Skip to content
Log in

Author

Yue Wang

View on Semantic Scholar

2 papers indexed here

We haven’t gathered this author’s papers yet. Follow them and we’ll fetch their work.

Not the right person? Other researchers publish under this name.

Book Open access Jul 2026

Precision at Scale: An End-to-End Graph-based Framework for Mitigating Network Interference in TikTok A/B Tests

Large-scale A/B tests on social platforms suffer from inherent network interference, violating the Stable Unit Treatment Value Assumption (SUTVA) and distorting measured results. Industrial practices for mitigating network interference face a foundational trade-off. For scalability, they often rely on clustering static graphs, which serve as imperfect proxies for true interference pathways. Conversely, theoretically sound methods remain computationally intractable at production scale. This paper presents a production-ready framework deployed at TikTok, which integrates three core contributions to address these challenges: 1) Learned Interference Graph (LIG): Estimates interference probabilities using dynamic interaction patterns for more context-aware modeling. 2) Scalable Community Partitioning (SCP): A Spark-optimized ParLeiden implementation that performs billion-node graph clustering daily and generalizes effectively across diverse interaction types, achieving a purity score of 0.898 for group chat interactions. 3) Sensitivity-Enhanced Estimation (SEE): A multivariate system leveraging Controlled-experiment Using Pre-Experiment Data (CUPED) to mitigate variance inflation from cluster-based randomization. In live production tests, our framework reduces interference rates by 68.8%, correcting a biased treatment effect estimate from +1.44% to a statistically significant +2.08%. It also enables previously undetectable cross-ecosystem measurements, revealing a +0.2% lift in creator upload volumes driven by user-side treatments.

Yu-Han Li, Jian-Yu Ni, Ao Li et al. · 0 citations
Preprint Jul 2026

PATS: Policy-Aware Training Scaffolding for Agentic Reinforcement Learning

In long-horizon LLM agent reinforcement learning, weak policies often repeat similar failures, producing uninformative rollout trajectories and limiting effective policy optimization. Existing skill-centric methods improve exploration by optimizing, filtering, or internalizing reusable skills. However, they remain centered on the skills themselves rather than being designed as adaptive training-time support for the evolving policy. To address this, we propose a policy-centric training paradigm that reframes skills as a dynamic training scaffold. Our framework, PATS, converts rollout groups from the latest policy into evidence cards and uses task-specific evaluation to adjust the context used in subsequent rollouts. Concrete guidance helps weak policies to complete challenging tasks. As policy improves, redundant context is revised or removed to reduce reliance on explicit guidance while preserving useful rollout variation. The policy is optimized with environmental rewards using standard RLVR, and the training scaffold is discarded at deployment. Across ALFWorld, WebShop, and seven search-augmented QA benchmarks, PATS achieves performance competitive with SOTA baselines while using 25%-50% fewer tokens.

Yipeng Shi, Zhi-Peng Ma, Yue Wang et al. · 0 citations