arXiv subjects

Yufei Sun

Publications and source records attributed to Yufei Sun.

15 recordsLinked to original sources

arXiv2026

SPT: Skills as Pre-Training Data for Agentic Language Models

Agentic (tool-using) language models are mainly trained on tool-call traces and agent trajectories during post-training. These data provide direct behavioral supervision, but producing them requires task environments, execution, and verification, making broad tool and task coverage expensive. Publicly available skills offer another source of training data: they encode reusable tool semantics and workflows but are typically used only as inference-time context. We introduce Skill Pre-Training (SPT), a mid-training method that applies causal language modeling to SkillCorpus, a collection of public multi-file skill packages, optionally mixed with general data. To preserve relations among files within each package, we also introduce Reference Insert, a reference-aware assembly strategy that places supporting files near their mentions in the primary instruction. Experiments across multiple model scales and post-training recipes show that SPT consistently improves agentic performance over mid-training on general or trajectory data, while largely preserving general performance. Data mixture experiments show additional benefits from combining skill data with general annealing corpora. These results indicate that skill packages are a valuable data source for pre-training agentic language models.

cs.CL↗

arXiv2026

ForeAct: Steering Your VLA with Efficient Visual Foresight Planning

Vision-Language-Action (VLA) models convert high-level language instructions into concrete, executable actions, a task that is especially challenging in open-world environments. We present Visual Foresight Planning (ForeAct), a general and efficient planner that guides a VLA step-by-step using imagined future observations and subtask descriptions. With an imagined future observation, the VLA can focus on visuo-motor inference rather than high-level semantic reasoning, leading to improved accuracy and generalization. Our planner comprises a highly efficient foresight image generation module that predicts a high-quality 640$\times$480 future observation from the current visual input and language instruction within only 0.33s on an H100 GPU, together with a vision-language model that reasons over the task and produces subtask descriptions for both the generator and the VLA. Importantly, state-of-the-art VLAs can integrate our planner seamlessly by simply augmenting their visual inputs, without any architectural modification. The foresight generator is pretrained on over 1 million multi-task, cross-embodiment episodes, enabling it to learn robust embodied dynamics. We evaluate our framework on a benchmark that consists of 11 diverse, multi-step real-world tasks. It achieves an average success rate of 87.4%, demonstrating a +40.9% absolute improvement over the $\pi_0$ baseline (46.5%) and a +30.3% absolute improvement over $\pi_0$ augmented with textual subtask guidance (57.1%).

cs.RO↗

arXiv2026

A Two-Stage GPU Kernel Tuner Combining Semantic Refactoring and Search-Based Optimization

GPU code optimization is a key performance bottleneck for HPC workloads as well as large-model training and inference. Although compiler optimizations and hand-written kernels can partially alleviate this issue, achieving near-hardware-limit performance still relies heavily on manual code refactoring and parameter tuning. Recent progress in LLM-agent-based kernel generation and optimization has been reported, yet many approaches primarily focus on direct code rewriting, where parameter choices are often implicit and hard to control, or require human intervention, leading to unstable performance gains. This paper introduces a template-based rewriting layer on top of an agent-driven iterative loop: kernels are semantically refactored into explicitly parameterizable templates, and template parameters are then optimized via search-based autotuning, yielding more stable and higher-quality speedups. Experiments on a set of real-world kernels demonstrate speedups exceeding 3x in the best case. We extract representative CUDA kernels from SGLang as evaluation targets; the proposed agentic tuner iteratively performs templating, testing, analysis, and planning, and leverages profiling feedback to execute constrained parameter search under hardware resource limits. Compared to agent-only direct rewriting, the template-plus-search design significantly reduces the randomness of iterative optimization, making the process more interpretable and enabling a more systematic approach toward high-performance configurations. The proposed method can be further extended to OpenCL, HIP, and other backends to deliver automated performance optimization for real production workloads.

cs.CL↗

arXiv2026

MVSS: A Unified Framework for Multi-View Structured Survey Generation

Scientific surveys require not only summarizing large bodies of literature, but also organizing them into clear and coherent conceptual structures. However, existing automatic survey generation methods typically focus on linear text generation and struggle to explicitly model hierarchical relations among research topics and structured methodological comparisons, resulting in substantial gaps in structural organization and evidence presentation compared to expert-written surveys. To address this limitation, we propose MVSS, a multi-view structured survey generation framework that jointly generates and aligns citation-grounded hierarchical trees, structured comparison tables, and survey text. MVSS follows a structure-first paradigm: it first constructs a tree that captures the conceptual organization of a research domain, then generates comparison tables constrained by the tree structure, and finally uses both the tree and tables as joint structural constraints to guide outline construction and survey text generation. This design enables complementary and aligned multi-view representations across structure, comparison, and narrative. In addition, we introduce a dedicated evaluation framework that systematically assesses generated surveys from multiple dimensions, including structural quality, comparative completeness, and citation fidelity. Through large-scale experiments on 76 computer science topics, we demonstrate that MVSS significantly outperforms existing methods in survey organization and evidence grounding, and achieves performance comparable to expert-written surveys across multiple evaluation metrics.

cs.CL↗

arXiv2025

MM-UAVBench: How Well Do Multimodal Large Language Models See, Think, and Plan in Low-Altitude UAV Scenarios?

While Multimodal Large Language Models (MLLMs) have exhibited remarkable general intelligence across diverse domains, their potential in low-altitude applications dominated by Unmanned Aerial Vehicles (UAVs) remains largely underexplored. Existing MLLM benchmarks rarely cover the unique challenges of low-altitude scenarios, while UAV-related evaluations mainly focus on specific tasks such as localization or navigation, without a unified evaluation of MLLMs'general intelligence. To bridge this gap, we present MM-UAVBench, a comprehensive benchmark that systematically evaluates MLLMs across three core capability dimensions-perception, cognition, and planning-in low-altitude UAV scenarios. MM-UAVBench comprises 19 sub-tasks with over 5.7K manually annotated questions, all derived from real-world UAV data collected from public datasets. Extensive experiments on 16 open-source and proprietary MLLMs reveal that current models struggle to adapt to the complex visual and cognitive demands of low-altitude scenarios. Our analyses further uncover critical bottlenecks such as spatial bias and multi-view understanding that hinder the effective deployment of MLLMs in UAV scenarios. We hope MM-UAVBench will foster future research on robust and reliable MLLMs for real-world UAV intelligence.

cs.CV↗

arXiv2025

VLASH: Real-Time VLAs via Future-State-Aware Asynchronous Inference

Vision-Language-Action models (VLAs) are becoming increasingly capable across diverse robotic tasks. However, these models are typically deployed under synchronous inference, where the robot waits for model inference to complete before acting, and cannot perceive or respond to environmental changes during action execution. This not only introduces noticeable action stalls, but also significantly increases reaction latency, fundamentally limiting the applicability of VLAs to dynamic, real-time tasks. Asynchronous inference offers a promising solution to achieve continuous and low-latency control by enabling robots to execute actions and perform inference simultaneously. However, because the robot and environment continue to evolve during inference, a temporal misalignment arises between the prediction and execution intervals. This leads to significant action instability, while existing asynchronous methods either degrade accuracy or introduce runtime overhead to mitigate it. We propose VLASH, a simple yet effective method for asynchronous VLA inference that delivers smooth, accurate, and fast reaction control without architectural changes or additional runtime overhead. VLASH leverages the future execution-time state by rolling the robot state forward with the previous action chunk, thereby bridging the gap between prediction and execution. Experiments show that VLASH reduces reaction latency by up to 11.8x compared to synchronous inference and consistently outperforms all asynchronous baselines in accuracy. With action quantization, it further achieves 1.5-2.0x task completion speedup with minimal accuracy loss. Moreover, it empowers state-of-the-art VLAs such as $\pi_{0.5}$ to handle fast-reaction, high-precision tasks including playing ping-pong and playing whack-a-mole, where traditional synchronous inference fails. Code is available at https://github.com/mit-han-lab/vlash

cs.RO↗

arXiv2025

LiveSecBench: A Dynamic and Event-Driven Safety Benchmark for Chinese Language Model Applications

We introduce LiveSecBench, a continuously updated safety benchmark specifically for Chinese-language LLM application scenarios. LiveSecBench constructs a high-quality and unique dataset through a pipeline that combines automated generation with human verification. By periodically releasing new versions to expand the dataset and update evaluation metrics, LiveSecBench provides a robust and up-to-date standard for AI safety. In this report, we introduce our second release v251215, which evaluates across five dimensions (Public Safety, Fairness & Bias, Privacy, Truthfulness, and Mental Health Safety.) We evaluate 57 representative LLMs using an ELO rating system, offering a leaderboard of the current state of Chinese LLM safety. The result is available at https://livesecbench.intokentech.cn/.

cs.CL↗

arXiv2025

RedNote-Vibe: A Dataset for Capturing Temporal Dynamics of AI-Generated Text in Lifestyle Social Media

We introduce RedNote-Vibe, a dataset spanning five years (pre-LLM to July 2025) sourced from lifestyle platform RedNote (Xiaohongshu), capturing the temporal dynamics of content creation and is enriched with comprehensive engagement metrics. To address the detection challenge posed by RedNote-Vibe, we propose the \textbf{PsychoLinguistic AIGT Detection Framework (PLAD)}. Grounded in cognitive psychology, PLAD leverages deep psychological signatures for robust and interpretable detection. Our experiments demonstrate PLAD's superior performance and reveal insights into content dynamics: (1) human content continues to outperform AI in emotionally resonant domains; (2) AI content is more homogeneous and rarely produces breaking posts, however, this human-AI gap narrows for arousing higher-investment interactions; and (3) most interestingly, a small group of users who strategically utilize AI tools can achieve higher engagement outcomes. The dataset is available at https://github.com/ydli-ai/RedNote-Vibe

cs.CL↗

arXiv2025

Invariant Graph Transformer for Out-of-Distribution Generalization

Graph Transformers (GTs) have demonstrated great effectiveness across various graph analytical tasks. However, the existing GTs focus on training and testing graph data originated from the same distribution, but fail to generalize under distribution shifts. Graph invariant learning, aiming to capture generalizable graph structural patterns with labels under distribution shifts, is potentially a promising solution, but how to design attention mechanisms and positional and structural encodings (PSEs) based on graph invariant learning principles remains challenging. To solve these challenges, we introduce Graph Out-Of-Distribution generalized Transformer (GOODFormer), aiming to learn generalized graph representations by capturing invariant relationships between predictive graph structures and labels through jointly optimizing three modules. Specifically, we first develop a GT-based entropy-guided invariant subgraph disentangler to separate invariant and variant subgraphs while preserving the sharpness of the attention function. Next, we design an evolving subgraph positional and structural encoder to effectively and efficiently capture the encoding information of dynamically changing subgraphs during training. Finally, we propose an invariant learning module utilizing subgraph node representations and encodings to derive generalizable graph representations that can to unseen graphs. We also provide theoretical justifications for our method. Extensive experiments on benchmark datasets demonstrate the superiority of our method over state-of-the-art baselines under distribution shifts.

cs.LG↗

arXiv2025

GraphRAG-R1: Graph Retrieval-Augmented Generation with Process-Constrained Reinforcement Learning

Graph Retrieval-Augmented Generation (GraphRAG) has shown great effectiveness in enhancing the reasoning abilities of LLMs by leveraging graph structures for knowledge representation and modeling complex real-world relationships. However, existing GraphRAG methods still face significant bottlenecks when handling complex problems that require multi-hop reasoning, as their query and retrieval phases are largely based on pre-defined heuristics and do not fully utilize the reasoning potentials of LLMs. To address this problem, we propose GraphRAG-R1, an adaptive GraphRAG framework by training LLMs with process-constrained outcome-based reinforcement learning (RL) to enhance the multi-hop reasoning ability. Our method can decompose complex problems, autonomously invoke retrieval tools to acquire necessary information, and perform effective reasoning. Specifically, we utilize a modified version of Group Relative Policy Optimization (GRPO) that supports rollout-with-thinking capability. Next, we design two process-constrained reward functions. To handle the shallow retrieval problem, we design a Progressive Retrieval Attenuation (PRA) reward to encourage essential retrievals. Then, to handle the over-thinking problem, we design Cost-Aware F1 (CAF) reward to balance the model performance with computational costs. We further design a phase-dependent training strategy, containing three training stages corresponding to cold start and these two rewards. Lastly, our method adopts a hybrid graph-textual retrieval to improve the reasoning capacity. Extensive experimental results demonstrate that GraphRAG-R1 boosts LLM capabilities in solving complex reasoning problems compared to state-of-the-art GraphRAG methods on both in-domain and out-of-domain datasets. Furthermore, our framework can be flexibly integrated with various existing retrieval methods, consistently delivering performance improvements.

cs.LG↗

arXiv2025

MrM: Black-Box Membership Inference Attacks against Multimodal RAG Systems

Multimodal retrieval-augmented generation (RAG) systems enhance large vision-language models by integrating cross-modal knowledge, enabling their increasing adoption across real-world multimodal tasks. These knowledge databases may contain sensitive information that requires privacy protection. However, multimodal RAG systems inherently grant external users indirect access to such data, making them potentially vulnerable to privacy attacks, particularly membership inference attacks (MIAs). % Existing MIA methods targeting RAG systems predominantly focus on the textual modality, while the visual modality remains relatively underexplored. To bridge this gap, we propose MrM, the first black-box MIA framework targeted at multimodal RAG systems. It utilizes a multi-object data perturbation framework constrained by counterfactual attacks, which can concurrently induce the RAG systems to retrieve the target data and generate information that leaks the membership information. Our method first employs an object-aware data perturbation method to constrain the perturbation to key semantics and ensure successful retrieval. Building on this, we design a counterfact-informed mask selection strategy to prioritize the most informative masked regions, aiming to eliminate the interference of model self-knowledge and amplify attack efficacy. Finally, we perform statistical membership inference by modeling query trials to extract features that reflect the reconstruction of masked semantics from response patterns. Experiments on two visual datasets and eight mainstream commercial visual-language models (e.g., GPT-4o, Gemini-2) demonstrate that MrM achieves consistently strong performance across both sample-level and set-level evaluations, and remains robust under adaptive defenses.

cs.CV↗

arXiv2025

HPSS: Heuristic Prompting Strategy Search for LLM Evaluators

Since the adoption of large language models (LLMs) for text evaluation has become increasingly prevalent in the field of natural language processing (NLP), a series of existing works attempt to optimize the prompts for LLM evaluators to improve their alignment with human judgment. However, their efforts are limited to optimizing individual factors of evaluation prompts, such as evaluation criteria or output formats, neglecting the combinatorial impact of multiple factors, which leads to insufficient optimization of the evaluation pipeline. Nevertheless, identifying well-behaved prompting strategies for adjusting multiple factors requires extensive enumeration. To this end, we comprehensively integrate 8 key factors for evaluation prompts and propose a novel automatic prompting strategy optimization method called Heuristic Prompting Strategy Search (HPSS). Inspired by the genetic algorithm, HPSS conducts an iterative search to find well-behaved prompting strategies for LLM evaluators. A heuristic function is employed to guide the search process, enhancing the performance of our algorithm. Extensive experiments across four evaluation tasks demonstrate the effectiveness of HPSS, consistently outperforming both human-designed evaluation prompts and existing automatic prompt optimization methods. Our code is available at https://github.com/thu-coai/HPSS.

cs.CL↗

arXiv2024

GNNRL-Smoothing: A Prior-Free Reinforcement Learning Model for Mesh Smoothing

Mesh smoothing methods can enhance mesh quality by eliminating distorted elements, leading to improved convergence in simulations. To balance the efficiency and robustness of traditional mesh smoothing process, previous approaches have employed supervised learning and reinforcement learning to train intelligent smoothing models. However, these methods heavily rely on labeled dataset or prior knowledge to guide the models' learning. Furthermore, their limited capacity to enhance mesh connectivity often restricts the effectiveness of smoothing. In this paper, we first systematically analyze the learning mechanisms of recent intelligent smoothing methods and propose a prior-free reinforcement learning model for intelligent mesh smoothing. Our proposed model integrates graph neural networks with reinforcement learning to implement an intelligent node smoothing agent and introduces, for the first time, a mesh connectivity improvement agent. We formalize mesh optimization as a Markov Decision Process and successfully train both agents using Twin Delayed Deep Deterministic Policy Gradient and Double Dueling Deep Q-Network in the absence of any prior data or knowledge. We verified the proposed model on both 2D and 3D meshes. Experimental results demonstrate that our model achieves feature-preserving smoothing on complex 3D surface meshes. It also achieves state-of-the-art results among intelligent smoothing methods on 2D meshes and is 7.16 times faster than traditional optimization-based smoothing methods. Moreover, the connectivity improvement agent can effectively enhance the quality distribution of the mesh.

cs.LG↗

arXiv2024

Deep Feature Surgery: Towards Accurate and Efficient Multi-Exit Networks

Multi-exit network is a promising architecture for efficient model inference by sharing backbone networks and weights among multiple exits. However, the gradient conflict of the shared weights results in sub-optimal accuracy. This paper introduces Deep Feature Surgery (\methodname), which consists of feature partitioning and feature referencing approaches to resolve gradient conflict issues during the training of multi-exit networks. The feature partitioning separates shared features along the depth axis among all exits to alleviate gradient conflict while simultaneously promoting joint optimization for each exit. Subsequently, feature referencing enhances multi-scale features for distinct exits across varying depths to improve the model accuracy. Furthermore, \methodname~reduces the training operations with the reduced complexity of backpropagation. Experimental results on Cifar100 and ImageNet datasets exhibit that \methodname~provides up to a \textbf{50.00\%} reduction in training time and attains up to a \textbf{6.94\%} enhancement in accuracy when contrasted with baseline methods across diverse models and tasks. Budgeted batch classification evaluation on MSDNet demonstrates that DFS uses about $\mathbf{2}\boldsymbol{\times}$ fewer average FLOPs per image to achieve the same classification accuracy as baseline methods on Cifar100. The code is available at https://github.com/GongCheng1919/dfs.

cs.CV↗

arXiv2021

Learning Fair Face Representation With Progressive Cross Transformer

Face recognition (FR) has made extraordinary progress owing to the advancement of deep convolutional neural networks. However, demographic bias among different racial cohorts still challenges the practical face recognition system. The race factor has been proven to be a dilemma for fair FR (FFR) as the subject-related specific attributes induce the classification bias whilst carrying some useful cues for FR. To mitigate racial bias and meantime preserve robust FR, we abstract face identity-related representation as a signal denoising problem and propose a progressive cross transformer (PCT) method for fair face recognition. Originating from the signal decomposition theory, we attempt to decouple face representation into i) identity-related components and ii) noisy/identity-unrelated components induced by race. As an extension of signal subspace decomposition, we formulate face decoupling as a generalized functional expression model to cross-predict face identity and race information. The face expression model is further concretized by designing dual cross-transformers to distill identity-related components and suppress racial noises. In order to refine face representation, we take a progressive face decoupling way to learn identity/race-specific transformations, so that identity-unrelated components induced by race could be better disentangled. We evaluate the proposed PCT on the public fair face recognition benchmarks (BFW, RFW) and verify that PCT is capable of mitigating bias in face recognition while achieving state-of-the-art FR performance. Besides, visualization results also show that the attention maps in PCT can well reveal the race-related/biased facial regions.

cs.CV↗