Reasoning & post-training
RLVR, rollout efficiency, reward signals, and the internal dynamics of model reasoning.
Research workbench
Research
I study model behavior through controlled experiments, with particular interest in reasoning, post-training, evaluation, and scientific machine learning.
Independent research · Ongoing · 2026
Detectability of Doomed Reasoning Chains
Linear probes can predict eventual correctness above chance, but separating trajectory signal from problem difficulty is the central open issue.
- verifier-graded rollouts
- 16K
- pooled probe AUC
- 0.708
- token savings
- 7.4%
at 80% kill precision; 0.93 advantage correlation
Question
Can hidden states reveal that a reasoning trajectory is unlikely to succeed early enough to save rollout compute?
Method
Linear probes over residual-stream states from 16K verifier-graded MATH rollouts, followed by simulated probe-guided truncation.
Open issue
Problem difficulty can masquerade as trajectory-level failure signal. Calibration and out-of-distribution generalization remain unresolved.
Working questions
Research interests
Evaluation & interpretability
Measurements that distinguish real model behavior from dataset artifacts and confounding.
Scientific ML
Models whose design reflects complex biological, clinical, and temporal data-generating processes.
Selected writing
Publications
Peer-reviewed work across genomics, deep learning, and cellular trajectory analysis.
- 2025
Comparative Analysis of AAV Serotypes for Transduction of Olfactory Sensory Neurons
Belfort B, Jia J*
Frontiers in NeuroscienceCo-first author
- 2023
Deep learning for detecting and elucidating human T-cell leukemia
Xu H, Jia J*, Jeong H, Zhao Z
Cell PatternsCo-first author
- 2021
Investigating cellular trajectories in COVID-19 severity
Jeong H, Jia J, Dai Y, Simon L
Genes
* Co-first authorship where indicated.
Find my work on Google Scholar