Algorithm Engineer, Agent Post-Training Data

What you’ll do

1. 01Agent post-training data systems: design and iterate on data for agent SFT, preference learning, and reinforcement learning across task planning, tool use, multi-turn interaction, memory, reflection, and error correction. Establish data quality assessment and filtering mechanisms, and optimize data distribution, difficulty, and diversity.
2. 02Data algorithm design and implementation: develop algorithms for agent data generation, trajectory collection, model self-labeling, rejection sampling, and difficulty evolution. Build pipelines for task generation, environment interaction, trajectory collection, and automated verification to produce high-value post-training data more efficiently.
3. 03Agent evaluation and capability diagnosis: design evaluation datasets and scoring criteria covering task completion, planning and decision-making, tool use, long-horizon interaction, and adaptation to environments. Combine programmatic verification, environment feedback, and LLM-as-a-judge to analyze failure patterns and trajectories and identify model capability gaps.
4. 04Data flywheel and engineering implementation: connect task construction, environment execution, trajectory filtering, model evaluation, and data iteration into a closed loop. Turn failure cases, human expert feedback, and evaluation results into targeted data production plans and pipelines.
5. 05Technical innovation and reusable resources: track advances in LLM agent data and evaluation, including Trajectory Synthesis, Process Reward, Verifiable Reward, Environment Scaling, and agent benchmark. Tackle core technical challenges and develop reusable data algorithm components, datasets, and evaluation benchmarks.

What we look for

1. 01A bachelor’s degree or higher in computer science, artificial intelligence, machine learning, or a related field, with R&D experience in LLM post-training data, agent data, or model evaluation.
2. 02Solid foundations in machine learning, natural language processing, and data mining, with an understanding of post-training workflows such as SFT, preference alignment, and reinforcement learning, as well as agent planning, tool use, and environment interaction.
3. 03Experience building instruction data, preference data, agent interaction trajectories, or evaluation datasets, and familiarity with data generation, quality filtering, automated verification, and identifying the causes of model failures.
4. 04Proficiency in Python and common LLM development tools, with strong data processing, algorithm implementation, and experimental analysis skills, and the ability to independently design data plans and validate their effectiveness.
5. 05Strong problem-framing and collaboration skills, with the ability to translate model capability gaps or business scenarios into actionable data and evaluation plans.

Nice to have

• Experience building data for Code agent, Browser agent, or Tool agent systems, or for agents that perform long-horizon tasks.
• Hands-on experience with sandbox environments, programmatic verification, a Reward Model, a Verifier, or LLM-as-a-judge.
• Training experience with SFT, DPO, GRPO, RLHF / RLAIF, or agentic RL.
• High-quality papers, open-source projects, or experience in algorithm competitions.