Qihang (Lucas) Jin
M.Eng. Student at USTC · Research Assistant with the Liam Paninski Lab

Qihang (Lucas) Jin 金启航

Columbia University

Research on neural–behavior modeling, representation geometry, and the reliability and adaptation of multimodal systems. Applying to PhD programs starting Fall 2027.

Publications Research Email
Google Scholar GitHub LinkedIn Twitter / X
Qihang (Lucas) Jin
Research

NeuroAI & Neural–Behavior Modeling

Multimodal representations connecting behavior, multi-view video, and neural population activity.

Representation Geometry & Spatial Coding

Auditing topology, geometry, decodability, and out-of-distribution usability in learned spatial representations.

Reliable & Adaptive Multimodal Systems

Diagnosing evaluation failures, routing bottlenecks, and continual-adaptation trade-offs in multimodal models.
Research Experience
Dec. 2025 – present

Columbia University — Liam Paninski Lab

Research Assistant
  • SABLE: cross-view geometry initialization, keypoint-assisted camera calibration, and point-cloud visualization for 3D animal-behavior reconstruction, with controlled ablations over 28-keypoint versus paws-and-nose supervision on Cheese3D.
  • Longitudinal IBL neural-behavior analysis: initial quality control across 60 sessions and POSSM baseline reproduction.
Aug. 2026 – present

Same Estimate, Different Update?

First-Author Research Project
  • Investigating whether task-trained recurrent networks maintaining the same decoded content can nonetheless implement different observation-response laws for future sensory evidence.
  • Designed a causal analysis framework combining mechanism-known calibration systems, history-matched counterfactual observation probes, and content-preserving hidden-state interventions to identify and validate low-rank sensory-revision control subspaces in task-trained recurrent networks.
  • Developed a planned reliability-OOD evaluation framework testing whether operator mismatch from a circular Bayesian reference predicts signed over- or under-correction — with matched-model controls, cross-architecture replication (RNN/GRU/LSTM), and a two-stage impair+rescue falsification protocol.
Education
Publications

* denotes equal contribution.

Projects
github.com/Lucas-Jin-Qh
Loading repositories…
Contact
Email
qihangjin@mail.ustc.edu.cn
Location
Hefei, Anhui, China
Affiliations
University of Science and Technology of China
Columbia University — Liam Paninski Lab
Profiles
GitHub Google Scholar LinkedIn Twitter / X
CV
Download PDF