Institution profile

TikTok

Industry researchnorthamerica · us
Official website
Research library81linked papers
Opportunities887open roles
Selected work

Representative Papers

SimpleRL-Zoo: Investigating and Taming Zero Reinforcement Learning for Open Base Models in the Wild

Mar 24, 2025arXiv.org

This work investigates the universality and training dynamics of zero-shot reinforcement learning (Zero RL) across heterogeneous foundation models. Method: We systematically evaluate whether chain-of-thought (CoT) reasoning emerges directly from base models—without explicit CoT supervision—across ten open-source models spanning diverse architectures and scales. We introduce two key design principles: format reward shaping and query difficulty control, and integrate rule-based RL, implicit CoT supervision, joint monitoring of response length and verification behavior, and a cross-model training dynamics analysis framework. Contribution/Results: We observe, for the first time, a “reasoning insight moment” in non-Qwen small-scale models. Our analysis reveals a non-monotonic relationship between model scale and training dynamics. Experiments demonstrate significant improvements in reasoning accuracy and response length across most models. To foster reproducibility, we open-source all code, fine-tuned models, and analytical tools.

25 citations4 influentialRead paper

Dr. Kernel: Reinforcement Learning Done Right for Triton Kernel Generations

Feb 05, 2026

This work addresses the challenges of reward hacking and lazy optimization that hinder large language models in generating high-performance, correct Triton kernels. To overcome these issues, the authors propose Dr.Kernel-14B, a 14-billion-parameter model trained via a multi-round reinforcement learning framework within the distributed GPU environment KernelGYM. The approach introduces the TRLOO algorithm to eliminate policy gradient bias caused by self-containment in GRPO, integrates profiling-based rewards with a mismatch correction mechanism to mitigate lazy optimization, and employs rejection sampling to enhance generation quality. Experimental results demonstrate that Dr.Kernel-14B achieves speedups of over 1.2× on 31.6% of kernels in KernelBench Level-2, outperforming Claude-4.5-Sonnet (26.7%) and GPT-5 (28.6%); this success rate rises to 47.8% when selecting the best-performing kernel across multiple generations.

1 citations1 influentialRead paper

When Rules Fall Short: Agent-Driven Discovery of Emerging Content Issues in Short Video Platforms

Jan 14, 2026

This study addresses the rapidly evolving content risks on short-video platforms, where traditional manual detection mechanisms suffer from low efficiency and delayed policy updates, thereby limiting governance effectiveness. The authors propose the first automated problem discovery framework based on a multimodal large language model (LLM) agent. The approach identifies potentially violating videos, applies a two-stage clustering method to generate emerging issue clusters, and automatically produces executable annotation policies. This work pioneers the use of multimodal LLM agents for defining and detecting novel content risks, enabling rapid iteration of annotation strategies. Experimental results demonstrate that, compared to manual processes, the proposed method improves the F1 score for emerging issue detection by over 20%, reduces associated video views by approximately 15%, and significantly shortens the policy update cycle.

1 citationsRead paper

AgentPS: Agentic Process Supervision for Multi-modal Content Quality Assurance through Multi-round QA

Dec 15, 2024arXiv.org

Multimodal large language models (MLLMs) exhibit limited performance on content moderation tasks requiring fine-grained logical reasoning. Method: We propose an agent-driven process supervision framework that deeply integrates structured sequential reasoning with multi-round question-answering (QA) fine-tuning—marking the first end-to-end embedding of process supervision into the multi-round QA training pipeline. The framework enables LLMs to autonomously generate high-quality supervision signals, eliminating reliance on manual annotations while preserving both reasoning fidelity and industrial-scale scalability. Results: Experiments on a private TikTok dataset demonstrate substantial improvements over state-of-the-art baselines. Notably, using LLM-generated labels alone retains over 90% of the performance gain achieved with human annotations, validating the framework’s effectiveness and practical deployability in large-scale real-world scenarios.

1 citationsRead paper
Recent publications

Latest Papers

DAEP: Difficulty-Aware Evidence Planning for Medical Video Corpus Temporal Answer Grounding

Aug 07, 2026

This work addresses the challenge posed by the disparity between simple and complex questions in medical video question answering by proposing a difficulty-aware multimodal evidence planning approach. The method integrates caption, visual, and procedural context evidence to rank candidate videos, then generates and refines temporal segments through high-scoring anchor expansion. Its key innovation lies in transforming explicit question difficulty labels into dynamic control strategies during inference, enabling adaptive adjustment of modality weights, evidence aggregation schemes, boundary thresholds, and reranking intensity. Evaluated on the NLPCC 2026 Shared Task, the proposed model achieved first place with an average score of 0.2728. Ablation studies further demonstrate that the introduced mechanisms yield particularly significant performance gains on complex questions.

0 citationsRead paper