Support Local Variables
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
Selected work
Representative Papers
SimGym: A Framework for A/B Test Simulation in E-Commerce with Traffic-Grounded VLM Agents
Traditional A/B testing in e-commerce is time-consuming, requires diverting real user traffic, and often degrades user experience. This work proposes an intelligent agent simulation framework grounded in vision-language models (VLMs), which constructs traffic-driven user profiles from production clickstream data and simulates end-to-end shopping journeys across control and treatment groups within live-browser environments. By integrating a situational memory mechanism and a behavior alignment evaluation protocol, the framework achieves high-fidelity user emulation. It uniquely combines multimodal perception, browser-level interaction, and behavioral modeling, and has been validated across multiple product categories on major e-commerce platforms. The simulated outcomes align with real user add-to-cart behavior in 77% of cases, reducing experimental cycles from weeks to under one hour.
0 citationsRead paper
SimPersona: Learning Discrete Buyer Personas from Raw Clickstreams for Grounded E-Commerce Agents
Existing e-commerce agents struggle to capture the heterogeneity of real-world buyers, often collapsing into a single “average buyer” strategy and relying on manually crafted persona prompts that are brittle and inefficient. This work proposes a method to automatically learn interpretable, discrete buyer types directly from raw clickstream data, encoding them via a behavior-aware vector-quantized variational autoencoder (VQ-VAE) into compact persona tokens residing in the vocabulary of a large language model. These tokens are then integrated with real browsing trajectories to fine-tune agent behavior. The approach enables personalized persona assignment without retraining and faithfully reconstructs merchant-specific buyer distributions. Evaluated on data from 42 live stores and 8.37 million buyers, the method achieves 78% alignment with real buyer conversion rates and outperforms a baseline model with eight times more parameters in goal-oriented tasks. The full pipeline—from clickstream processing to agent training—is publicly released.
0 citationsRead paper
The Wristband Gaussian Loss: Deterministic, Composable Latents via a Sphere-Interval Decomposition
This work proposes a deterministic Gaussianization method that eschews sampling, KL divergence, and iterative transport. By decomposing input vectors into direction and radius—where the radius is transformed via the cumulative distribution function—the approach maps data to a spherical–interval product space (a “wristband”) and employs an energy minimization mechanism based on the Neumann reflection kernel to achieve Gaussianization. Key contributions include the first deterministic Gaussianization loss, spectral regularization combining spherical harmonics with cosine Mercer modes, and formally verified theoretical guarantees in Lean 4. The method integrates a 1D Wasserstein radial term, moment penalties, invertible flows, and learnable key attention, while constructing Gaussian reference batches via Hungarian recursive averaging. It achieves state-of-the-art Gaussianization performance on complex, non-independent radial–angular joint distributions in both 10D and 128D settings and supports counterfactual generation with either independent or dependent factors.
0 citationsRead paper
SimGym: Traffic-Grounded Browser Agents for Offline A/B Testing in E-Commerce
Traditional A/B testing in e-commerce relies on real user traffic, resulting in prolonged experiment cycles and potential degradation of user experience. This work proposes SimGym, a large language model (LLM)-driven browser agent framework that constructs high-fidelity synthetic buyers by extracting user personas and intents from production data, enabling offline simulation of their interactions under both control and treatment conditions. SimGym represents the first approach to integrate LLM-powered agents with real-world user behavioral patterns, effectively replicating the impact of UI changes without involving actual users. Validated on a major e-commerce platform, SimGym reduces experiment duration from weeks to under an hour—even without alignment fine-tuning—demonstrating substantial gains in testing efficiency and scalability.
0 citationsRead paper
Recent publications
Latest Papers
Support Local Variables
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
SimGym: A Framework for A/B Test Simulation in E-Commerce with Traffic-Grounded VLM Agents
Traditional A/B testing in e-commerce is time-consuming, requires diverting real user traffic, and often degrades user experience. This work proposes an intelligent agent simulation framework grounded in vision-language models (VLMs), which constructs traffic-driven user profiles from production clickstream data and simulates end-to-end shopping journeys across control and treatment groups within live-browser environments. By integrating a situational memory mechanism and a behavior alignment evaluation protocol, the framework achieves high-fidelity user emulation. It uniquely combines multimodal perception, browser-level interaction, and behavioral modeling, and has been validated across multiple product categories on major e-commerce platforms. The simulated outcomes align with real user add-to-cart behavior in 77% of cases, reducing experimental cycles from weeks to under one hour.
0 citationsRead paper
SimPersona: Learning Discrete Buyer Personas from Raw Clickstreams for Grounded E-Commerce Agents
Existing e-commerce agents struggle to capture the heterogeneity of real-world buyers, often collapsing into a single “average buyer” strategy and relying on manually crafted persona prompts that are brittle and inefficient. This work proposes a method to automatically learn interpretable, discrete buyer types directly from raw clickstream data, encoding them via a behavior-aware vector-quantized variational autoencoder (VQ-VAE) into compact persona tokens residing in the vocabulary of a large language model. These tokens are then integrated with real browsing trajectories to fine-tune agent behavior. The approach enables personalized persona assignment without retraining and faithfully reconstructs merchant-specific buyer distributions. Evaluated on data from 42 live stores and 8.37 million buyers, the method achieves 78% alignment with real buyer conversion rates and outperforms a baseline model with eight times more parameters in goal-oriented tasks. The full pipeline—from clickstream processing to agent training—is publicly released.
0 citationsRead paper
The Wristband Gaussian Loss: Deterministic, Composable Latents via a Sphere-Interval Decomposition
This work proposes a deterministic Gaussianization method that eschews sampling, KL divergence, and iterative transport. By decomposing input vectors into direction and radius—where the radius is transformed via the cumulative distribution function—the approach maps data to a spherical–interval product space (a “wristband”) and employs an energy minimization mechanism based on the Neumann reflection kernel to achieve Gaussianization. Key contributions include the first deterministic Gaussianization loss, spectral regularization combining spherical harmonics with cosine Mercer modes, and formally verified theoretical guarantees in Lean 4. The method integrates a 1D Wasserstein radial term, moment penalties, invertible flows, and learnable key attention, while constructing Gaussian reference batches via Hungarian recursive averaging. It achieves state-of-the-art Gaussianization performance on complex, non-independent radial–angular joint distributions in both 10D and 128D settings and supports counterfactual generation with either independent or dependent factors.
0 citationsRead paper
SimGym: Traffic-Grounded Browser Agents for Offline A/B Testing in E-Commerce
Traditional A/B testing in e-commerce relies on real user traffic, resulting in prolonged experiment cycles and potential degradation of user experience. This work proposes SimGym, a large language model (LLM)-driven browser agent framework that constructs high-fidelity synthetic buyers by extracting user personas and intents from production data, enabling offline simulation of their interactions under both control and treatment conditions. SimGym represents the first approach to integrate LLM-powered agents with real-world user behavioral patterns, effectively replicating the impact of UI changes without involving actual users. Validated on a major e-commerce platform, SimGym reduces experiment duration from weeks to under an hour—even without alignment fine-tuning—demonstrating substantial gains in testing efficiency and scalability.
0 citationsRead paper