字节跳动
校招风控垂域大模型基座和Agentic AI研究-风控(新加坡)Research on Risk Control Vertical LLM Foundation and Agent-Risk Control (Singapore)
新加坡 · 校招 · 2027 届 · 博士
长期有效
任职要求
- 2027届毕业,获得博士学位,软件开发、计算机科学、计算机工程、网络安全或相关技术专业
- 具备出色的编码能力,在数据结构与算法方面基础扎实;需熟练掌握Python,熟悉PyTorch或TensorFlow(TF)者优先
- 具备优秀的问题界定、分析与解决能力;在AAAI、NeurIPS、SIGKDD、SIGIR等CCF-A类期刊或会议上发表过论文者优先
- 具备良好的沟通能力与团队协作能力;热爱技术,愿意与团队共同迎接挑战,拥有创新驱动力。
- Individuals who are completing or recently completed a PhD in Software Development, Computer Science, Computer Engineering, Cybersecurity, or a related technical discipline
- Excellent coding skills and a solid foundation in data structures and algorithms; proficiency in Python is required, and familiarity with PyTorch or TensorFlow (TF) is preferred
- Outstanding ability to define, analyze, and solve problems; candidates with publications in CCF-A category journals or conferences such as AAAI, NeurIPS, SIGKDD, SIGIR, etc., are preferred
- Strong resilience, excellent communication and teamwork skills; passionate about technology, willing to embrace challenges with the team, and a drive for innovation.
岗位描述
团队介绍:风控算法团队深耕于抖音(短视频、直播、电商、本地生活等)、今日头条等多个字节系业务,涵盖内容、交易、流量、账号等多个场景的底线风险治理、黑灰产对抗及复杂反作弊等关键任务。风控场景具有对抗性极强、逻辑推理链路长及海量多模态数据等挑战。目前,团队正积极推进大语言模型(LLM/MLLM)与 Agentic AI 技术的深入研究与工业化落地。通过探索前沿的模型训练范式、自主规划的 Agent 架构以及严谨的 Harness Engineering 体系,我们致力于打造下一代智能风控基础设施,为十亿级用户的社区生态保驾护航。
课题介绍:AI 技术的飞速发展大幅提升了风控场景下的风险识别与对抗能力,但仍面临多维度挑战:主流 SOTA 大模型在强对抗风险内容理解及 AIGC 深度伪造识别上存在明显瓶颈;面对黑产手法的瞬息万变,风控亟需建设基于 Agent 的自主对抗方案,以强化对抗时效并降低运营成本;针对风控场景依赖数据多、判别规则复杂的挑战,需重点突破跨模态长上下文信息抽取及复杂合规标准的指令遵循难题。本课题将通过大模型全链路优化、智能 Agent 体系建设与范式创新,系统性提升全场景风控的智能化水平。
课题挑战:
- 1、打造具有优势的风控垂域大模型基座;
- 2、Agentic AI架构研究与演进;
- 3、大模型结构化数据的理解和应用;
- 4、科学评估与Harness Engineering体系构建。
课题价值:
- 1、打造具有优势的风控垂域大模型,显著增强对复杂、跨领域风险的识别,从底层降低风险漏过率;
- 2、定义风控场景 Agentic AI 的落地范式,构建自演化对抗体系,在持续对抗新型风险的同时,大幅降低人工运营成本。
Topic Content
Current leading large models struggle with understanding highly adversarial risky content and identifying AI-generated content (AIGC). At the same time, underground activities online change quickly and need fast responses. To meet this challenge, risk control needs to develop agent-based autonomous systems that can fight threats more effectively and reduce operational costs. Risk control also faces challenges due to the large amount of data and complex rules involved. It needs better ways to extract cross-modal long-context information and follow complex compliance instructions. This topic aims to improve risk control intelligence across all scenarios by optimizing large models from end to end, building smart agent systems, and creating new paradigms.
Topic Challenges
- 1.Insufficient understanding of underground industry variants, AIGC, and other adversarial content by general large models;
- 2.Challenges in long-context comprehension, information extraction, and instruction adherence;
- 3.Research and Evolution of Agentic AI Architecture;
- 4.Integrating fragmented risk control knowledge into agent-usable skills.
Topic Value
- 1.Develop agent-based approaches that can adapt and fight new risks on their own, cutting operational costs;
- 2.Improve recall of long-tail and adversarial samples to reduce leakage.
你的简历匹配这个岗位吗?
上传简历,AI 会对照这份 JD 指出你的差距和需要补强的地方
该岗位可能会问的问题
基于 算法 岗位的常见面试问题整理
请解释 Transformer 模型的自注意力机制(Self-Attention)
查看答题思路
1. Self-Attention = 序列中每个 token 与所有 token 计算相关性权重 2. Q=Wq·x, K=Wk·x, V=Wv·x; Attention(Q,K,V)=softmax(QK^T/√d_k)V 3. 除以 √d_k 防止点积过大导致 softmax 梯度消失 4. 多头注意力:多组 QKV 并行,拼接后线性变换,捕获不同子空间特征 5. 对比 RNN:可并行计算,捕获长距离依赖
你做过的最有深度的算法/模型项目是什么
查看答题思路
1. 一句话说清任务:分类/生成/检索/推荐 2. 模型选型:为什么选这个而非那个 3. 数据:数据量、来源、清洗策略 4. 创新点:你做的改进是什么(不是调参) 5. 结果:核心指标提升(准确率/召回/延迟)
过拟合是什么?怎么判断和解决
查看答题思路
1. 过拟合=训练集表现好、验证集表现差,模型记住了噪声 2. 判断:train loss 持续下降而 val loss 上升 3. 解决:更多数据 / 正则化(L1/L2) / Dropout / 早停 / 数据增强 / 降低模型复杂度
解释梯度下降的原理,SGD 和 Adam 的区别
查看答题思路
1. 梯度下降:沿负梯度方向更新参数,使损失降低 2. SGD:每次用随机小批量估算梯度,引入随机性帮助跳出局部最优 3. Adam = Momentum + RMSProp:自适应学习率 + 动量加速 4. Adam 收敛快但可能泛化不如 SGD;CV 常用 SGD,NLP 多用 Adam
给你 10 万条未标注的文本数据,如何找出其中 10 个主题
查看答题思路
1. 预处理:分词、去停用词 2. 向量化:TF-IDF 或 Sentence Embedding 3. 聚类:K-Means(需指定 K=10)或 HDBSCAN(自动确定类数) 4. 或 LDA 主题模型直接输出主题-词分布 5. 评估:看每个簇的高频词判断主题合理性,迭代调整
更多面试题库功能即将上线
信息来源:企业官方招聘页。投递前请以官网信息为准。