商汤AI LLM Development Lead

Team Introduction: TikTok Content Security Algorithm Research Team The International Content Safety Algorithm Research Team is dedicated to maintaining a safe and trustworthy environment for users of ByteDance's international products. We develop and iterate on machine learning models and information systems to identify risks earlier, respond to incidents faster, and monitor potential threats more effectively. The team also leads the development of foundational large models for products. In the R&D process, we tackle key challenges such as data compliance, model reasoning capability, and multilingual performance optimization. Our goal is to build secure, compliant, and high-performance models that empower various business scenarios across the platform, including content moderation, search, and recommendation. Research Project Background: In recent years, Large Language Models (LLMs) have achieved remarkable progress across various domains of natural language processing (NLP) and artificial intelligence. These models have demonstrated impressive capabilities in tasks such as language generation, question answering, and text translation. However, reasoning remains a key area for further improvement. Current approaches to enhancing reasoning abilities often rely on large amounts of Supervised Fine-Tuning (SFT) data. However, acquiring such high-quality SFT data is expensive and poses a significant barrier to scalable model development and deployment. To address this, OpenAI's o1 series of models have made progress by increasing the length of the Chain-of-Thought (CoT) reasoning process. While this technique has proven effective, how to efficiently scale this approach in practical testing remains an open question. Recent research has explored alternative methods such as Process-based Reward Model (PRM), Reinforcement Learning (RL), and Monte Carlo Tree Search (MCTS) to improve reasoning. However, these approaches still fall short of the general reasoning performance achieved by OpenAI's o1 series of models. Notably, the recent DeepSeek R1 paper suggests that pure RL methods can enable LLM to autonomously develop reasoning skills without relying on the expensive SFT data, revealing the substantial potential of RL in advancing LLM capabilities. 团队介绍： 国际化内容安全算法研究团队致力于为字节跳动国际化产品的用户维护安全可信赖环境，通过开发、迭代机器学习模型和信息系统以更早、更快发掘风险、监控风险、响应紧急事件，团队同时负责产品基座大模型的研发，我们在研发过程中需要解决数据合规、模型推理能力、多语种性能优化等方面的问题，从而为平台上的内容审核、搜索、推荐等多项业务提供安全合规，性能优越的基座模型。 课题介绍: 课题背景： 近年来，大规模语言模型（Large Language Models, LLM）在自然语言处理和人工智能的各个领域都取得了显著的进展。这些模型展示了强大的能力，例如在生成语言、回答问题、翻译文本等任务上表现优异。然而，LLM 的推理能力仍有很大的提升空间。在现有的研究中，通常依赖于大量的监督微调（Supervised Fine-Tuning, SFT）数据来增强模型的推理性能。然而，高质量 SFT 数据的获取成本高昂，这对模型的开发和应用带来了极大的限制。 为了提升推理能力，OpenAI 的 o1 系列模型通过增加思维链（Chain-of-Thought, CoT）的推理过程长度取得了一定的成功。这种方法虽然有效，但在实际测试时如何高效地进行扩展仍是一个开放的问题。一些研究尝试使用基于过程的奖励模型（Process-based Reward Model, PRM）、强化学习（Reinforcement Learning, RL）以及蒙特卡洛树搜索算法（Monte Carlo Tree Search, MCTS）等方法来解决推理问题，然而这些方法尚未能达到 OpenAI o1 系列模型的通用推理性能水平。最近deepseek r1在论文中提到通过纯强化学习的方法，可以使得 LLM 自主发展推理能力，而无需依赖昂贵的 SFT 数据。这一系列的工作都揭示着强化学习对LLM的巨大潜力。 课题挑战： 1、Reward模型的设计：在强化学习过程中，设计一个合适的reward模型是关键。Reward模型需要准确地反映推理过程的效果，并引导模型逐步提升其推理能力。这不仅要求对不同任务精准设定评估标准，还要确保reward模型能够在训练过程中动态调整，以适应模型性能的变化和提高。 2、稳定的训练过程：在缺乏高质量SFT数据的情况下，如何确保强化学习过程中的稳定训练是一个重大挑战。强化学习过程通常涉及大量的探索和试错，这可能导致训练不稳定甚至模型性能下降。需要开发具有鲁棒性的训练方法，以保证模型在训练过程中的稳定性和效果。 3、如何从数学和代码任务上拓展到自然语言任务上：现有的推理强化方法主要应用在数学和代码这些CoT数据量相对丰富的任务上。然而，自然语言任务的开放性和复杂性更高，如何将成功的RL策略从这些相对简单的任务拓展到自然语言处理任务上，要求对数据处理和RL方法进行深入的研究和创新，以实现跨任务的通用推理能力。 4、推理效率的提升：在保证推理性能的前提下，提升推理效率也是一个重要挑战。推理过程的效率直接影响到模型在实际应用中的可用性和经济性。可以考虑利用知识蒸馏技术，将复杂模型的知识传递给较小的模型，以减少计算资源消耗。另外，使用长思维链（Long Chain-of-Thought, Long-CoT）技术来改进短思维链（Short-CoT）模型，也是一种潜在的方法，以在保证推理质量的同时提升推理速度。

Team Introduction： The Search Team is primarily responsible for the innovation of search algorithm and architecture research and development (R&D) for products such as Douyin, Toutiao, and Xigua Video, as well as businesses like E-commerce and Local Services. We leverage cutting-edge machine learning technologies for end-to-end modeling and continuously push for breakthroughs. We also focus on the construction and performance optimization of distributed and machine learning systems — ranging from memory and disk optimization to innovations in index compression and exploration of recall and ranking algorithms — providing students with ample opportunities to grow and develop themselves. The main areas of work include: 1. Exploring Cutting-Edge NLP Technologies: From basic tasks like word segmentation and Named Entity Recognition (NER) to advanced business functions like text and multimodal pre-training, query analysis, and fundamental relevance modeling, we apply deep learning models throughout the pipeline where every detail presents a challenge. 2. Cross-Modal Matching Technologies: Applying deep learning techniques that combine Computer Vision (CV) and Natural Language Processing (NLP) in search, we aim to achieve powerful semantic understanding and retrieval capabilities for multimodal video search. 3. Large-Scale Streaming Machine Learning Technologies: Utilising large-scale machine learning to address recommendation challenges in search, making the search more personalized and intuitive in understanding user needs. 4. Architecture for data at the scale of hundreds of billions: Conducting in-depth research and innovation in all aspects, from large-scale offline computing and performance and scheduling optimization of distributed systems to building high-availability, high-throughput, and low-latency online services. 5. Recommendation Technologies: Leveraging ultra-large-scale machine learning to build industry-leading search recommendation systems and continuously explore and innovate in search recommendation technologies. 团队介绍： 字节跳动搜索团队主要负责抖音、今日头条、西瓜视频等产品以及电商、生活服务等业务的搜索算法创新和架构研发工作。我们使用前沿的机器学习技术进行端到端建模并不断创新突破，同时专注于分布式系统、机器学习系统的构建和性能优化，从内存、Disk等优化到索引压缩、召回、排序等算法的探索，充分给同学们提供成长自我的机会。 主要工作方向包括： 1、探索前沿的NLP技术：从基础的分词、NER，文本、多模态预训练，到业务上的Query分析、基础相关性等，全链路应用深度学习模型，每个细节都充满挑战； 2、跨模态匹配技术：在搜索中应用CV+NLP深度学习技术，实现多模态视频搜索强大的语义理解和检索能力； 3、大规模流式机器学习技术：应用大规模机器学习，解决搜索中的推荐问题，让搜索更加个性化更加懂你； 4、千亿级数据规模的架构：从大规模离线计算，分布式系统的性能、调度优化，到构建高可用、高吞吐和低延迟的在线服务的方方面面都有深入研究和创新； 5、推荐技术：基于超大规模机器学习技术，构建业界领先的搜索推荐系统，对搜索推荐技术进行探索和创新。 课题背景/目标： 随着大模型技术的快速发展，智能搜索领域迎来了新的机遇和挑战。传统搜索技术在面对海量数据、多模态信息以及用户复杂需求时，逐渐暴露出模型容量不足、语义理解能力有限、资源利用率低等问题。基于大模型的智能搜索构建旨在通过引入大模型技术，提升搜索系统的智能化水平，优化用户体验，并解决超大规模检索、复杂语义理解、资源高效利用等核心问题。具体目标包括： 1、探索大模型与排序算法的结合，提升个性化排序的精度和用户体验； 2、研究生成式检索算法，解决百亿乃至千亿级别候选库的超大规模检索问题； 3、利用大语言模型（LLM）提升复杂多义query的搜索满意度； 4、构建高性能、低资源消耗的大规模批流一体检索和计算系统，提升资源利用率。 课题挑战/必要性： 1、个性化排序的挑战：传统排序算法难以充分利用多模态信息（如文本、图像、视频等），且模型复杂度有限，无法满足用户对精准化和个性化搜索的需求； 2、超大规模检索的挑战：传统判别式模型在千亿级别候选库的检索中，面临模型容量不足、索引效率低下等问题，亟需新一代检索算法； 3、复杂query理解的挑战：用户搜索需求日益复杂，传统搜索引擎难以准确理解长难句、多义query的语义，导致搜索结果满意度低； 4、资源利用率的挑战：搜索系统存储和计算分离的架构导致资源利用率低，如何在保证性能的同时优化资源使用成为关键问题； 5、基于大模型的智能搜索构建是解决上述挑战的必要途径。通过引入大模型技术，可以显著提升搜索系统的语义理解能力、检索效率和资源利用率，从而为用户提供更精准、更高效的搜索体验。 课题内容： 1、个性化排序大模型研究； 2、超大规模生成式检索算法研究； 3、基于LLM提升复杂多义query的搜索满意度； 4、高性能大规模批流一体检索和计算系统。 涉及的研究方向：排序大模型、生成式检索与跨模态融合、大语言模型（LLM）与复杂query理解、高性能计算与存储架构。

• Lead the architecture, design, and development of intelligent agent systems that integrate LLMs with real-world applications. • Drive full-stack engineering implementation, including backend services, API integration, database design, and task orchestration. • Select and optimize system components such as message queues, middleware, vector databases, and caching frameworks to meet performance and scalability targets. • Work closely with product and research teams to translate AI agent logic (e.g., tool-use, planning, reasoning) into robust, production-grade systems. • Take ownership of system performance tuning, including concurrency handling, throughput optimization, and service reliability. • Guide the team through best practices in code quality, CI/CD pipelines, and system observability. • Build and lead a team of engineers to deliver high-quality agent-driven applications from prototype to deployment.

Strategic Leadership: · Research and formulate the development of AI technology and business strategies, covering areas such as autonomous driving, embodied intelligence, physical AI, generative AI, etc. · Drive cross-functional collaboration to identify emerging opportunities and define roadmaps for AI adoption and innovation. Technology Trend Analysis: · Monitor global AI advancements, competitive landscapes, and regulatory changes (e.g., AI ethics, compliance frameworks) to inform strategic decision-making. · Evaluate cutting-edge technologies (e.g., LLMs, reinforcement learning, multimodal AI) and assess their potential business impact. Ecosystem Development: · Build partnerships with academia, industry consortia, and startups to foster co-innovation and leverage external expertise. · Establish and maintain collaborative relationships with research institutions, industry associations, and other key stakeholders, and represent the company at conferences, technical committees, etc. Insight Delivery: · Produce strategic and technical insight reports and presentations for executive leadership to inform global perspectives for business planning, and translate insights into strategic recommendations.