{"code":0,"data":{"job_post_detail":{"id":"7687904703179213099","title":"agent后训练算法实习生-2027届","description":"工作内容\n\n1. Agent 后训练全流程建设： 负责基于大语言模型的通用 Agent 的后训练（Post-training）工作，覆盖 SFT、偏好对齐（DPO/ORPO/KTO 等）、RLHF/RLAIF、RLVR 等阶段，针对 Tool-use、多轮规划、长程推理、错误恢复等核心能力进行定向训练；\n2. Agentic SFT 数据工程： 构建高质量 Agent 轨迹数据生产线，包括真实交互轨迹采集、模型自博弈（Self-play）生成、轨迹筛选与清洗、拒绝采样（Rejection Sampling）、多样性增强等策略，持续迭代数据配方与课程学习方案；\n3. Preference \u0026 Reward Modeling： 训练面向 Agent 行为的 Reward Model / Preference Model，覆盖任务完成率、工具调用正确性、推理链质量、步骤效率、安全性等多维信号；探索 Process Reward Model（PRM）对中间推理/动作步骤的细粒度建模；\n4. RL for Agent： 研究并落地适合 Agent 场景的强化学习算法，解决长 horizon 信用分配、稀疏奖励、工具环境交互不稳定、离线-在线 RL 融合等关键问题；在代码执行、浏览器操作、API 调用等交互式环境中训练 agent 策略；\n5. 多轮交互与记忆能力训练： 针对 Agent 的多轮对话记忆、长上下文工具使用、从反馈中学习（online adaptation）等能力，设计专项训练任务与对齐方案；\n6. 训练 Infrastructure 优化： 基于 DeepSpeed / Megatron / vLLM / SGLang / veRL 等框架，优化大规模分布式后训练 pipeline 的稳定性、吞吐效率和显存利用率，支撑百亿到千亿参数模型的训推一体迭代；\n7. 评测与归因： 建立系统化的 Agent 能力评测基准（覆盖工具使用、规划、推理、鲁棒性、安全性等维度），结合自动评测与人工评估对训练效果进行量化分析，badcase 归因后驱动下一轮数据与训练策略迭代；\n8. 前沿复现与创新： 跟踪并复现 Agent 后训练方向的最新工作（如 OpenAI o-series、DeepSeek-R1、Tool-integrated Reasoning、Agent Q*、Reinforced Fine-tuning 等），结合业务场景提出创新方案并落地验证。","requirement":"基本要求\n- 计算机、人工智能、数学、统计等相关专业在读硕士/博士（优秀本科生亦可）；\n- 扎实的机器学习与深度学习基础，深入理解 Transformer 架构与 LLM 训练/推理原理；\n- 熟悉强化学习基础算法（PPO、DQN、Policy Gradient、Actor-Critic 等），了解 LLM 对齐算法（RLHF、DPO、GRPO、KTO 等）的原理与实现细节；\n- 熟练使用 Python 与 PyTorch，有实际的大模型训练/微调经验（SFT、LoRA、RLHF 任一）；\n- 了解分布式训练相关技术（数据并行、张量并行、流水线并行、ZeRO 等），具备多卡/多机训练实操经验；\n- 数学功底扎实，能独立推导优化目标、分析 loss 设计与梯度行为。","city_info":{"code":"CT_11","name":"北京","en_name":"Beijing","i18n_name":"北京"},"recruit_type":{"id":"202","name":"实习","en_name":"Intern","i18n_name":"实习","depth":2,"parent":{"id":"2","name":"校招","en_name":"Campus","i18n_name":"校招","depth":1,"children":[],"active_status":1,"selectability":1},"children":[],"active_status":1,"selectability":1},"publish_time":1789979972059,"channel_online_status":1,"code":"A16404","job_id":"7687904703179016491","city_list":[{"code":"CT_11","name":"北京","en_name":"Beijing","location_type":3,"i18n_name":"北京","py_name":"beijing","mdm_code":"MDCY00000001","node_status":1}],"job_post_info":{"recruitment_type":{"id":"202","name":"实习","en_name":"Intern","i18n_name":"实习","depth":2,"parent":{"id":"2","name":"校招","en_name":"Campus","i18n_name":"校招","depth":1,"children":[],"active_status":1,"selectability":1},"children":[],"active_status":1,"selectability":1},"HighlightList":[],"JobChannelPublishList":[],"job_post_object_value_map":{},"address_list":[],"city_list":[],"correlation_job_list":[],"tag_list":[],"storefront_list":[],"target_major_list":[],"job_post_process_time_list":[],"job_level_id_list":[]},"city_info_list_for_delivery":[{"code":"CT_11","name":"北京","en_name":"Beijing","location_type":3,"i18n_name":"北京","py_name":"beijing","mdm_code":"MDCY00000001","node_status":1}],"tag_list":[],"storefront_mode":1,"storefront_list":[],"process_type":2},"recommend_job_post_List":[]},"message":"ok","error":null}