arXiv subjects

Chao Jin

Publications and source records attributed to Chao Jin.

At least 19 recordsLinked to original sources

arXiv2026

FESC: Remodeling Long-Context Private Inference with Encrypted State-Space Models

Processing long, sensitive documents with machine-learning models requires efficient, privacy-preserving long-context inference. Prior private inference systems optimize or distribute encrypted Transformer attention, but its quadratic token-pair work remains the bottleneck as sequence length grows. Selective state-space models (SSMs) offer linear-time recurrence, yet direct encrypted implementation incurs linear multiplicative depth, sequence-wide state residency, or dense FHE-MPC conversion. We present Factorized Encrypted Scan-Contract (FESC), a hybrid FHE-MPC system for private long-context selective SSM inference. Its factorized scan-contract keeps input-dependent transitions compact across conversion boundaries, composes them without dense expansion, streams state chunks on demand, and contracts outputs before conversion. We demonstrate interface compatibility of the scan-contract implementation across invariant and selective SSM architectures. For our Mamba-2 instantiation, we design GPU-optimized CKKS kernels for linear computations, MPC protocols for SiLU, softplus, exponential, and RMSNorm, with approximation-aware fine-tuning. To our knowledge, FESC is the first private long-document inference system to complete native end-to-end execution at $L \geq 1{,}024$ on a single GPU. At $L = 2{,}048$, a 12-layer Mamba-base model completes inference in 77.3 minutes on one A100 GPU with a peak memory footprint of 32.7 GB, while maintaining near-plaintext accuracy on the evaluated long-document tasks.

cs.CR↗

arXiv2026

ExpertPlex: A High-Goodput Disaggregated Serving System for MoE LLMs with Adaptive Persistent Kernels

LLMs scale Mixture-of-Experts (MoE) parameters for superior intelligence, but massive weights and dynamic computation impede efficient serving. Existing instance-level prefill-decode disaggregation isolates the phases on separate full-model replicas. As MoE weights grow, each instance may span tens to hundreds of GPUs, making resource allocation increasingly coarse. Configured prefill-to-decode ratios thus often mismatch demand, overprovisioning one phase while overloading the other. Prefill-decode colocation avoids this duplication, but existing Green Context solutions partition each GPU by phase and fix phase resources during a kernel. They cannot track resource changes across operations or layerwise variation in routed expert load, causing head-of-line blocking or idle reserved resources. Partitioning every GPU also leaves each phase with fewer local resources, forces wider parallelism and more communication, and lets prefill and decode traffic interfere on the shared network. We present ExpertPlex, which shares massive MoE experts across phases while disaggregating lightweight attention modules. Expert sharing eliminates over 95% of duplicate model weights and multiplexes dynamically sparse computation, while attention disaggregation reduces attention communication cost. ExpertPlex further uses (1) adaptive persistent kernels to schedule dynamic expert computation at tile granularity for efficient, isolated execution; (2) attention-initiated MoE communication to avoid network interference and enable cross-phase communication-computation overlap; and (3) a tile-to-cluster model to optimize these mechanisms for maximum goodput. Experiments serving MiniMax-M2.7 and GLM-5.1-FP8 show that ExpertPlex improves goodput by up to 2.01$\times$ over instance-level prefill-decode disaggregation and 1.66$\times$ over prefill-decode colocation.

cs.DC↗

arXiv2026

AdaThinking-E: One-Token Entropy Regulation for Adaptive Thinking

Multimodal large language models have demonstrated strong document reasoning capabilities by incorporating explicit thinking processes. While this capability significantly improves performance on challenging tasks, current models apply such deep reasoning uniformly to all questions, resulting in unnecessary computational overhead for simple task. This not only degrades user experience but also negatively impact accuracy on benchmark datasets. We identify the critical need for adaptive thinking mechanisms that can intelligently determine when to engage reasoning based on question complexity. To address this, we propose AdaThinking-E, a novel reinforcement learning framework that learns adaptive thinking through one-token entropy regulation. Our key insight is that model confidence in the decision to engage thinking (or not) can be quantified through entropy analysis of the predicted probability distribution at critical decision tokens. This observation motivates our entropy-governed reward mechanism: the training process naturally transitions from high-entropy exploration, where the model experiments with different thinking strategies, to low-entropy convergence with confident, generalizable decision-making policies. Crucially, this approach enables models to intrinsically discover when to think without requiring manual intervention or external difficulty labels. Extensive experiments demonstrate that our approach enables models to be both accurate on complex problems and efficient on simple ones across diverse document tasks.

cs.CL↗

arXiv2026

UltraEP: Unleash MoE Training and Inference on Rack-Scale Nodes with Near-Optimal Load Balancing

Large-scale expert parallelism (EP) is becoming pivotal for training and serving frontier MoE models, but it also amplifies device-level expert load imbalance into compute stragglers, token all-to-all bottlenecks, and activation-memory spikes. Existing balancers redistribute experts periodically based on historical load, which becomes unreliable for production deployments with non-stationary load patterns. We present UltraEP, the first exact-load, real-time balancer for large-EP MoE training and serving prefill on rack-scale nodes (RSNs). Leveraging the extended scale-up connectivity among dozens of GPUs within RSNs, UltraEP rebalances every microbatch and layer on critical paths, which requires nontrivial co-design of plan solving and expert replication communication to minimize exposed overhead. To this end, UltraEP eagerly reacts to post-gating load with an efficient quota-driven planner, and executes the resulting irregular expert-state transfers with RSN-native persistent tile streaming and relay-based fan-out mitigation. We evaluate UltraEP in a multi-RSN deployment of up to 256 GPUs, using cutting-edge MoE models from 106B to 671B parameters. Averaged across training and serving, UltraEP achieves 94.3% of the force-balanced ideal throughput, delivering 1.49$\times$ improvement over no-balancing, while reducing the final inter-rank imbalance from 1.30$-$4.01 to 1.01$-$1.04.

cs.DC↗

arXiv2026

BigMac: Breaking the Pareto Frontier of Compute and Memory in Multimodal LLM Training

Training multimodal large language models (MLLMs) is challenged by both model and data heterogeneity. Existing systems redesign the training pipeline to address these challenges, but remain bound by a Pareto frontier between compute and memory efficiency, improving one only at the expense of the other. We present BigMac, a new training pipeline for multimodal LLMs. The core idea of BigMac is to elegantly nest the encoder and generator computation into the original LLM pipeline, forming a dependency-safe nested pipeline structure. With this design, BigMac reduces the activation memory complexity of the encoder and generator to O(1) while keeping the activation memory complexity of the LLM unchanged. At the same time, it achieves the same computational efficiency as the idealized setting with unlimited memory. As a result, BigMac breaks the Pareto frontier between computational efficiency and memory usage, enabling simultaneous optimization of both computation and memory in MLLM training. We evaluate BigMac on multiple MLLMs and training workloads. Experimental results show that BigMac achieves a 1.08$\times$-1.9$\times$ training speedup over baseline systems while maintaining stable memory usage as batch size increases.

cs.LG↗

arXiv2026

ReLibra: Routing-Replay-Guided Load Balancing for MoE Training in Reinforcement Learning

Load imbalance is a long-standing challenge in Mixture-of-Experts (MoE) training and is exacerbated in reinforcement learning (RL) for LLMs, where hot experts can shift frequently across micro-batches. Existing MoE training systems rely on historical loads to predict future expert demand, making them less effective under sharp fluctuations. We propose ReLibra, an MoE RL training system that exploits a unique opportunity in RL's rollout-training workflow, routing replay, to enable fine-grained load balancing at micro-batch granularity. Because rollout and training process the same tokens with the same MoE parameters, the token-to-expert routing decisions are known before training starts. Leveraging this information, ReLibra places two MoE load-balancing mechanisms at inter- and intra-batch timescales, matching their communication patterns to hierarchical network bandwidths. At the inter-batch timescale, ReLibra performs expert reordering to redistribute experts for batch-level cross-node balancing; at the intra-batch timescale, it dynamically performs expert replication within a node to absorb micro-batch-level load fluctuations. Experiments on diverse MoE LLMs and RL workloads show that ReLibra improves training throughput by up to 1.6$\times$ over Megatron-LM and by up to 1.2$\times$ over EPLB, even when EPLB is given oracle loads. Moreover, ReLibra remains within 6%-10% of the throughput of an idealized balanced baseline.

cs.LG↗

arXiv2026

HalluClear: Diagnosing, Evaluating and Mitigating Hallucinations in GUI Agents

While progress in GUI agents has been largely driven by industrial-scale training, ungrounded hallucinations often trigger cascading failures in real-world deployments.Unlike general VLM domains, the GUI agent field lacks a hallucination-focused suite for fine-grained diagnosis, reliable evaluation, and targeted mitigation.To bridge this gap, we introduce HalluClear, a comprehensive suite for hallucination mitigation in GUI agents as a complement to computation-intensive scaling. HalluClear comprises: (1) a GUI-specific hallucination taxonomy derived from empirical failure analysis; (2) a calibrated three-stage evaluation workflow which enhances VLM-as-a-judge reliability via expert-annotated benchmarking and ensemble credibility estimation; and (3) a mitigation scheme based on closed-loop structured reasoning, enabling lightweight continual post-training with cold-start initialization for both generalist and GUI-specialist agents. Experiments across representative agents and public benchmarks demonstrate that post-training on only 9K samples within our suite can significantly reduce hallucinations, thereby improving grounding and action fidelity, offering a compute-efficient pathway to robust GUI automation.

cs.AI↗

arXiv2026

EncFormer: Secure and Efficient Transformer Inference over Encrypted Data

Transformer inference in machine-learning-as-a-service (MLaaS) raises privacy concerns for sensitive user inputs. Prior secure solutions that combine fully homomorphic encryption (FHE) and secure multiparty computation (MPC) are bottlenecked by inefficient FHE kernels, communication-heavy MPC protocols, and expensive FHE-MPC conversions. We present EncFormer, a two-party private Transformer inference framework that introduces Stage Compatible Patterns so that FHE kernels compose efficiently, reducing repacking and conversions. EncFormer also provides a cost analysis model built around a minimal-conversion baseline, enabling principled selection of FHE-MPC boundaries. To further reduce communication, EncFormer proposes a secure complex CKKS-MPC conversion protocol and designs communication-efficient MPC protocols for nonlinearities. With GPU optimizations, evaluations on GPT- and BERT-style models show that EncFormer achieves 1.4x-30.4x lower online MPC communication and 1.3x-9.8x lower end-to-end latency against prior hybrid FHE-MPC systems, and 1.9x-3.5x lower end-to-end latency on BERT-base than FHE-only pipelines under a matched backend, while maintaining near-plaintext accuracy on selected GLUE tasks.

cs.CR↗

arXiv2026

Are GUI Agents Focused Enough? Automated Distraction via Semantic-level UI Element Injection

Existing red-teaming studies on GUI agents face two fundamental limitations: adversarial perturbations require white-box access unavailable in commercial deployments, while prompt injection is increasingly neutralized by stronger safety alignment. To study robustness under a more practical threat model, we propose Semantic-level UI Element Injection, a black-box red-teaming paradigm that overlays safety-aligned and harmless UI elements onto screenshots to misdirect the agent's visual grounding. Our method couples a modular Editor--Overlapper--Victim pipeline with iterative search that samples multiple candidate edits, keeps the best cumulative overlay, and adapts future prompt strategies based on previous failures. Experiments across 19 victim models spanning 8 model families show that strategic optimization substantially outperforms random injection (3.5-6.9x on the most robust victims) and transfers near-perfectly across architectures, confirming model-agnostic visual-semantic vulnerabilities. After the first successful attack, the victim still clicks the attacker-controlled icon in over 15\% of subsequent independent trials versus below 1% for random injection, establishing that strategically placed icons act as persistent attractors that causally redirect grounding rather than introducing incidental clutter.

cs.CR↗

arXiv2026

Heddle: A Distributed Orchestration System for Agentic RL Rollout

Agentic Reinforcement Learning (RL) enables LLMs to solve complex tasks by alternating between a data-collection rollout phase and a policy training phase. During rollout, the agent generates trajectories, i.e., multi-step interactions between LLMs and external tools. Yet, frequent tool calls induce long-tailed trajectory generation that bottlenecks rollouts. This stems from step-centric designs that ignore trajectory context, triggering three system problems for long-tail trajectory generation: queueing delays, interference overhead, and inflated per-token time. We propose Heddle, a trajectory-centric system to optimize the when, where, and how of agentic rollout execution. Heddle integrates three core mechanisms: trajectory-level scheduling using runtime prediction and progressive priority to minimize cumulative queueing; trajectory-aware placement via presorted dynamic programming and opportunistic migration during idle tool call intervals to minimize interference; and trajectory-adaptive resource manager that dynamically tunes model parallelism to accelerate the per-token time of long-tail trajectories while maintaining high throughput for short trajectories. Evaluations across diverse agentic RL workloads demonstrate that Heddle effectively neutralizes the long-tail bottleneck, achieving up to 2.5$\times$ higher end-to-end rollout throughput compared to state-of-the-art baselines.

cs.LG↗

arXiv2025

MegaScale-MoE: Large-Scale Communication-Efficient Training of Mixture-of-Experts Models in Production

We present MegaScale-MoE, a production system tailored for the efficient training of large-scale mixture-of-experts (MoE) models. MoE emerges as a promising architecture to scale large language models (LLMs) to unprecedented sizes, thereby enhancing model performance. However, existing MoE training systems experience a degradation in training efficiency, exacerbated by the escalating scale of MoE models and the continuous evolution of hardware. Recognizing the pivotal role of efficient communication in enhancing MoE training, MegaScale-MoE customizes communication-efficient parallelism strategies for attention and FFNs in each MoE layer and adopts a holistic approach to overlap communication with computation at both inter- and intra-operator levels. Additionally, MegaScale-MoE applies communication compression with adjusted communication patterns to lower precision, further improving training efficiency. When training a 352B MoE model on 1,440 NVIDIA Hopper GPUs, MegaScale-MoE achieves a training throughput of 1.41M tokens/s, improving the efficiency by 1.88$\times$ compared to Megatron-LM. We share our operational experience in accelerating MoE training and hope that by offering our insights in system design, this work will motivate future research in MoE systems.

cs.LG↗

arXiv2025

StreamRL: Scalable, Heterogeneous, and Elastic RL for LLMs with Disaggregated Stream Generation

Reinforcement learning (RL) has become the core post-training technique for large language models (LLMs). RL for LLMs involves two stages: generation and training. The LLM first generates samples online, which are then used to derive rewards for training. The conventional view holds that the colocated architecture, where the two stages share resources via temporal multiplexing, outperforms the disaggregated architecture, in which dedicated resources are assigned to each stage. However, in real-world deployments, we observe that the colocated architecture suffers from resource coupling, where the two stages are constrained to use the same resources. This coupling compromises the scalability and cost-efficiency of colocated RL in large-scale training. In contrast, the disaggregated architecture allows for flexible resource allocation, supports heterogeneous training setups, and facilitates cross-datacenter deployment. StreamRL is designed with disaggregation from first principles and fully unlocks its potential by addressing two types of performance bottlenecks in existing disaggregated RL frameworks: pipeline bubbles, caused by stage dependencies, and skewness bubbles, resulting from long-tail output length distributions. To address pipeline bubbles, StreamRL breaks the traditional stage boundary in synchronous RL algorithms through stream generation and achieves full overlapping in asynchronous RL. To address skewness bubbles, StreamRL employs an output-length ranker model to identify long-tail samples and reduces generation time via skewness-aware dispatching and scheduling. Experiments show that StreamRL improves throughput by up to 2.66x compared to existing state-of-the-art systems, and improves cost-effectiveness by up to 1.33x in a heterogeneous, cross-datacenter setting.

cs.LG↗

arXiv2025

Task Assignment and Exploration Optimization for Low Altitude UAV Rescue via Generative AI Enhanced Multi-agent Reinforcement Learning

The integration of emerging uncrewed aerial vehicles (UAVs) with artificial intelligence (AI) and ground-embedded robots (GERs) has transformed emergency rescue operations in unknown environments. However, the high computational demands often exceed a single UAV's capacity, making it difficult to continuously provide stable high-level services. To address this, this paper proposes a cooperation framework involving UAVs, GERs, and airships. The framework enables resource pooling through UAV-to-GER (U2G) and UAV-to-airship (U2A) links, offering computing services for offloaded tasks. Specifically, we formulate the multi-objective problem of task assignment and exploration as a dynamic long-term optimization problem aiming to minimize task completion time and energy use while ensuring stability. Using Lyapunov optimization, we transform it into a per-slot deterministic problem and propose HG-MADDPG, which combines the Hungarian algorithm with a GDM-based multi-agent deep deterministic policy gradient. Simulations demonstrate significant improvements in offloading efficiency, latency, and system stability over baselines.

cs.AI↗

arXiv2025

MegaScale-Infer: Serving Mixture-of-Experts at Scale with Disaggregated Expert Parallelism

Mixture-of-Experts (MoE) showcases tremendous potential to scale large language models (LLMs) with enhanced performance and reduced computational complexity. However, its sparsely activated architecture shifts feed-forward networks (FFNs) from being compute-intensive to memory-intensive during inference, leading to substantially lower GPU utilization and increased operational costs. We present MegaScale-Infer, an efficient and cost-effective system for serving large-scale MoE models. MegaScale-Infer disaggregates attention and FFN modules within each model layer, enabling independent scaling, tailored parallelism strategies, and heterogeneous deployment for both modules. To fully exploit disaggregation in the presence of MoE's sparsity, MegaScale-Infer introduces ping-pong pipeline parallelism, which partitions a request batch into micro-batches and shuttles them between attention and FFNs for inference. Combined with distinct model parallelism for each module, MegaScale-Infer effectively hides communication overhead and maximizes GPU utilization. To adapt to disaggregated attention and FFN modules and minimize data transmission overhead (e.g., token dispatch), MegaScale-Infer provides a high-performance M2N communication library that eliminates unnecessary GPU-to-CPU data copies, group initialization overhead, and GPU synchronization. Experimental results indicate that MegaScale-Infer achieves up to 1.90x higher per-GPU throughput than state-of-the-art solutions.

cs.DC↗

arXiv2025

HydraServe: Minimizing Cold Start Latency for Serverless LLM Serving in Public Clouds

With the proliferation of large language model (LLM) variants, developers are turning to serverless computing for cost-efficient LLM deployment. However, public cloud providers often struggle to provide performance guarantees for serverless LLM serving due to significant cold start latency caused by substantial model sizes and complex runtime dependencies. To address this problem, we present HydraServe, a serverless LLM serving system designed to minimize cold start latency in public clouds. HydraServe proactively distributes models across servers to quickly fetch them, and overlaps cold-start stages within workers to reduce startup latency. Additionally, HydraServe strategically places workers across GPUs to avoid network contention among cold-start instances. To minimize resource consumption during cold starts, HydraServe further introduces pipeline consolidation that can merge groups of workers into individual serving endpoints. Our comprehensive evaluations under diverse settings demonstrate that HydraServe reduces the cold start latency by 1.7$\times$-- 4.7$\times$ and improves service level objective attainment by 1.43$\times$--1.74$\times$ compared to baselines.

cs.DC↗

arXiv2025

Contextuality of Quantum Error-Correcting Codes

Universal fault-tolerant quantum computation requires overcoming the Eastin--Knill theorem on quantum error correction (QEC) codes that protect information from noise. This is often accomplished through strategies like magic state distillation, which prepares computational resources -- namely, magic states -- whose power is rooted in quantum contextuality, a fundamental nonclassical feature generalizing Bell nonlocality. Yet, the broader role of contextuality in enabling universality, including its significance as an inherent feature of QEC codes and protocols themselves, has remained largely unexplored. In this work, we develop a rigorous framework for contextuality in QEC and prove three main results. Fundamentally, we show that subsystem stabilizer codes with two or more gauge qubits are strongly contextual in their partial closure, while others are noncontextual, establishing a clear criterion for identifying contextual codes. Mathematically, we unify Abramsky--Brandenburger's sheaf-theoretic and Kirby--Love's tree-based definitions of contextuality, resolving a conjecture of Kim and Abramsky. Practically, we prove that many widely studied code-switching protocols which admit universal transversal gate sets, such as the doubled color codes introduced by Bravyi and Cross, are necessarily strongly contextual in their partial closure. Collectively, our results establish quantum contextuality as an intrinsic characteristic of fault-tolerant quantum codes and protocols, complementing entanglement and magic as resources for scalable quantum computation. For quantum coding theorists, this provides a new invariant: contextuality classifies which subsystem stabilizer codes can participate in universal fault-tolerant protocols. These findings position contextuality not only as a foundational concept but also as a practical guide for the design and analysis of future QEC architectures.

quant-ph↗

arXiv2025

Patch Triplet Similarity Purification for Guided Real-World Low-Dose CT Image Denoising

Image denoising of low-dose computed tomography (LDCT) is an important problem for clinical diagnosis with reduced radiation exposure. Previous methods are mostly trained with pairs of synthetic or misaligned LDCT and normal-dose CT (NDCT) images. However, trained with synthetic noise or misaligned LDCT/NDCT image pairs, the denoising networks would suffer from blurry structure or motion artifacts. Since non-contrast CT (NCCT) images share the content characteristics to the corresponding NDCT images in a three-phase scan, they can potentially provide useful information for real-world LDCT image denoising. To exploit this aspect, in this paper, we propose to incorporate clean NCCT images as useful guidance for the learning of real-world LDCT image denoising networks. To alleviate the issue of spatial misalignment in training data, we design a new Patch Triplet Similarity Purification (PTSP) strategy to select highly similar patch (instead of image) triplets of LDCT, NDCT, and NCCT images for network training. Furthermore, we modify two image denoising transformers of SwinIR and HAT to accommodate the NCCT image guidance, by replacing vanilla self-attention with cross-attention. On our collected clinical dataset, the modified transformers trained with the data selected by our PTSP strategy show better performance than 15 comparison methods on real-world LDCT image denoising. Ablation studies validate the effectiveness of our NCCT image guidance and PTSP strategy. We will publicly release our data and code.

eess.IV↗

arXiv2024

From Algorithm to Hardware: A Survey on Efficient and Safe Deployment of Deep Neural Networks

Deep neural networks (DNNs) have been widely used in many artificial intelligence (AI) tasks. However, deploying them brings significant challenges due to the huge cost of memory, energy, and computation. To address these challenges, researchers have developed various model compression techniques such as model quantization and model pruning. Recently, there has been a surge in research of compression methods to achieve model efficiency while retaining the performance. Furthermore, more and more works focus on customizing the DNN hardware accelerators to better leverage the model compression techniques. In addition to efficiency, preserving security and privacy is critical for deploying DNNs. However, the vast and diverse body of related works can be overwhelming. This inspires us to conduct a comprehensive survey on recent research toward the goal of high-performance, cost-efficient, and safe deployment of DNNs. Our survey first covers the mainstream model compression techniques such as model quantization, model pruning, knowledge distillation, and optimizations of non-linear operations. We then introduce recent advances in designing hardware accelerators that can adapt to efficient model compression approaches. Additionally, we discuss how homomorphic encryption can be integrated to secure DNN deployment. Finally, we discuss several issues, such as hardware evaluation, generalization, and integration of various compression approaches. Overall, we aim to provide a big picture of efficient DNNs, from algorithm to hardware accelerators and security perspectives.

cs.LG↗