arXiv subjects

Quanlu Zhang

Publications and source records attributed to Quanlu Zhang.

At least 19 recordsLinked to original sources

arXiv2026

NeoHorse-1: Towards Recursive Self-Improvement via Agentic Post-Training with Routing Harness

Recursive self-improvement (RSI) requires a concrete mechanism through which an AI system observes its capabilities and converts that evidence into the next round of learning. We present NeoHorse-1, a family of agent-native models developed to explore this path through agentic post-training. Our system combines a heterogeneous model pool with intelligent routing, recording the predicted capability demand, selected service tier, and subsequent interaction for each user turn. These records are converted into training examples that preserve interleaved reasoning, tool calls, and harness context, and are admitted through structural validation, six-dimensional semantic evaluation, and subscene-level labeling. Routing signals organize supervised fine-tuning into a three-stage curriculum and extend to routing-guided on-policy distillation, where a teacher supervises student-generated responses under the same progression. Capability-guided allocation then converts evaluation feedback into the next training mixture, closing an evaluation-selection-update loop in which what the system learns to do shapes what it learns from next. Across eleven benchmarks covering harness-based agents, tool use, coding, and instruction following, post-training raises the macro-average from 58.94 to 64.87 at 4B and from 65.60 to 69.04 at 9B, substantially narrowing the aggregate gap between the post-trained 4B model and the 9B base model. NeoHorse-1 provides an initial prototype of this feedback-driven process and a path toward harness-mediated RSI across successive iterations.

cs.CL↗

arXiv2026

STEAM: Self-Supervised Temporal Ensemble Advantage Modeling for Real-World Robot Learning

Real-world robot learning increasingly relies on heterogeneous data, but demonstrations and rollouts often mix useful progress with stalls, corrections, and suboptimal behavior. Effective policy learning therefore requires frame-level advantages that distinguish reliable local progress from failures and regressions. We propose Self-supervised Temporal Ensemble Advantage Modeling (STEAM), a label-free method that learns such advantages from expert demonstrations. STEAM trains an ensemble of temporal-offset predictors on frame pairs within expert trajectories, using the normalized temporal offset between two frames as a self-supervised signal. Each predictor maps a frame pair to a distribution over temporal offsets, which is converted into a scalar advantage. STEAM then takes the minimum advantage across the ensemble to score mixed-quality rollout data conservatively. Across real-world bimanual towel folding, chip checkout, cola restocking, and single-arm pick-and-place tasks, STEAM identifies stalls, failures, and recoveries. When combined with CFGRL, STEAM further improves policy success rate by 59%, 54.3%, 23% and 16.2% over baselines, respectively.

cs.RO↗

arXiv2026

DynaTrain: Fast Online Parallelism Switching for Elastic LLM Training

Modern large language model (LLM) training is inherently dynamic: resource fluctuations, RLHF phase shifts, and cluster elasticity continually reshape the optimal parallelism layout, posing a significant challenge to existing training frameworks built around a static execution model. We present DynaTrain, a distributed training system for sub-second, online reconfiguration across arbitrary multi-dimensional parallelism. At its core, we propose a Virtual Parameter Space (VPS) abstraction that unifies all distributed training states under one logical coordinate space, turning any parallelism configuration into a deterministic mapping and collapsing complex transition into manageable geometric intersections. On top of VPS, a state routing-and-transition layer executes rank-local transfers under a memory-aware, deadlock-free schedule, and an Elastic Device Manager overlaps new-world construction with ongoing training to mask topology-change cost. On dense and MoE models up to 235B parameters, DynaTrain reconfigures a 70B dense model in under 2s and a 235B MoE model in 4.36s, outperforming state-of-the-art checkpoint-based and elastic systems by up to three orders of magnitude while preserving correctness.

cs.LG↗

arXiv2026

WoVR: World Models as Reliable Simulators for Post-Training VLA Policies with RL

Reinforcement learning (RL) promises to unlock capabilities beyond imitation learning for Vision--Language--Action (VLA) models, but its requirement for massive real-world interaction prevents direct deployment on physical robots. Recent work attempts to use learned world models as simulators for policy optimization, yet closed-loop imagined rollouts inevitably suffer from hallucination and long-horizon error accumulation. Such errors not only degrade visual fidelity, but also mislead policy optimization by providing unreliable learning signals. We propose WoVR, a reliable world-model-based RL framework for post-training VLA policies. Instead of assuming a faithful world model, WoVR explicitly regulates how RL interacts with imperfect imagined dynamics. It improves rollout stability through a controllable action-conditioned video world model, reshapes imagined interaction to reduce effective error depth via Keyframe-Initialized Rollouts, and maintains policy--simulator alignment through World Model-Policy co-evolution. Extensive experiments demonstrate that WoVR enables stable long-horizon imagined rollouts and effective policy optimization, achieving superior LIBERO performance and consistent real-world gains across multiple robotic platforms. These results show that world models can serve as practical simulators for RL when hallucination is explicitly controlled. Additional visualization results are available at https://wovr-corl.github.io.

cs.RO↗

arXiv2026

RLinf-USER: A Unified and Extensible System for Real-World Online Policy Learning in Embodied AI

Online policy learning directly in the physical world is a promising yet challenging direction for embodied intelligence. Unlike simulation, real-world systems cannot be arbitrarily accelerated, cheaply reset, or massively replicated, suggesting that real-world policy learning is not merely an algorithmic problem, but inherently a systems problem. We present USER, a \underline{U}nified and extensible \underline{S}yst\underline{E}m for real-world online policy lea\underline{R}ning. On the systems side, USER introduces a hardware abstraction layer for unified robot management and an adaptive communication plane that enables efficient cloud-edge training. On the learning side, USER adopts a fully asynchronous training framework, designs a persistent and cache-aware replay buffer, and provides extensible abstractions for rewards, algorithms, and policies. Experiments in both simulation and the real world demonstrate that USER supports multi-robot coordination, heterogeneous manipulators, cloud-edge training with large models, and long-running asynchronous training. Together, these capabilities establish USER as a unified and extensible systems foundation for real-world online policy learning.

cs.RO↗

arXiv2026

WideSeek-R1: Exploring Width Scaling for Broad Information Seeking via Multi-Agent Reinforcement Learning

Recent advancements in Large Language Models (LLMs) have largely focused on depth scaling, where a single agent solves long-horizon problems with multi-turn reasoning and tool use. However, as tasks grow broader, the key bottleneck shifts from individual competence to organizational capability. In this work, we explore a complementary dimension of width scaling with multi-agent systems to address broad information seeking. Existing multi-agent systems often rely on hand-crafted workflows and turn-taking interactions that fail to parallelize work effectively. To bridge this gap, we propose WideSeek-R1, a lead-agent-subagent framework trained via multi-agent reinforcement learning (MARL) to synergize scalable orchestration and parallel execution. By utilizing a shared LLM with isolated contexts and specialized tools, WideSeek-R1 jointly optimizes the lead agent and parallel subagents on a curated dataset of 20k broad information-seeking tasks. Extensive experiments show that WideSeek-R1-4B achieves an item F1 score of 40.0% on the WideSearch benchmark, which is comparable to the performance of single-agent DeepSeek-R1-671B. Furthermore, WideSeek-R1-4B exhibits consistent performance gains as the number of parallel subagents increases, highlighting the effectiveness of width scaling.

cs.AI↗

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

Reducing Latency of LLM Search Agent via Speculation-based Algorithm-System Co-Design

LLM-based search agents achieve strong performance but suffer from severe latency, as each step requires serialized LLM reasoning followed by action of tool execution. We revisit this bottleneck through the lens of speculation. While traditional predict-verify speculation paradigm can break serial execution, its benefit remains limited, as it retains the full original workload and adds extra inference overhead. We observe that early agent steps often involve simple evidence-gathering, where correct actions can often be predicted without full reasoning. Building on these observations, we present SPAgent, an algorithm-system co-design framework that expands the role of speculation in search agents to reduce latency. Algorithmically, SPAgent introduces a two-phase adaptive speculation mechanism that selectively omits verification when safe. System-wise, a two-level scheduler regulates speculative requests based on engine load to ensure speculation remains beneficial. We implement SPAgent in real-world systems. Across extensive experimental settings, SPAgent achieves up to $1.65\times$ end-to-end speedup while maintaining same or even achieving higher accuracy, enabling practical deployment of multi-step search agents.

cs.AI↗

arXiv2025

$\pi_\texttt{RL}$: Online RL Fine-tuning for Flow-based Vision-Language-Action Models

Vision-Language-Action (VLA) models enable robots to understand and perform complex tasks from multimodal input. Although recent work explores using reinforcement learning (RL) to automate the laborious data collection process in scaling supervised fine-tuning (SFT), applying RL to large-scale flow-based VLAs (\eg, $\pi_0$, $\pi_{0.5}$) remains challenging due to intractable action log-likelihoods raised from flow matching. We address this challenge with $\pi_{\texttt{RL}}$, featuring two technical approaches: (1) \textbf{Flow-Noise} models the denoising process as a discrete-time MDP with a learnable noise network for exact log-likelihood computation. (2) \textbf{Flow-SDE} integrates denoising with agent-environment interaction, formulating a two-layer MDP that employs ODE-to-SDE conversion for efficient RL exploration. We evaluate $\pi_{\texttt{RL}}$ across various benchmarks, with experiments demonstrating that RL yields significant performance improvements in both in-distribution and out-of-distribution settings.

cs.LG↗

arXiv2025

RLinf-VLA: A Unified and Efficient Framework for Reinforcement Learning of Vision-Language-Action Models

Recent studies have demonstrated the potential of reinforcement learning (RL) to improve the task performance of vision-language-action (VLA) models through interaction. However, current efforts remain fragmented, lacking a unified platform for fair comparison across architectures and algorithms, as well as an efficient system design for scalable training. Therefore, we present RLinf-VLA, a unified and efficient framework for scalable RL training of VLA models. RLinf-VLA standardizes the integration of diverse VLA architectures, RL algorithms, and heterogeneous simulators through a unified interface, enabling extensibility and reproducibility. To improve efficiency, the framework adopts a flexible resource allocation architecture for rendering, inference, and training in RL pipelines. In particular, RLinf-VLA introduces a hybrid fine-grained pipeline allocation strategy that achieves a 1.61$\times$-1.88$\times$ training speedup on ManiSkill. Using this framework, RL-trained models achieve strong performance across embodied benchmarks, including 98.11% success on 130 LIBERO tasks, 97.66% success on 25 ManiSkill tasks, and 84.63% average success across 6 RoboTwin tasks. In addition, RLinf-VLA distills a set of effective practices for RL-based VLA training. We envision RLinf-VLA as a foundational framework for efficient, unified, and reproducible research in embodied intelligence.

cs.RO↗

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

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↗

arXiv2024

HETHUB: A Distributed Training System with Heterogeneous Cluster for Large-Scale Models

Training large-scale models relies on a vast number of computing resources. For example, training the GPT-4 model (1.8 trillion parameters) requires 25000 A100 GPUs . It is a challenge to build a large-scale cluster with one type of GPU-accelerator. Using multiple types of GPU-accelerators to construct a large-scale cluster is an effective way to solve the problem of insufficient homogeneous GPU-accelerators. However, the existing distributed training systems for large-scale models only support homogeneous GPU-accelerators, not support heterogeneous GPU-accelerators. To address the problem, this paper proposes a distributed training system with hybrid parallelism, HETHUB, for large-scale models, which supports heterogeneous cluster, including AMD, Nvidia GPU and other types of GPU-accelerators . It introduces a distributed unified communicator to realize the communication between heterogeneous GPU-accelerators, a distributed performance predictor, and an automatic parallel planner to develop and train models efficiently with heterogeneous GPU-accelerators. Compared to the distributed training system with homogeneous GPU-accelerators, our system can support six combinations of heterogeneous GPU-accelerators. We train the Llama-140B model on a heterogeneous cluster with 768 GPU-accelerators(128 AMD and 640 GPU-accelerator A). The experiment results show that the optimal performance of our system in the heterogeneous cluster has achieved up to 97.49% of the theoretical upper bound performance.

cs.DC↗

arXiv2024

You Only Cache Once: Decoder-Decoder Architectures for Language Models

We introduce a decoder-decoder architecture, YOCO, for large language models, which only caches key-value pairs once. It consists of two components, i.e., a cross-decoder stacked upon a self-decoder. The self-decoder efficiently encodes global key-value (KV) caches that are reused by the cross-decoder via cross-attention. The overall model behaves like a decoder-only Transformer, although YOCO only caches once. The design substantially reduces GPU memory demands, yet retains global attention capability. Additionally, the computation flow enables prefilling to early exit without changing the final output, thereby significantly speeding up the prefill stage. Experimental results demonstrate that YOCO achieves favorable performance compared to Transformer in various settings of scaling up model size and number of training tokens. We also extend YOCO to 1M context length with near-perfect needle retrieval accuracy. The profiling results show that YOCO improves inference memory, prefill latency, and throughput by orders of magnitude across context lengths and model sizes. Code is available at https://aka.ms/YOCO.

cs.CL↗

arXiv2023

Efficient Large Language Models: A Survey

Large Language Models (LLMs) have demonstrated remarkable capabilities in important tasks such as natural language understanding and language generation, and thus have the potential to make a substantial impact on our society. Such capabilities, however, come with the considerable resources they demand, highlighting the strong need to develop effective techniques for addressing their efficiency challenges. In this survey, we provide a systematic and comprehensive review of efficient LLMs research. We organize the literature in a taxonomy consisting of three main categories, covering distinct yet interconnected efficient LLMs topics from model-centric, data-centric, and framework-centric perspective, respectively. We have also created a GitHub repository where we organize the papers featured in this survey at https://github.com/AIoT-MLSys-Lab/Efficient-LLMs-Survey. We will actively maintain the repository and incorporate new research as it emerges. We hope our survey can serve as a valuable resource to help researchers and practitioners gain a systematic understanding of efficient LLMs research and inspire them to contribute to this important and exciting field.

cs.CL↗

arXiv2023

AutoTaskFormer: Searching Vision Transformers for Multi-task Learning

Vision Transformers have shown great performance in single tasks such as classification and segmentation. However, real-world problems are not isolated, which calls for vision transformers that can perform multiple tasks concurrently. Existing multi-task vision transformers are handcrafted and heavily rely on human expertise. In this work, we propose a novel one-shot neural architecture search framework, dubbed AutoTaskFormer (Automated Multi-Task Vision TransFormer), to automate this process. AutoTaskFormer not only identifies the weights to share across multiple tasks automatically, but also provides thousands of well-trained vision transformers with a wide range of parameters (e.g., number of heads and network depth) for deployment under various resource constraints. Experiments on both small-scale (2-task Cityscapes and 3-task NYUv2) and large-scale (16-task Taskonomy) datasets show that AutoTaskFormer outperforms state-of-the-art handcrafted vision transformers in multi-task learning. The entire code and models will be open-sourced.

cs.CV↗

arXiv2023

ElasticViT: Conflict-aware Supernet Training for Deploying Fast Vision Transformer on Diverse Mobile Devices

Neural Architecture Search (NAS) has shown promising performance in the automatic design of vision transformers (ViT) exceeding 1G FLOPs. However, designing lightweight and low-latency ViT models for diverse mobile devices remains a big challenge. In this work, we propose ElasticViT, a two-stage NAS approach that trains a high-quality ViT supernet over a very large search space that supports a wide range of mobile devices, and then searches an optimal sub-network (subnet) for direct deployment. However, prior supernet training methods that rely on uniform sampling suffer from the gradient conflict issue: the sampled subnets can have vastly different model sizes (e.g., 50M vs. 2G FLOPs), leading to different optimization directions and inferior performance. To address this challenge, we propose two novel sampling techniques: complexity-aware sampling and performance-aware sampling. Complexity-aware sampling limits the FLOPs difference among the subnets sampled across adjacent training steps, while covering different-sized subnets in the search space. Performance-aware sampling further selects subnets that have good accuracy, which can reduce gradient conflicts and improve supernet quality. Our discovered models, ElasticViT models, achieve top-1 accuracy from 67.2% to 80.0% on ImageNet from 60M to 800M FLOPs without extra retraining, outperforming all prior CNNs and ViTs in terms of accuracy and latency. Our tiny and small models are also the first ViT models that surpass state-of-the-art CNNs with significantly lower latency on mobile devices. For instance, ElasticViT-S1 runs 2.62x faster than EfficientNet-B0 with 0.1% higher accuracy.

cs.CV↗

arXiv2023

SpaceEvo: Hardware-Friendly Search Space Design for Efficient INT8 Inference

The combination of Neural Architecture Search (NAS) and quantization has proven successful in automatically designing low-FLOPs INT8 quantized neural networks (QNN). However, directly applying NAS to design accurate QNN models that achieve low latency on real-world devices leads to inferior performance. In this work, we find that the poor INT8 latency is due to the quantization-unfriendly issue: the operator and configuration (e.g., channel width) choices in prior art search spaces lead to diverse quantization efficiency and can slow down the INT8 inference speed. To address this challenge, we propose SpaceEvo, an automatic method for designing a dedicated, quantization-friendly search space for each target hardware. The key idea of SpaceEvo is to automatically search hardware-preferred operators and configurations to construct the search space, guided by a metric called Q-T score to quantify how quantization-friendly a candidate search space is. We further train a quantized-for-all supernet over our discovered search space, enabling the searched models to be directly deployed without extra retraining or quantization. Our discovered models establish new SOTA INT8 quantized accuracy under various latency constraints, achieving up to 10.1% accuracy improvement on ImageNet than prior art CNNs under the same latency. Extensive experiments on diverse edge devices demonstrate that SpaceEvo consistently outperforms existing manually-designed search spaces with up to 2.5x faster speed while achieving the same accuracy.

cs.CV↗