSCIL paradigm, paving the way for future research.

cs.LG↗

arXiv2025

Multi-task Adversarial Attacks against Black-box Model with Few-shot Queries

Current multi-task adversarial text attacks rely on abundant access to shared internal features and numerous queries, often limited to a single task type. As a result, these attacks are less effective against practical scenarios involving black-box feedback APIs, limited queries, or multiple task types. To bridge this gap, we propose \textbf{C}luster and \textbf{E}nsemble \textbf{M}ulti-task Text Adversarial \textbf{A}ttack (\textbf{CEMA}), an effective black-box attack that exploits the transferability of adversarial texts across different tasks. CEMA simplifies complex multi-task scenarios by using a \textit{deep-level substitute model} trained in a \textit{plug-and-play} manner for text classification, enabling attacks without mimicking the victim model. This approach requires only a few queries for training, converting multi-task attacks into classification attacks and allowing attacks across various tasks. CEMA generates multiple adversarial candidates using different text classification methods and selects the one that most effectively attacks substitute models. In experiments involving multi-task models with two, three, or six tasks--spanning classification, translation, summarization, and text-to-image generation--CEMA demonstrates significant attack success with as few as 100 queries. Furthermore, CEMA can target commercial APIs (e.g., Baidu and Google Translate), large language models (e.g., ChatGPT 4o), and image-generation models (e.g., Stable Diffusion V2), showcasing its versatility and effectiveness in real-world applications.

cs.CR↗

arXiv2025

No Query, No Access

Textual adversarial attacks mislead NLP models, including Large Language Models (LLMs), by subtly modifying text. While effective, existing attacks often require knowledge of the victim model, extensive queries, or access to training data, limiting real-world feasibility. To overcome these constraints, we introduce the \textbf{Victim Data-based Adversarial Attack (VDBA)}, which operates using only victim texts. To prevent access to the victim model, we create a shadow dataset with publicly available pre-trained models and clustering methods as a foundation for developing substitute models. To address the low attack success rate (ASR) due to insufficient information feedback, we propose the hierarchical substitution model design, generating substitute models to mitigate the failure of a single substitute model at the decision boundary. Concurrently, we use diverse adversarial example generation, employing various attack methods to generate and select the adversarial example with better similarity and attack effectiveness. Experiments on the Emotion and SST5 datasets show that VDBA outperforms state-of-the-art methods, achieving an ASR improvement of 52.08\% while significantly reducing attack queries to 0. More importantly, we discover that VDBA poses a significant threat to LLMs such as Qwen2 and the GPT family, and achieves the highest ASR of 45.99% even without access to the API, confirming that advanced NLP models still face serious security risks. Our codes can be found at https://anonymous.4open.science/r/VDBA-Victim-Data-based-Adversarial-Attack-36EC/

cs.CL↗

arXiv2025

Navigating the Risks of Using Large Language Models for Text Annotation in Social Science Research

Large language models (LLMs) have the potential to revolutionize computational social science, particularly in automated textual analysis. In this paper, we conduct a systematic evaluation of the promises and risks associated with using LLMs for text classification tasks, using social movement studies as an example. We propose a framework for social scientists to incorporate LLMs into text annotation, either as the primary coding decision-maker or as a coding assistant. This framework offers researchers tools to develop the potential best-performing prompt, and to systematically examine and report the validity and reliability of LLMs as a methodological tool. Additionally, we evaluate and discuss its epistemic risks associated with validity, reliability, replicability, and transparency. We conclude with several practical guidelines for using LLMs in text annotation tasks and offer recommendations for more effectively communicating epistemic risks in research.

cs.CL↗

arXiv2025

Hypergraphs with a quarter uniform Turán density

The uniform Turán density $π_{1}(F)$ of a $3$-uniform hypergraph $F$ is the supremum over all $d$ for which there is an $F$-free hypergraph with the property that every linearly sized subhypergraph with density at least $d$. Determining $π_{1}(F)$ for given hypergraphs $F$ was suggested by Erdős and Sós in 1980s. In particular, they raised the questions of determining $π_{1}(K_4^{(3)-})$ and $π_{1}(K_4^{(3)})$. The former question was solved recently in [Israel J. Math. 211 (2016), 349-366] and [J. Eur. Math. Soc. 20 (2018), 1139-1159], while the latter is still a major open problem. In addition to $K_4^{(3)-}$, there are very few hypergraphs whose uniform Turán density has been determined. In this paper, we give a sufficient condition for $3$-uniform hypergraphs $F$ satisfying $π_{1}(F)=1/4$. In particular, currently all known $3$-uniform hypergraphs whose uniform Turán density is $1/4$, such as $K_4^{(3)-}$ and the $3$-uniform hypergraphs $F^{\star}_5$ studied in [arXiv:2211.12747], satisfy this condition. Moreover, we find some intriguing $3$-uniform hypergraphs whose uniform Turán density is also $1/4$.

math.CO↗

arXiv2025

A strengthening on consecutive odd cycles in graphs of given minimum degree

Liu and Ma [J. Combin. Theory Ser. B, 2018] conjectured that every $2$-connected non-bipartite graph with minimum degree at least $k+1$ contains $\lceil k/2\rceil $ cycles with consecutive odd lengths. In particular, they showed that this conjecture holds when $k$ is even. In this paper, we confirm this conjecture for any $k\in \mathbb N$. Moreover, we also improve some previous results about cycles of consecutive lengths.

math.CO↗

arXiv2025

SurfAAV: Design and Implementation of a Novel Multimodal Surfing Aquatic-Aerial Vehicle

Despite significant advancements in the research of aquatic-aerial robots, existing configurations struggle to efficiently perform underwater, surface, and aerial movement simultaneously. In this paper, we propose a novel multimodal surfing aquatic-aerial vehicle, SurfAAV, which efficiently integrates underwater navigation, surface gliding, and aerial flying capabilities. Thanks to the design of the novel differential thrust vectoring hydrofoil, SurfAAV can achieve efficient surface gliding and underwater navigation without the need for a buoyancy adjustment system. This design provides flexible operational capabilities for both surface and underwater tasks, enabling the robot to quickly carry out underwater monitoring activities. Additionally, when it is necessary to reach another water body, SurfAAV can switch to aerial mode through a gliding takeoff, flying to the target water area to perform corresponding tasks. The main contribution of this letter lies in proposing a new solution for underwater, surface, and aerial movement, designing a novel hybrid prototype concept, developing the required control laws, and validating the robot's ability to successfully perform surface gliding and gliding takeoff. SurfAAV achieves a maximum surface gliding speed of 7.96 m/s and a maximum underwater speed of 3.1 m/s. The prototype's surface gliding maneuverability and underwater cruising maneuverability both exceed those of existing aquatic-aerial vehicles.

cs.RO↗

arXiv2025

RollingQ: Reviving the Cooperation Dynamics in Multimodal Transformer

Multimodal learning faces challenges in effectively fusing information from diverse modalities, especially when modality quality varies across samples. Dynamic fusion strategies, such as attention mechanism in Transformers, aim to address such challenge by adaptively emphasizing modalities based on the characteristics of input data. However, through amounts of carefully designed experiments, we surprisingly observed that the dynamic adaptability of widely-used self-attention models diminishes. Model tends to prefer one modality regardless of data characteristics. This bias triggers a self-reinforcing cycle that progressively overemphasizes the favored modality, widening the distribution gap in attention keys across modalities and deactivating attention mechanism's dynamic properties. To revive adaptability, we propose a simple yet effective method Rolling Query (RollingQ), which balances attention allocation by rotating the query to break the self-reinforcing cycle and mitigate the key distribution gap. Extensive experiments on various multimodal scenarios validate the effectiveness of RollingQ and the restoration of cooperation dynamics is pivotal for enhancing the broader capabilities of widely deployed multimodal Transformers. The source code is available at https://github.com/GeWu-Lab/RollingQ_ICML2025.

cs.LG↗

arXiv2025

DriveAgent: Multi-Agent Structured Reasoning with LLM and Multimodal Sensor Fusion for Autonomous Driving

We introduce DriveAgent, a novel multi-agent autonomous driving framework that leverages large language model (LLM) reasoning combined with multimodal sensor fusion to enhance situational understanding and decision-making. DriveAgent uniquely integrates diverse sensor modalities-including camera, LiDAR, GPS, and IMU-with LLM-driven analytical processes structured across specialized agents. The framework operates through a modular agent-based pipeline comprising four principal modules: (i) a descriptive analysis agent identifying critical sensor data events based on filtered timestamps, (ii) dedicated vehicle-level analysis conducted by LiDAR and vision agents that collaboratively assess vehicle conditions and movements, (iii) environmental reasoning and causal analysis agents explaining contextual changes and their underlying mechanisms, and (iv) an urgency-aware decision-generation agent prioritizing insights and proposing timely maneuvers. This modular design empowers the LLM to effectively coordinate specialized perception and reasoning agents, delivering cohesive, interpretable insights into complex autonomous driving scenarios. Extensive experiments on challenging autonomous driving datasets demonstrate that DriveAgent is achieving superior performance on multiple metrics against baseline methods. These results validate the efficacy of the proposed LLM-driven multi-agent sensor fusion framework, underscoring its potential to substantially enhance the robustness and reliability of autonomous driving systems.

cs.RO↗

arXiv2025

UniAP: Unifying Inter- and Intra-Layer Automatic Parallelism by Mixed Integer Quadratic Programming

Distributed learning is commonly used for training deep learning models, especially large models. In distributed learning, manual parallelism (MP) methods demand considerable human effort and have limited flexibility. Hence, automatic parallelism (AP) methods have recently been proposed for automating the parallel strategy optimization process. Existing AP methods suffer from sub-optimal solutions because they do not jointly optimize the two categories of parallel strategies (i.e., inter-layer parallelism and intra-layer parallelism). In this paper, we propose a novel AP method called UniAP, which unifies inter- and intra-layer automatic parallelism by mixed integer quadratic programming. To the best of our knowledge, UniAP is the first parallel method that can jointly optimize the two categories of parallel strategies to find an optimal solution. Experimental results show that UniAP outperforms state-of-the-art methods by up to 3.80$\times$ in throughput and reduces strategy optimization time by up to 107$\times$ across five Transformer-based models.

cs.LG↗

arXiv2025

Realization of a Dirac-vortex topological photonic crystal fiber

Photonic crystal fibers (PCFs) that trap and guide light using photonic bandgaps have revolutionized modern optics with enormous scientific innovations and technological applications spanning many disciplines. Recently, inspired by the discovery of topological phases of matter, Dirac-vortex topological PCFs have been theoretically proposed with intriguing topological properties and unprecedented opportunities in optical fiber communications. However, due to the substantial challenges of fabrication and characterization, experimental demonstration of Dirac-vortex topological PCFs has thus far remained elusive. Here, we report the experimental realization of a Dirac-vortex topological PCF using the standard stack-and-draw fabrication process with silica glass capillaries. Moreover, we experimentally observe that Dirac-vortex single-polarization single-mode bounds to and propagates along the fiber core in the full communication window (1260-1675nm). Our study pushes the research frontier of PCFs and provides a new avenue to enhance their performance and functionality further.

physics.optics↗