I am a researcher on the ByteDance Seed team, focused on coding agents and the reinforcement-learning post-training that drives them. My work centers on software-engineering agents, long-horizon agents, and agent harnesses and sandbox infrastructure.

Publications

The following papers include Shulin Xin in their named author rosters. Coding- and agent-focused work is listed first.

1. Multi-SWE-bench: A Multilingual Benchmark for Issue Resolving
   Shulin Xin (co-author), 2025, arXiv. paper · code · project
   A multilingual benchmark for resolving real issue-tracker tasks across real repositories.
2. SWE-Mirror: Scaling Issue-Resolving Datasets by Mirroring Issues Across Repositories
   Shulin Xin (co-author), 2025, arXiv. paper
   Scales verifiable training data by mirroring human GitHub issues into preconfigured environments.
3. AInsteinBench: Benchmarking Coding Agents on Scientific Repositories
   Shulin Xin (co-author), 2025, arXiv. paper · code
   A benchmark for coding agents across real scientific software repositories.
4. Seed-Coder: Let the Code Model Curate Data for Itself
   ByteDance Seed; Shulin Xin (co-author), 2025, arXiv technical report. paper · code
   An open 8B code-model family built on model-driven data curation.
5. FullStack Bench: Evaluating LLMs as Full Stack Coders
   ByteDance Seed Foundation Code Team; Shulin Xin (co-author), 2024, arXiv. paper · code · sandbox
   An execution-based, multilingual evaluation of LLMs as broad-domain full-stack coders.
6. UI-TARS-2 Technical Report: Advancing GUI Agent with Multi-Turn Reinforcement Learning
   ByteDance Seed; Shulin Xin (co-author), 2025, arXiv. paper · project
   Advancing GUI agents with multi-turn reinforcement learning.
7. Seed1.5-Thinking: Advancing Superb Reasoning Models with Reinforcement Learning
   ByteDance Seed; Shulin Xin (co-author), 2025, arXiv. paper
   Advancing strong reasoning models through reinforcement learning.

Team Model Cards & Reports

ByteDance Seed team-authored reports listed on my Google Scholar profile; these are distinct from papers with individually named author rosters.

1. Seed2.0 Model Card: Towards Intelligence Frontier for Real-World Complexity
   ByteDance Seed, 2026. paper
   A model card for reasoning, visual understanding, search, and real-world agentic capabilities.
2. Seed1.8 Model Card: Towards Generalized Real-World Agency
   ByteDance Seed, 2026. paper
   A team model card on multi-turn, tool-using real-world agency.
3. Seed-OSS Open-Source Models
   ByteDance Seed Team, 2025. model card · repo
   An open-weight, long-context reasoning and agent model family.
4. Technical Introduction to the Seed1.6 Model Series
   ByteDance Seed, 2025. Scholar record
   A technical introduction to the Seed1.6 model series.

Open Source

Public repositories and releases associated with my co-authored research and the broader Seed team.

• Multi-SWE-bench — benchmark and harness for multilingual issue resolution.
• AInsteinBench — coding-agent benchmark on scientific repositories.
• Seed-Coder — open 8B code-model family release.
• FullStackBench — execution-based full-stack coding evaluation.
• SandboxFusion — sandbox infrastructure for execution-based evaluation.
• UI-TARS — project repository for the UI-TARS GUI-agent series.
• Seed-OSS — open-weight Seed model releases.

Education

Master's degree, Beihang University, 2024.

Experience

• 2022 — Megvii Technology
• 2022–2023 — Alibaba, DAMO Academy / Tongyi
• 2024–Present — ByteDance Seed

Technical Focus & Skills

• RL post-training — reinforcement-learning post-training for code and reasoning agents.
• SWE / execution-based evaluation — issue-resolving benchmarks and execution-verified evaluation.
• Long-horizon / multi-turn agents — extended-horizon, tool-using agent behavior.
• Harness & environment infrastructure — agent harnesses, execution environments, and rollout infrastructure.

Public tooling ecosystem (team projects): SandboxFusion, containerized execution environments, and the benchmark and model-release repositories above.