arXiv subjects

Zechen Hu

Publications and source records attributed to Zechen Hu.

12 recordsLinked to original sources

arXiv2026

Mergeable Model-Side Aggregation States for Long-Context Language Models

A known limitation of long-context language models is their increasingly unreliable performance in non-additive, set-based aggregation as context length grows. Examples include cardinality estimation, set relationships, and grouped statistics, which widely exist in logs, program outputs, tables, and multi-turn conversations. To provide the aggregation state required by these tasks, we introduce a model-side aggregation interface that maintains compact Hash-based HyperLogLog (HLL) sketch states alongside a frozen language model. While the model processes the context, an extractor maps each relevant record to a canonical identity. The identity is then hashed and updates the HLL state. These states can be merged across context segments and/or read out directly for downstream reasoning, avoiding an additional generate-execute-return cycle. We validate the proposed approach by setting the HLL state size as 2 KiB (2,048 registers), which does not increase with context length or set cardinality. In a distinct-count experiment involving one million records, the mean relative error was 1.6%. In a separate merge test, states built from as many as 256 segments produced exactly the same readout as a single pass over the same stream. On 3,969 aggregate-then-reason tasks from 174 source windows, the fixed-budget interface reached 99.2% accuracy on Gemma 4 (31B, BF16), compared with 100.0% under exact aggregation; the paired gap was 0.8 percentage points (95% window-cluster CI: 0.5-1.3 points). On a matched set of 174 items, our method improved over direct full-context reasoning by 63.2 points on Qwen and 56.3 points on Gemma. The corresponding gains over chain-of-thought (CoT) reasoning were 60.9 and 63.2 points, respectively. On a fixed 1,200-task Oolong-Synth subset, our method reached 91.1% on Qwen and 99.3% on Gemma. Code is available at https://github.com/songdc98/sketchops.

cs.CL↗

arXiv2026

Smooth Scaling Laws Hide Stepwise Token Learning

Language model loss follows remarkably regular scaling laws over model and data size, yet it remains unclear why the aggregate loss should exhibit a power-law form. Existing explanations often attribute this regularity to a heavy-tailed spectrum of pattern difficulty in natural language, but this view has not been directly validated at token-level granularity in large-scale real-data training. We present a token-level framework that decomposes scaling laws into localized learning events of individual contextualized tokens. By fitting token loss trajectories with sigmoids, we show that token learning is concentrated in localized transitions, giving rise to a learning-time spectrum that dominates the scaling-law shape. Across more than one hundred pre-training runs on large and diverse real-language corpora with modern LLM architectures, scaling up to 6B parameters and 300B training tokens, the measured learning-time spectrum quantitatively reconstructs the validation loss derivative along the training-step $T$, data-scale $D$, and model-scale $M$ axes. We further show that the same signal is actionable: by reshaping the training distribution according to when tokens become learnable, we alter the optimization trajectory and achieve 11\% faster validation-loss reduction. These results provide direct empirical evidence that scaling laws are governed primarily by the distribution of token-level learning times, and that this distribution can be used not only to explain scaling behavior but also to improve training performance.

cs.CL↗

arXiv2025

Efficient Multi-turn RL for GUI Agents via Decoupled Training and Adaptive Data Curation

Vision-language model (VLM) based GUI agents show promise for automating complex desktop and mobile tasks, but face significant challenges in applying reinforcement learning (RL): (1) slow multi-turn interactions with GUI environments for policy rollout, and (2) insufficient high-quality agent-environment interactions for policy learning. To address these challenges, we propose DART, a Decoupled Agentic RL Training framework for GUI agents, which coordinates heterogeneous modules in a highly decoupled manner. DART separates the training system into four asynchronous modules: environment cluster, rollout service, data manager, and trainer. This design enables non-blocking communication, asynchronous training, rollout-wise trajectory sampling, and per-worker model synchronization, significantly improving the system efficiency: 1.6*GPU utilization for rollout, 1.9* training throughput, and 5.5* environment utilization. To facilitate effective learning from abundant samples, we introduce an adaptive data curation scheme: (1) pre-collecting successful trajectories for challenging tasks to supplement sparse success in online sampling; (2) dynamically adjusting rollout numbers and trajectory lengths based on task difficulty; (3) training selectively on high-entropy steps to prioritize critical decisions; (4) stabilizing learning via truncated importance sampling for policy mismatch between policy rollout and updating. On the OSWorld benchmark, DART-GUI-7B achieves a 42.13% task success rate, a 14.61% absolute gain over the base model, and 7.34% higher than open-source SOTA. We will fully open-source our training framework, data, and model checkpoints via computer-use-agents.github.io/dart-gui, which we believe is a timely contribution to the open-source community of agentic RL training.

cs.LG↗

arXiv2025

Robust Online Calibration for UWB-Aided Visual-Inertial Navigation with Bias Correction

This paper presents a novel robust online calibration framework for Ultra-Wideband (UWB) anchors in UWB-aided Visual-Inertial Navigation Systems (VINS). Accurate anchor positioning, a process known as calibration, is crucial for integrating UWB ranging measurements into state estimation. While several prior works have demonstrated satisfactory results by using robot-aided systems to autonomously calibrate UWB systems, there are still some limitations: 1) these approaches assume accurate robot localization during the initialization step, ignoring localization errors that can compromise calibration robustness, and 2) the calibration results are highly sensitive to the initial guess of the UWB anchors' positions, reducing the practical applicability of these methods in real-world scenarios. Our approach addresses these challenges by explicitly incorporating the impact of robot localization uncertainties into the calibration process, ensuring robust initialization. To further enhance the robustness of the calibration results against initialization errors, we propose a tightly-coupled Schmidt Kalman Filter (SKF)-based online refinement method, making the system suitable for practical applications. Simulations and real-world experiments validate the improved accuracy and robustness of our approach.

cs.RO↗

arXiv2025

CARoL: Context-aware Adaptation for Robot Learning

Using Reinforcement Learning (RL) to learn new robotic tasks from scratch is often inefficient. Leveraging prior knowledge has the potential to significantly enhance learning efficiency, which, however, raises two critical challenges: how to determine the relevancy of existing knowledge and how to adaptively integrate them into learning a new task. In this paper, we propose Context-aware Adaptation for Robot Learning (CARoL), a novel framework to efficiently learn a similar but distinct new task from prior knowledge. CARoL incorporates context awareness by analyzing state transitions in system dynamics to identify similarities between the new task and prior knowledge. It then utilizes these identified similarities to prioritize and adapt specific knowledge pieces for the new task. Additionally, CARoL has a broad applicability spanning policy-based, value-based, and actor-critic RL algorithms. We validate the efficiency and generalizability of CARoL on both simulated robotic platforms and physical ground vehicles. The simulations include CarRacing and LunarLander environments, where CARoL demonstrates faster convergence and higher rewards when learning policies for new tasks. In real-world experiments, we show that CARoL enables a ground vehicle to quickly and efficiently adapt policies learned in simulation to smoothly traverse real-world off-road terrain.

cs.RO↗

arXiv2025

Mocap-2-to-3: Multi-view Lifting for Monocular Motion Recovery with 2D Pretraining

Human motion recovery for real-world interaction demands both precise action details and metric-scale trajectories. Recovering absolute human pose from monocular input presents a viable solution, but faces two main challenges: (1) models' reliance on 3D training data from constrained environments limits their out-of-distribution generalization; and (2) the inherent difficulty of estimating metric-scale poses from monocular observations. This paper introduces Mocap-2-to-3, a novel framework that differs from prior HMR methods by recovering absolute poses from monocular input and leveraging abundant 2D data to enhance 3D motion recovery. To effectively utilize the action priors and diversity in large-scale 2D datasets, we reformulate 3D motion as a multi-view synthesis process and divide the training into two stages: a single-view diffusion model is first pre-trained on extensive 2D data, followed by multi-view fine-tuning on 3D data, thus achieving a combination of strong priors and geometric constraints. Furthermore, to recover absolute poses, we introduce a novel human motion representation that decouples the learning of local pose and global movements, while encoding ground geometric priors to accelerate convergence, thereby yielding more precise positioning in the physical world. Experiments on in-the-wild benchmarks show that our method outperforms state-of-the-art approaches in both camera-space motion realism and world-grounded human positioning, while exhibiting strong generalization capability.

cs.CV↗

arXiv2024

Motion-2-To-3: Leveraging 2D Motion Data for 3D Motion Generations

Text-driven human motion synthesis has showcased its potential for revolutionizing motion design in the movie and game industry. Existing methods often rely on 3D motion capture data, which requires special setups, resulting in high costs for data acquisition, ultimately limiting the diversity and scope of human motion. In contrast, 2D human videos offer a vast and accessible source of motion data, covering a wider range of styles and activities. In this paper, we explore the use of 2D human motion extracted from videos as an alternative data source to improve text-driven 3D motion generation. Our approach introduces a novel framework that disentangles local joint motion from global movements, enabling efficient learning of local motion priors from 2D data. We first train a single-view 2D local motion generator on a large dataset of text-2D motion pairs. Then we fine-tune the generator with 3D data, transforming it into a multi-view generator that predicts view-consistent local joint motion and root dynamics. Evaluations on the well-acknowledged dataset and novel text prompts demonstrate that our method can efficiently utilize 2D data, supporting a wider range of realistic 3D human motion generation. Our code is publicly available at https://zju3dv.github.io/Motion-2-to-3/.

cs.CV↗

arXiv2024

World-Grounded Human Motion Recovery via Gravity-View Coordinates

We present a novel method for recovering world-grounded human motion from monocular video. The main challenge lies in the ambiguity of defining the world coordinate system, which varies between sequences. Previous approaches attempt to alleviate this issue by predicting relative motion in an autoregressive manner, but are prone to accumulating errors. Instead, we propose estimating human poses in a novel Gravity-View (GV) coordinate system, which is defined by the world gravity and the camera view direction. The proposed GV system is naturally gravity-aligned and uniquely defined for each video frame, largely reducing the ambiguity of learning image-pose mapping. The estimated poses can be transformed back to the world coordinate system using camera rotations, forming a global motion sequence. Additionally, the per-frame estimation avoids error accumulation in the autoregressive methods. Experiments on in-the-wild benchmarks demonstrate that our method recovers more realistic motion in both the camera space and world-grounded settings, outperforming state-of-the-art methods in both accuracy and speed. The code is available at https://zju3dv.github.io/gvhmr/.

cs.CV↗

arXiv2024

Learning Coordinated Maneuver in Adversarial Environments

This paper aims to solve the coordination of a team of robots traversing a route in the presence of adversaries with random positions. Our goal is to minimize the overall cost of the team, which is determined by (i) the accumulated risk when robots stay in adversary-impacted zones and (ii) the mission completion time. During traversal, robots can reduce their speed and act as a `guard' (the slower, the better), which will decrease the risks certain adversary incurs. This leads to a trade-off between the robots' guarding behaviors and their travel speeds. The formulated problem is highly non-convex and cannot be efficiently solved by existing algorithms. Our approach includes a theoretical analysis of the robots' behaviors for the single-adversary case. As the scale of the problem expands, solving the optimal solution using optimization approaches is challenging, therefore, we employ reinforcement learning techniques by developing new encoding and policy-generating methods. Simulations demonstrate that our learning methods can efficiently produce team coordination behaviors. We discuss the reasoning behind these behaviors and explain why they reduce the overall team cost.

cs.RO↗

arXiv2024

Bi-CL: A Reinforcement Learning Framework for Robots Coordination Through Bi-level Optimization

In multi-robot systems, achieving coordinated missions remains a significant challenge due to the coupled nature of coordination behaviors and the lack of global information for individual robots. To mitigate these challenges, this paper introduces a novel approach, Bi-level Coordination Learning (Bi-CL), that leverages a bi-level optimization structure within a centralized training and decentralized execution paradigm. Our bi-level reformulation decomposes the original problem into a reinforcement learning level with reduced action space, and an imitation learning level that gains demonstrations from a global optimizer. Both levels contribute to improved learning efficiency and scalability. We note that robots' incomplete information leads to mismatches between the two levels of learning models. To address this, Bi-CL further integrates an alignment penalty mechanism, aiming to minimize the discrepancy between the two levels without degrading their training efficiency. We introduce a running example to conceptualize the problem formulation and apply Bi-CL to two variations of this example: route-based and graph-based scenarios. Simulation results demonstrate that Bi-CL can learn more efficiently and achieve comparable performance with traditional multi-agent reinforcement learning baselines for multi-robot coordination.

cs.RO↗

arXiv2024

Scaling Team Coordination on Graphs with Reinforcement Learning

This paper studies Reinforcement Learning (RL) techniques to enable team coordination behaviors in graph environments with support actions among teammates to reduce the costs of traversing certain risky edges in a centralized manner. While classical approaches can solve this non-standard multi-agent path planning problem by converting the original Environment Graph (EG) into a Joint State Graph (JSG) to implicitly incorporate the support actions, those methods do not scale well to large graphs and teams. To address this curse of dimensionality, we propose to use RL to enable agents to learn such graph traversal and teammate supporting behaviors in a data-driven manner. Specifically, through a new formulation of the team coordination on graphs with risky edges problem into Markov Decision Processes (MDPs) with a novel state and action space, we investigate how RL can solve it in two paradigms: First, we use RL for a team of agents to learn how to coordinate and reach the goal with minimal cost on a single EG. We show that RL efficiently solves problems with up to 20/4 or 25/3 nodes/agents, using a fraction of the time needed for JSG to solve such complex problems; Second, we learn a general RL policy for any $N$-node EGs to produce efficient supporting behaviors. We present extensive experiments and compare our RL approaches against their classical counterparts.

cs.RO↗

arXiv2023

Team Coordination on Graphs with State-Dependent Edge Cost

This paper studies a team coordination problem in a graph environment. Specifically, we incorporate "support" action which an agent can take to reduce the cost for its teammate to traverse some edges that have higher costs otherwise. Due to this added feature, the graph traversal is no longer a standard multi-agent path planning problem. To solve this new problem, we propose a novel formulation by posing it as a planning problem in the joint state space: the joint state graph (JSG). Since the edges of JSG implicitly incorporate the support actions taken by the agents, we are able to now optimize the joint actions by solving a standard single-agent path planning problem in JSG. One main drawback of this approach is the curse of dimensionality in both the number of agents and the size of the graph. To improve scalability in graph size, we further propose a hierarchical decomposition method to perform path planning in two levels. We provide complexity analysis as well as a statistical analysis to demonstrate the efficiency of our algorithm.

cs.MA↗