Selected work

Representative Papers

AMBER: Training Long-Horizon Web Agents through Append-Only Memory

Oct 05, 2026

This study addresses the challenge of critical information loss in long-horizon web agents due to context overflow, where existing overwrite-based memory mechanisms struggle to retain facts and feedback under sparse rewards. We propose AMBER, a framework introducing a novel append-only free-form memory mechanism that enables agents to jointly learn reasoning, acting, and memory writing. This architecture ensures the permanent preservation of key evidence and facilitates end-to-end reinforcement training without costly supervised data. Experiments demonstrate that AMBER improves success rates by 4.09% over baselines on WebArena Lite and increases multi-turn task completion by 4.8%, outperforming methods reliant on extensive annotated data at significantly lower cost.

0 citationsRead paper

TreeWalker: Partial Evaluation for Grouped Tree-Ensemble Inference

Oct 02, 2026

This study addresses the redundant computation caused by shared features during grouped inference in tree ensemble models by proposing an optimization framework based on partial evaluation. The method decouples feature static and dynamic components, leveraging bitmask partitioning to enable single-pass inference. Furthermore, it introduces a theory of structured work decomposition, proving that the per-row computational cost asymptotically approaches the theoretical lower bound as group size increases. The proposed system maintains full compatibility with standard LightGBM and XGBoost models. Experimental results demonstrate speedups ranging from 2.5× to 7.8× on Intel platforms, with even more pronounced gains on Arm architectures, while preserving floating-point precision comparable to native implementations.

0 citationsRead paper

SimTrace: Grounded Multimodal User Trajectories Generation for Online User Modeling

Sep 29, 2026

This study addresses the challenge of acquiring fine-grained online user trajectories, which is hindered by privacy constraints and data scarcity, while existing datasets suffer from over-abstraction or insufficient scale. To overcome these limitations, this work proposes SimTrace, a framework that anonymizes real interactions and constructs digital twins of web environments, leveraging computer-use agents to generate high-fidelity, multimodal synthetic clickstream trajectories that balance privacy preservation with semantic authenticity. Experimental results demonstrate that SimTrace outperforms baselines on seven of eight fidelity metrics, achieving downstream task performance comparable to models trained on real data. Furthermore, data augmentation yields an 11.0% relative improvement in next-action prediction accuracy, providing shareable data support for virtual customer development.

0 citationsRead paper

T-RoPE: Time-Aware Rotary Position Embedding for Sequential Recommendation

Sep 24, 2026

This study addresses the limitation of Rotary Position Embedding (RoPE) in recommender systems, which encodes only sequential order while neglecting temporal periodicity and calendar phases. To overcome this, we propose a time-aware rotary position encoding method. Specifically, it employs a multi-scale frequency bank and non-stationary keys to break temporal translation invariance, and introduces learnable temporal coefficients for shifted query alignment, enabling non-stationary temporal modeling while preserving interface compatibility. Furthermore, timestamp-based angular rotation and a linear-complexity algorithm are designed. Experiments demonstrate that the proposed approach significantly improves performance across multiple benchmark and industrial datasets, achieving a 0.33% increase in conversion rate in online A/B testing.

0 citationsRead paper

Support Local Variables

Sep 01, 2026

为实现更高级的优化并鼓励外部贡献，提出了一种新的基于方法的JIT编译器ZJIT，通过将局部变量提升为SSA值来优化Ruby的局部变量。

0 citationsRead paper

Recent publications

Latest Papers

AMBER: Training Long-Horizon Web Agents through Append-Only Memory

Oct 05, 2026

This study addresses the challenge of critical information loss in long-horizon web agents due to context overflow, where existing overwrite-based memory mechanisms struggle to retain facts and feedback under sparse rewards. We propose AMBER, a framework introducing a novel append-only free-form memory mechanism that enables agents to jointly learn reasoning, acting, and memory writing. This architecture ensures the permanent preservation of key evidence and facilitates end-to-end reinforcement training without costly supervised data. Experiments demonstrate that AMBER improves success rates by 4.09% over baselines on WebArena Lite and increases multi-turn task completion by 4.8%, outperforming methods reliant on extensive annotated data at significantly lower cost.

0 citationsRead paper

TreeWalker: Partial Evaluation for Grouped Tree-Ensemble Inference

Oct 02, 2026

This study addresses the redundant computation caused by shared features during grouped inference in tree ensemble models by proposing an optimization framework based on partial evaluation. The method decouples feature static and dynamic components, leveraging bitmask partitioning to enable single-pass inference. Furthermore, it introduces a theory of structured work decomposition, proving that the per-row computational cost asymptotically approaches the theoretical lower bound as group size increases. The proposed system maintains full compatibility with standard LightGBM and XGBoost models. Experimental results demonstrate speedups ranging from 2.5× to 7.8× on Intel platforms, with even more pronounced gains on Arm architectures, while preserving floating-point precision comparable to native implementations.

0 citationsRead paper

SimTrace: Grounded Multimodal User Trajectories Generation for Online User Modeling

Sep 29, 2026

This study addresses the challenge of acquiring fine-grained online user trajectories, which is hindered by privacy constraints and data scarcity, while existing datasets suffer from over-abstraction or insufficient scale. To overcome these limitations, this work proposes SimTrace, a framework that anonymizes real interactions and constructs digital twins of web environments, leveraging computer-use agents to generate high-fidelity, multimodal synthetic clickstream trajectories that balance privacy preservation with semantic authenticity. Experimental results demonstrate that SimTrace outperforms baselines on seven of eight fidelity metrics, achieving downstream task performance comparable to models trained on real data. Furthermore, data augmentation yields an 11.0% relative improvement in next-action prediction accuracy, providing shareable data support for virtual customer development.

0 citationsRead paper

T-RoPE: Time-Aware Rotary Position Embedding for Sequential Recommendation

Sep 24, 2026

This study addresses the limitation of Rotary Position Embedding (RoPE) in recommender systems, which encodes only sequential order while neglecting temporal periodicity and calendar phases. To overcome this, we propose a time-aware rotary position encoding method. Specifically, it employs a multi-scale frequency bank and non-stationary keys to break temporal translation invariance, and introduces learnable temporal coefficients for shifted query alignment, enabling non-stationary temporal modeling while preserving interface compatibility. Furthermore, timestamp-based angular rotation and a linear-complexity algorithm are designed. Experiments demonstrate that the proposed approach significantly improves performance across multiple benchmark and industrial datasets, achieving a 0.33% increase in conversion rate in online A/B testing.

0 citationsRead paper

Support Local Variables

Sep 01, 2026

为实现更高级的优化并鼓励外部贡献，提出了一种新的基于方法的JIT编译器ZJIT，通过将局部变量提升为SSA值来优化Ruby的局部变量。

0 citationsRead paper