My research focuses on building autonomous and self-evolving agents that
continuously learn, adapt, and improve through procedural memory and
agentic reinforcement learning. I study how agents
can leverage both their own experience and knowledge shared by other agents to develop
reusable capabilities, adapt to evolving environments, and improve their reasoning and
decision-making over time. Beyond agent learning, I also study how to make foundation models
more efficient and trustworthy: how to reduce inference cost without sacrificing quality,
and how to explain what drives their outputs.
Before starting my Ph.D., I received an M.S. in Applied Statistics from the
University of Michigan–Ann Arbor, advised by Professors
Kerby Shedden and
Yang Chen, where my research focused on
high-dimensional statistics and time series.
I am actively looking for research internship and full-time research scientist / engineer /
applied scientist roles in 2027! Please feel free to contact me,
and referrals are greatly appreciated!
Updates
News
Two papers were accepted to NeurIPS 2026.
One paper was accepted to Findings of EMNLP 2026.
ForecastCompass was selected as a Spotlight at Forecast@ICML 2026.
Started an Applied Scientist internship at Amazon.
Received the ICML Gold Reviewer recognition.
Started an Applied Scientist internship at Amazon.
Two papers were accepted to ACL 2025.
One paper was accepted to The Web Conference 2024.
MemCollab contrasts reasoning trajectories from agents with different backbone
models to distill abstract reasoning constraints that capture shared task-level
invariants, building a single memory system that can be shared across heterogeneous agents.
25.4% average improvement on Qwen2.5-7B · 29.3% cross-family improvement · 38.5% fewer reasoning turns
Yurui Chang, Yongkang Du, Yuanpu Cao, Jinghui Chen, Lu Lin
ForecastCompass converts forecast–retrospective discrepancies into
hierarchical factor and calibration memories that can be reused across
future events and model backbones.
Shows that MLLM cascades expose a new attack surface: a universal border trigger
lowers the weak model's confidence and forces queries to be deferred to the
expensive strong model, manipulating compute allocation without targeting answer correctness.
Yuanpu Cao, Yongkang Du, Yurui Chang, Lu Lin, Jinghui Chen
AuditForecast anchors forecasts with a quantitative baseline model, then applies explicit
situational factor updates in odds space, turning prose-based judgment into an auditable,
structured process that improves accuracy and calibration at lower cost.
Monitoring Decoding detects hallucination-prone tokens during generation
and selectively revises them through monitor-guided tree decoding instead
of repeatedly sampling complete responses.
+16% TruthfulQA · +13.6% GSM8K · up to 2.78× faster than Best-of-N
Yurui Chang, Bochuan Cao, Yujia Wang, Jinghui Chen, Lu Lin
JoPA formulates prompt attribution as discrete combinatorial optimization
and uses gradient-guided probabilistic search to identify prompt components
that jointly shape an LLM's complete generation.
Up to 100× more efficient than leave-one-out and integrated-gradient baselines
Graph Distribution Matching synthesizes compact interpretive graphs by
matching feature-space distributions throughout GNN training, revealing
global class-level patterns while preserving model fidelity.
Up to +38.5 model fidelity over GNNInterpreter · >90% predictive accuracy on 5 of 6 datasets · ~4× faster than XGNN
Jennifer A. Schaub, Christopher L. O'Connor, Meghan Dailey, Andrew W. Hlynka, Yurui Chang, Deborah Postiff, Samuel D. Kaffenberger, Ganesh S. Palapattu, Brenda W. Gillespie, Jeffrey B. Hodgin, Kerby Shedden, Markus Bitzer