Personal homepage launched.
Xihang Shan · 单夕航
I am committed to building AI systems that test before they trust.
I am Xihang Shan (单夕航), a mathematics undergraduate at Xiamen University. My research connects machine learning, AI agents, causal decision-making, graph reasoning, and mathematical modeling. I am grateful to my undergraduate mentors, Prof. Da Zhou and Prof. Ye Luo, for their guidance.
01Machine Learning
02AI Agents & Decision-Making
03Causal & Graph Reasoning
04Mathematical Modeling
News
Recent updates
Summer research collaboration at HKUST with Prof. Can Yang.
Completed the NeurIPS 2026 rebuttal; post-rebuttal reviews: 4/4/5.
Three manuscripts submitted to AAAI 2027.
National-level undergraduate research program launched at Xiamen University.
Attended the Cryptography and Mathematics Summer School at the Institute of Information Engineering, Chinese Academy of Sciences, and was named an Outstanding Student.
Selected manuscripts
Research that questions its assumptions.
Across causal learning, agents, and graphs, I design controls that reveal when external knowledge helps—and when a model should refuse it.
NeurIPS 2026 · Under review01
PRCD-MAP: Learning How Much to Trust Imperfect Priors in Causal Discovery
Learns edge-specific trust from data, using useful prior structure while safely reverting to data-only discovery when the prior is misleading.
Post-rebuttal reviews: 4/4/5
AAAI 2027 · Under review02
Testing Before Trusting: Causal Skeptic Bandits under Unreliable Historical Evidence
Treats historical causal knowledge as a hypothesis to be tested online; diagnostic falsification and a causal veto keep misleading graphs from steering decisions.
AAAI 2027 · Under review03
DeltaNAR: Learning What to Reuse and What to Recompute in Neural Algorithmic Reasoning
Separates state that remains valid after a graph update from state that must be recomputed, enabling incremental reasoning without unsafe copying.
AAAI 2027 · Under review04
Failure-Learning Claims Need Controls: A Claim-Control Evaluation of Language Agents
Uses claim-matched controls to distinguish repair, transfer, target-solving, and executable validity from ordinary task success.
EMNLP 2026 · Under review05
Bounded Path Context: Visible Path History in LLM-Based KGQA
Uses bounded visible path history to reduce prompt exposure while preserving effective, auditable symbolic reasoning.
PRICAI 2026 · Under review06
Recipe-Controlled Decoder Audit for Structural Knowledge-Graph Completion
Shows how decoder choice and training recipe can confound structural KGC comparisons, motivating controlled reporting across architectures and datasets.
* Corresponding author.
Other projects
01 / Research agents
Research Memory Auditor
Bounded, claim-relevant memory views and fail-closed evidence auditing for research agents.
Open repository ↗02 / Discrete structure
Boolean Function Structure
Walsh-spectral affine approximation, derivative-guided affine covers, and feedforward sequence recovery.
Repository & report ↗03 / Spatial AI
LocSource
Conservative, auditable transcript-ownership proposals for Xenium post-segmentation analysis.
Repository & report ↗Education
Education
Xiamen University
B.S. in Mathematics and Applied Mathematics
School of Mathematical Sciences
Honors
Honors
Fujian First Prize
National Undergraduate Mathematical Modeling Contest · Team Leader
Fujian Third Prize
National Undergraduate Mathematical Modeling Contest · Team Leader
2025
Outstanding Student
Cryptography & Mathematics Summer School
Beyond research
Things I enjoy away from the screen.
01
Basketball
Long-time NBA fan · Always happy to talk basketball
02
Billiards
03
Bridge
04
Piano
05
Touch Rugby
Contact