arXiv subjects

Yihao Sun

Publications and source records attributed to Yihao Sun.

At least 19 recordsLinked to original sources

arXiv2026

Terascale Query Processing in the Browser: Rethinking GPU Acceleration

Recursive query computation, central to graph algorithms and relational databases, demands GPU acceleration due to its inherent computational intensity. While substantial prior work addresses GPU implementations of recursive queries that require fixed-point evaluation, existing systems are restricted to native execution environments. We introduce WGLog, the first web-browser-native GPU engine for compute-bound recursive database queries. WGLog is built entirely on WebGPU compute shaders, a cross-platform API that enables GPU acceleration in web browsers. WGLog leverages two key technical innovations. First, we replace hash-table-based joins with atomic-free sorted-array joins, eliminating the serialization bottleneck that hash tables suffer on skewed graphs. Second, we develop an asynchronous execution pipeline using WebGPU's indirect dispatch capability, which eliminates GPU-host synchronizations that would otherwise dominate per-iteration overhead. On representative workloads, WGLog delivers a 1.48--4.68x speedup over native GPU systems and orders-of-magnitude improvement over CPU and WebAssembly implementations.

cs.DB↗

arXiv2026

AcademiClaw: When Students Set Challenges for AI Agents

Benchmarks within the OpenClaw ecosystem have thus far evaluated exclusively assistant-level tasks, leaving the academic-level capabilities of OpenClaw largely unexamined. We introduce AcademiClaw, a bilingual benchmark of 80 complex, long-horizon tasks sourced directly from university students' real academic workflows -- homework, research projects, competitions, and personal projects -- that they found current AI agents unable to solve effectively. Curated from 230 student-submitted candidates through rigorous expert review, the final task set spans 25+ professional domains, ranging from olympiad-level mathematics and linguistics problems to GPU-intensive reinforcement learning and full-stack system debugging, with 16 tasks requiring CUDA GPU execution. Each task executes in an isolated Docker sandbox and is scored on task completion by multi-dimensional rubrics combining six complementary techniques, with an independent five-category safety audit providing additional behavioral analysis. Experiments on six frontier models show that even the best achieves only a 55\% pass rate. Further analysis uncovers sharp capability boundaries across task domains, divergent behavioral strategies among models, and a disconnect between token consumption and output quality, providing fine-grained diagnostic signals beyond what aggregate metrics reveal. We hope that AcademiClaw and its open-sourced data and code can serve as a useful resource for the OpenClaw community, driving progress toward agents that are more capable and versatile across the full breadth of real-world academic demands. All data and code are available at https://github.com/GAIR-NLP/AcademiClaw.

cs.AI↗

arXiv2026

Adversarial Imitation Learning with General Function Approximation: Theoretical Analysis and Practical Algorithms

Adversarial imitation learning (AIL), a prominent approach in imitation learning, has achieved significant practical success powered by neural network approximation. However, existing theoretical analyses of AIL are primarily confined to simplified settings, such as tabular and linear function approximation, and involve complex algorithmic designs that impede practical implementation. This creates a substantial gap between theory and practice. This paper bridges this gap by exploring the theoretical underpinnings of online AIL with general function approximation. We introduce a novel framework called optimization-based AIL (OPT-AIL), which performs online optimization for reward learning coupled with optimism-regularized optimization for policy learning. Within this framework, we develop two concrete methods: model-free OPT-AIL and model-based OPT-AIL. Our theoretical analysis demonstrates that both variants achieve polynomial expert sample complexity and interaction complexity for learning near-expert policies. To the best of our knowledge, they represent the first provably efficient AIL methods under general function approximation. From a practical standpoint, OPT-AIL requires only the approximate optimization of two objectives, thereby facilitating practical implementation. Empirical studies demonstrate that OPT-AIL outperforms previous state-of-the-art deep AIL methods across several challenging tasks.

cs.LG↗

arXiv2026

Scaling Worst-Case Optimal Datalog to GPUs

Datalog is a declarative logic-programming language used for complex analytic reasoning workloads such as program analysis and graph analytics. Datalog's popularity is due to its unique price-point, marrying logic-defined specification with the potential for massive data parallelism. While traditional engines are CPU-based, the memory-bound nature of Datalog has led to increasing interest in leveraging GPUs. These engines beat CPU-based engines by operationalizing iterated relational joins via SIMT-friendly join algorithms. Unfortunately, all existing GPU Datalog engines are built on binary joins, which are inadequate for the complex multi-way queries arising in production systems such as DOOP and ddisasm. For these queries, binary decomposition can incur the AGM bound asymptotic blowup in time and space, leading to OOM failures regardless of join order. Worst-Case Optimal Joins (WCOJ) avoid this blowup, but their attribute-at-a-time intersections map poorly to SIMT hardware under key skew, causing severe load imbalance across Streaming Multiprocessors (SMs). We present SRDatalog, the first GPU Datalog engine based on WCOJ. SRDatalog uses flat columnar storage and two-phase deterministic memory allocation to avoid the OOM failures of binary joins and the index-rebuild overheads of static WCOJ systems. To mitigate skew and hide hardware stalls, SRDatalog further employs root-level histogram-guided load balancing, structural helper-relation splitting, and stream-aligned rule multiplexing. On real-world program-analysis workloads, SRDatalog achieves geometric-mean speedups of 21x to 47x.

cs.DB↗

arXiv2026

Superset Decompilation

Reverse engineering tools remain monolithic and imperative compared to the advancement of modern compiler architectures: analyses are tied to a single mutable representation, making them difficult to extend or refine, and forcing premature choices between soundness and precision. We observe that decompilation is the reverse of compilation and can be structured as a sequence of modular passes, each performing a granular and clearly defined interpretation of the binary at a progressively higher level of abstraction. We formalize this as provenance-guided superset decompilation (PGSD), a framework that monotonically derives facts about the binary into a relation store. Instead of committing early to a single interpretation, the pipeline retains ambiguous interpretations as parallel candidates with provenance, deferring resolution until the final selection phase. Manifold implements PGSD as a declarative reverse engineering framework that lifts Linux ELF binaries to C99 through a granular intermediate representation in ~35K lines of Rust and Datalog. On GNU coreutils, Manifold's output quality matches Ghidra, IDA Pro, angr, and RetDec on multiple metrics while producing fewer compiler errors, and generalizes across compilers and optimization levels.

cs.PL↗

arXiv2026

Towards Practical World Model-based Reinforcement Learning for Vision-Language-Action Models

Vision-Language-Action (VLA) models show strong generalization for robotic control, but finetuning them with reinforcement learning (RL) is constrained by the high cost and safety risks of real-world interaction. Training VLA models in interactive world models avoids these issues but introduces several challenges, including pixel-level world modeling, multi-view consistency, and compounding errors under sparse rewards. Building on recent advances across large multimodal models and model-based RL, we propose VLA-MBPO, a practical framework to tackle these problems in VLA finetuning. Our approach has three key design choices: (i) adapting unified multimodal models (UMMs) for data-efficient world modeling; (ii) an interleaved view decoding mechanism to enforce multi-view consistency; and (iii) chunk-level branched rollout to mitigate error compounding. Theoretical analysis and experiments across simulation and real-world tasks demonstrate that VLA-MBPO significantly improves policy performance and sample efficiency, underscoring its robustness and scalability for real-world robotic deployment.

cs.RO↗

arXiv2026

Speedup Patch: Learning a Plug-and-Play Policy to Accelerate Embodied Manipulation

While current embodied policies exhibit remarkable manipulation skills, their execution remains unsatisfactorily slow as they inherit the tardy pacing of human demonstrations. Existing acceleration methods typically require policy retraining or costly online interactions, limiting their scalability for large-scale foundation models. In this paper, we propose Speedup Patch (SuP), a lightweight, policy-agnostic framework that enables plug-and-play acceleration using solely offline data. SuP introduces an external scheduler that adaptively downsamples action chunks provided by embodied policies to eliminate redundancies. Specifically, we formalize the optimization of our scheduler as a Constrained Markov Decision Process (CMDP) aimed at maximizing efficiency without compromising task performance. Since direct success evaluation is infeasible in offline settings, SuP introduces World Model based state deviation as a surrogate metric to enforce safety constraints. By leveraging a learned world model as a virtual evaluator to predict counterfactual trajectories, the scheduler can be optimized via offline reinforcement learning. Empirical results on simulation benchmarks (Libero, Bigym) and real-world tasks validate that SuP achieves an overall 1.8x execution speedup for diverse policies while maintaining their original success rates.

cs.RO↗

arXiv2025

Skypilot: Fine-Tuning LLM with Physical Grounding for AAV Coverage Search

Autonomous aerial vehicles (AAVs) have played a pivotal role in coverage operations and search missions. Recent advances in large language models (LLMs) offer promising opportunities to augment AAV intelligence. These advances help address complex challenges like area coverage optimization, dynamic path planning, and adaptive decision-making. However, the absence of physical grounding in LLMs leads to hallucination and reproducibility problems in spatial reasoning and decision-making. To tackle these issues, we present Skypilot, an LLM-enhanced two-stage framework that grounds language models in physical reality by integrating monte carlo tree search (MCTS). In the first stage, we introduce a diversified action space that encompasses generate, regenerate, fine-tune, and evaluate operations, coupled with physics-informed reward functions to ensure trajectory feasibility. In the second stage, we fine-tune Qwen3-4B on 23,000 MCTS-generated samples, achieving substantial inference acceleration while maintaining solution quality. Extensive numerical simulations and real-world flight experiments validate the efficiency and superiority of our proposed approach. Detailed information and experimental results are accessible at https://sky-pilot.top.

cs.RO↗

arXiv2025

The Three Regimes of Offline-to-Online Reinforcement Learning

Offline-to-online reinforcement learning (RL) has emerged as a practical paradigm that leverages offline datasets for pretraining and online interactions for fine-tuning. However, its empirical behavior is highly inconsistent: design choices of online fine-tuning that work well in one setting can fail completely in another. Guided by the stability--plasticity principle, we propose a framework that can explain this inconsistency: We argue that efficient fine-tuning must preserve the utility of the stronger offline prior, whether that is the pretrained policy or the offline dataset, while maintaining sufficient plasticity. This perspective identifies three regimes of online fine-tuning, each requiring distinct stability properties. We validate this framework through a large-scale empirical study, finding that the results strongly align with its predictions in 45 out of 63 cases, with only 3 opposite mismatches. This work provides a framework for guiding design choices in offline-to-online RL based on the relative performance of the offline dataset and the pretrained policy.

cs.LG↗

arXiv2025

Planning with Unified Multimodal Models

With the powerful reasoning capabilities of large language models (LLMs) and vision-language models (VLMs), many recent works have explored using them for decision-making. However, most of these approaches rely solely on language-based reasoning, which limits their ability to reason and make informed decisions. Recently, a promising new direction has emerged with unified multimodal models (UMMs), which support both multimodal inputs and outputs. We believe such models have greater potential for decision-making by enabling reasoning through generated visual content. To this end, we propose Uni-Plan, a planning framework built on UMMs. Within this framework, a single model simultaneously serves as the policy, dynamics model, and value function. In addition, to avoid hallucinations in dynamics predictions, we present a novel approach self-discriminated filtering, where the generative model serves as a self-discriminator to filter out invalid dynamics predictions. Experiments on embodied decision-making tasks show that Uni-Plan substantially improves success rates compared to VLM-based methods, while also showing strong data scalability, requiring no expert demonstrations and achieving better performance under the same training-data size. This work lays a foundation for future research in reasoning and decision-making with UMMs.

cs.CV↗

arXiv2025

TeraSim: Uncovering Unknown Unsafe Events for Autonomous Vehicles through Generative Simulation

Traffic simulation is essential for autonomous vehicle (AV) development, enabling comprehensive safety evaluation across diverse driving conditions. However, traditional rule-based simulators struggle to capture complex human interactions, while data-driven approaches often fail to maintain long-term behavioral realism or generate diverse safety-critical events. To address these challenges, we propose TeraSim, an open-source, high-fidelity traffic simulation platform designed to uncover unknown unsafe events and efficiently estimate AV statistical performance metrics, such as crash rates. TeraSim is designed for seamless integration with third-party physics simulators and standalone AV stacks, to construct a complete AV simulation system. Experimental results demonstrate its effectiveness in generating diverse safety-critical events involving both static and dynamic agents, identifying hidden deficiencies in AV systems, and enabling statistical performance evaluation. These findings highlight TeraSim's potential as a practical tool for AV safety assessment, benefiting researchers, developers, and policymakers. The code is available at https://github.com/mcity/TeraSim.

cs.RO↗

arXiv2025

Column-Oriented Datalog on the GPU

Datalog is a logic programming language widely used in knowledge representation and reasoning (KRR), program analysis, and social media mining due to its expressiveness and high performance. Traditionally, Datalog engines use either row-oriented or column-oriented storage. Engines like VLog and Nemo favor column-oriented storage for efficiency on limited-resource machines, while row-oriented engines like Souffle use advanced data structures with locking to perform better on multi-core CPUs. The advent of modern datacenter GPUs, such as the NVIDIA H100 with its ability to run over 16k threads simultaneously and high memory bandwidth, has reopened the debate on which storage layout is more effective. This paper presents the first column-oriented Datalog engines tailored to the strengths of modern GPUs. We present VFLog, a CUDA-based Datalog runtime library with a column-oriented GPU datastructure that supports all necessary relational algebra operations. Our results demonstrate over 200x performance gains over SOTA CPU-based column-oriented Datalog engines and a 2.5x speedup over GPU Datalog engines in various workloads, including KRR.

cs.DB↗

arXiv2024

Datalog with First-Class Facts

Datalog is a popular logic programming language for deductive reasoning tasks in a wide array of applications, including business analytics, program analysis, and ontological reasoning. However, Datalog's restriction to flat facts over atomic constants leads to challenges in working with tree-structured data, such as derivation trees or abstract syntax trees. To ameliorate Datalog's restrictions, popular extensions of Datalog support features such as existential quantification in rule heads (Datalog$^\pm$, Datalog$^\exists$) or algebraic data types (Souffl\'e). Unfortunately, these are imperfect solutions for reasoning over structured and recursive data types, with general existentials leading to complex implementations requiring unification, and ADTs unable to trigger rule evaluation and failing to support efficient indexing. We present DL$^{\exists!}$, a Datalog with first-class facts, wherein every fact is identified with a Skolem term unique to the fact. We show that this restriction offers an attractive price point for Datalog-based reasoning over tree-shaped data, demonstrating its application to databases, artificial intelligence, and programming languages. We implemented DL$^{\exists!}$ as a system \slog{}, which leverages the uniqueness restriction of DL$^{\exists!}$ to enable a communication-avoiding, massively-parallel implementation built on MPI. We show that Slog outperforms leading systems (Nemo, Vlog, RDFox, and Souffl\'e) on a variety of benchmarks, with the potential to scale to thousands of threads.

cs.DB↗

arXiv2024

WHALE: Towards Generalizable and Scalable World Models for Embodied Decision-making

World models play a crucial role in decision-making within embodied environments, enabling cost-free explorations that would otherwise be expensive in the real world. To facilitate effective decision-making, world models must be equipped with strong generalizability to support faithful imagination in out-of-distribution (OOD) regions and provide reliable uncertainty estimation to assess the credibility of the simulated experiences, both of which present significant challenges for prior scalable approaches. This paper introduces WHALE, a framework for learning generalizable world models, consisting of two key techniques: behavior-conditioning and retracing-rollout. Behavior-conditioning addresses the policy distribution shift, one of the primary sources of the world model generalization error, while retracing-rollout enables efficient uncertainty estimation without the necessity of model ensembles. These techniques are universal and can be combined with any neural network architecture for world model learning. Incorporating these two techniques, we present Whale-ST, a scalable spatial-temporal transformer-based world model with enhanced generalizability. We demonstrate the superiority of Whale-ST in simulation tasks by evaluating both value estimation accuracy and video generation fidelity. Additionally, we examine the effectiveness of our uncertainty estimation technique, which enhances model-based policy optimization in fully offline scenarios. Furthermore, we propose Whale-X, a 414M parameter world model trained on 970K trajectories from Open X-Embodiment datasets. We show that Whale-X exhibits promising scalability and strong generalizability in real-world manipulation scenarios using minimal demonstrations.

cs.LG↗

arXiv2024

Provably and Practically Efficient Adversarial Imitation Learning with General Function Approximation

As a prominent category of imitation learning methods, adversarial imitation learning (AIL) has garnered significant practical success powered by neural network approximation. However, existing theoretical studies on AIL are primarily limited to simplified scenarios such as tabular and linear function approximation and involve complex algorithmic designs that hinder practical implementation, highlighting a gap between theory and practice. In this paper, we explore the theoretical underpinnings of online AIL with general function approximation. We introduce a new method called optimization-based AIL (OPT-AIL), which centers on performing online optimization for reward functions and optimism-regularized Bellman error minimization for Q-value functions. Theoretically, we prove that OPT-AIL achieves polynomial expert sample complexity and interaction complexity for learning near-expert policies. To our best knowledge, OPT-AIL is the first provably efficient AIL method with general function approximation. Practically, OPT-AIL only requires the approximate optimization of two objectives, thereby facilitating practical implementation. Empirical studies demonstrate that OPT-AIL outperforms previous state-of-the-art deep AIL methods in several challenging tasks.

cs.LG↗

arXiv2024

Any-step Dynamics Model Improves Future Predictions for Online and Offline Reinforcement Learning

Model-based methods in reinforcement learning offer a promising approach to enhance data efficiency by facilitating policy exploration within a dynamics model. However, accurately predicting sequential steps in the dynamics model remains a challenge due to the bootstrapping prediction, which attributes the next state to the prediction of the current state. This leads to accumulated errors during model roll-out. In this paper, we propose the Any-step Dynamics Model (ADM) to mitigate the compounding error by reducing bootstrapping prediction to direct prediction. ADM allows for the use of variable-length plans as inputs for predicting future states without frequent bootstrapping. We design two algorithms, ADMPO-ON and ADMPO-OFF, which apply ADM in online and offline model-based frameworks, respectively. In the online setting, ADMPO-ON demonstrates improved sample efficiency compared to previous state-of-the-art methods. In the offline setting, ADMPO-OFF not only demonstrates superior performance compared to recent state-of-the-art offline approaches but also offers better quantification of model uncertainty using only a single ADM.

cs.LG↗

arXiv2024

Assemblage: Automatic Binary Dataset Construction for Machine Learning

Binary code is pervasive, and binary analysis is a key task in reverse engineering, malware classification, and vulnerability discovery. Unfortunately, while there exist large corpora of malicious binaries, obtaining high-quality corpora of benign binaries for modern systems has proven challenging (e.g., due to licensing issues). Consequently, machine learning based pipelines for binary analysis utilize either costly commercial corpora (e.g., VirusTotal) or open-source binaries (e.g., coreutils) available in limited quantities. To address these issues, we present Assemblage: an extensible cloud-based distributed system that crawls, configures, and builds Windows PE binaries to obtain high-quality binary corpuses suitable for training state-of-the-art models in binary analysis. We have run Assemblage on AWS over the past year, producing 890k Windows PE and 428k Linux ELF binaries across 29 configurations. Assemblage is designed to be both reproducible and extensible, enabling users to publish "recipes" for their datasets, and facilitating the extraction of a wide array of features. We evaluated Assemblage by using its data to train modern learning-based pipelines for compiler provenance and binary function similarity. Our results illustrate the practical need for robust corpora of high-quality Windows PE binaries in training modern learning-based binary analyses. Assemblage code is open sourced under the MIT license, and the dataset can be downloaded from https://assemblage-dataset.net

cs.CR↗

arXiv2023

Episodic Return Decomposition by Difference of Implicitly Assigned Sub-Trajectory Reward

Real-world decision-making problems are usually accompanied by delayed rewards, which affects the sample efficiency of Reinforcement Learning, especially in the extremely delayed case where the only feedback is the episodic reward obtained at the end of an episode. Episodic return decomposition is a promising way to deal with the episodic-reward setting. Several corresponding algorithms have shown remarkable effectiveness of the learned step-wise proxy rewards from return decomposition. However, these existing methods lack either attribution or representation capacity, leading to inefficient decomposition in the case of long-term episodes. In this paper, we propose a novel episodic return decomposition method called Diaster (Difference of implicitly assigned sub-trajectory reward). Diaster decomposes any episodic reward into credits of two divided sub-trajectories at any cut point, and the step-wise proxy rewards come from differences in expectation. We theoretically and empirically verify that the decomposed proxy reward function can guide the policy to be nearly optimal. Experimental results show that our method outperforms previous state-of-the-art methods in terms of both sample efficiency and performance.

cs.LG↗