Long-context understanding
Evidence retrieval, information compression, citation grounding, and multi-document reasoning.
AI Researcher · LLM Post-training
Zhaoyuan Xia 夏照源
M.S. Student in Artificial Intelligence at Peking University · Algorithm Intern at Baidu
I build language-model systems that can read, reason over, and act on long, information-dense contexts with better evidence grounding.
Building reliable AI systems
Profile
About me
I am a master's student in Artificial Intelligence at the School of Software and Microelectronics, Peking University, where I study long-context understanding, model post-training, and evaluation for language and vision-language models.
I am currently an algorithm intern at Baidu, working on long-context post-training for foundation models. Previously, I built prompt optimization agents at Baidu and geo-temporal reasoning benchmarks for vision-language models at SenseTime.
I received my B.S. in Computer Science and Technology from Central South University in 2024. I enjoy turning research ideas into measurable, deployable systems.
PyTorch
Transformers
SFT / RLHF / DPO
RAG
Agents
What I work on
Research interests
LLM post-training
SFT, GRPO, process rewards, and length curriculum for more capable and controllable language models.
AI agents
Self-evolving prompt optimization systems that diagnose errors, test hypotheses, and reuse experience.
VLM evaluation
Benchmarks and analysis for geographic, temporal, and cross-view reasoning in vision-language models.
Selected work
Publications
2026
2026
Where I have worked
Experience
Baidu · Wenxin Foundation Model
Algorithm Intern · Long-context post-training
- Proposed the H2S reading paradigm and made evidence localization, summarization, and answer generation explicit in one autoregressive trajectory.
- Built a structured training corpus across 11 benchmark families for long-context reasoning.
- Trained and evaluated 7B/14B models for evidence-grounded long-context understanding.
Baidu · Mobile Ecosystem Group
Algorithm Intern · Basic search
- Built PEAgent, a self-evolving prompt optimization system with error clustering, hypothesis testing, slot-level edits, rollback, and candidate selection.
- Designed teacher–student collaboration between strong models for diagnosis and planning and efficient models for batch evaluation.
- Improved authority, satisfaction, and relevance metrics by 10.5, 13.3, and 4.0 percentage points and helped launch the internal platform.
SenseTime · Smart City Group
Algorithm Research Intern · Multimodal AI
- Led the development of GTR-Bench for geo-temporal reasoning across maps, videos, and non-overlapping camera views.
- Evaluated more than 10 mainstream VLMs and analyzed gaps in spatial-temporal utilization, temporal prediction, and cross-view alignment.
- The work was released with an open-source codebase.
Selected projects
Projects
Open source · 2026
Highlight-Then-Summarize
Long-context post-training framework and dataset for compressing distributed evidence before answering.
LLMGRPO128K context
View project
Research
GTR-Bench
Benchmarking geographic and temporal reasoning in vision-language models across multi-view environments.
VLMBenchmarkVideo + map
View project
Industry research
PEAgent
A prompt optimization agent that turns model errors into testable hypotheses and iterative improvements.
AgentsPrompt optimizationEvaluation
Deployed internally
Beyond research
Honors
2023National Second Prize · Service Outsourcing Innovation CompetitionTeam leader
2023Peking University Challenge Cup · First PrizeSpecial contribution track
—Peking University “Lixing” Program · Outstanding IndividualUndergraduate honors
Get in touch
Let’s talk about
reliable AI systems.
I am interested in long-context reasoning, post-training, evaluation, and practical systems that make model behavior easier to inspect.
2401210464@stu.pku.edu.cn
Beijing, China