Baidu · Agora

Selected work

Representative Papers

Beyond Monolithic Architectures: A Multi-Agent Search and Knowledge Optimization Framework for Agentic Search

Jan 08, 2026arXiv.org

This work addresses the limitations of existing agent-based search systems, which suffer from prolonged reasoning chains, sparse rewards, and credit assignment difficulties due to their monolithic architecture, ultimately undermining learning stability. To overcome these challenges, the authors propose M-ASK, a novel framework that decouples the search task into two specialized agents: a search behavior agent responsible for action execution and a knowledge management agent tasked with maintaining a compressed contextual representation. These agents are jointly optimized through turn-level fine-grained rewards. By integrating a multi-agent architecture, large language model tool invocation, and context compression techniques, M-ASK significantly outperforms strong baselines on multi-hop question answering benchmarks, achieving both higher answer accuracy and markedly improved training stability.

3 citations1 influentialRead paper

OmegaUse: Building a General-Purpose GUI Agent for Autonomous Task Execution

Jan 28, 2026

This work addresses the challenge of developing general-purpose GUI agents capable of autonomously performing real-world tasks across mobile and desktop platforms, a goal hindered by the scarcity of high-quality interaction data and the lack of effective training methodologies. To overcome these limitations, the authors propose an automated synthetic data generation framework that integrates bottom-up exploration with top-down task-driven generation to produce high-fidelity, cross-platform interaction trajectories. They further introduce a decoupled two-stage training paradigm—comprising supervised fine-tuning (SFT) followed by Generalized Reinforcement Learning with Policy Optimization (GRPO)—leveraging a Mixture-of-Experts backbone and cross-platform GUI understanding techniques. The resulting model achieves state-of-the-art performance on multiple benchmarks, including ScreenSpot-V2 (96.3%), AndroidControl (79.1%), and the newly introduced OS-Nav suite (ChiM-Nav 74.24%, Ubu-Nav 55.9%), demonstrating the first unified architecture for high-performance cross-device autonomous operation.

2 citationsRead paper

Distributional Clarity: The Hidden Driver of RL-Friendliness in Large Language Models

Jan 11, 2026arXiv.org

This work investigates the significant performance disparities observed among large language models under identical reinforcement learning (RL) training, noting that some models struggle to benefit from such optimization. The study introduces, for the first time, “distributional sharpness” as a key structural property governing RL compatibility and quantifies it using the Silhouette Coefficient. Building on this insight, the authors propose a Silhouette-Aware Reweighting strategy that adaptively reweights low-sharpness samples during training to improve learning efficiency. Extensive experiments across six mathematical reasoning benchmarks demonstrate consistent performance gains, with improvements of up to 5.9 points on AIME24, thereby validating both the trainability and broad applicability of distributional sharpness as a guiding principle for RL-based model refinement.

1 citationsRead paper

VideoAR: Autoregressive Video Generation via Next-Frame & Scale Prediction

Jan 09, 2026arXiv.org

Existing video generation methods suffer from significant limitations in computational efficiency and long-term temporal consistency. This work proposes VideoAR, the first large-scale visual autoregressive framework for video generation, which decouples spatial and temporal dependencies by integrating multi-scale next-frame prediction with autoregressive modeling. Key innovations include a 3D multi-scale tokenizer, multi-scale temporal RoPE positional encoding, a cross-frame error correction mechanism, and a stochastic frame masking strategy, complemented by a three-stage progressive pretraining scheme to enhance spatiotemporal modeling. Experiments demonstrate that VideoAR reduces the Fréchet Video Distance (FVD) on UCF-101 from 99.5 to 88.6, decreases inference steps by over an order of magnitude, and achieves a VBench score of 81.74—matching the performance of diffusion models an order of magnitude larger in scale.

1 citationsRead paper

VoxelSplat: Dynamic Gaussian Splatting as an Effective Loss for Occupancy and Flow Prediction

Jun 05, 2025

To address performance bottlenecks in monocular 3D semantic occupancy and scene flow joint prediction—arising from occlusion and imbalanced dynamic object distribution—this paper proposes the first dynamic Gaussian rasterization-based regularization framework. It innovatively introduces 3D Gaussian splatting into occupancy training: leveraging its differentiable 2D projection for pixel-level semantic supervision, and constructing self-supervised scene flow targets via inter-frame Gaussian parameter changes—eliminating the need for explicit flow annotations. The method comprises sparse semantic Gaussian decoding and dynamic rasterization regularization, significantly improving joint prediction accuracy for both semantic occupancy and scene flow, achieving state-of-the-art results on benchmarks including nuScenes. Moreover, it is fully compatible with mainstream occupancy models and incurs zero inference overhead.

1 citationsRead paper

Recent publications