Undergraduate researcher in artificial intelligence

Hengyu Jin

I am interested in the reasoning capabilities and reliability of foundation models, particularly how they behave and interact when deployed as agents in complex environments.

Education

Tongji UniversityB.Eng. in Software Engineering · GPA 4.60 / 5.00

Current

Visiting Student · PRADA Lab, KAUSTResearch Intern · Aether AI
Curriculum Vitae Email me Google Scholar GitHub
Recent updates

Released our first-author preprint on latent reasoning faithfulness across training trajectories and submitted it to ACL Rolling Review.

RISER appeared in Findings of ACL 2026.

Kairotask was published at CHI 2026.

01 / Research

Research interests

My work centers on understanding how foundation models reason and building AI agents that behave reliably in practice.

01

Mechanistic interpretability

Understanding and intervening in internal representations through training-dynamics analysis and activation steering.

02

Reasoning in foundation models

Studying how models reason faithfully and efficiently across latent and explicit reasoning settings.

03

Reliable AI agents

Building verifiable environments for multi-user agents, alongside execution-time failure detection and recovery methods for embodied agents.

02 / Publications

Publications & preprints

Official paper pages are linked through arXiv, ACL Anthology, and publisher DOI records.

2026
First authorarXiv · Under review at ACL Rolling Review

Final Checkpoints Are Not Enough: Analyzing Latent Reasoning Faithfulness Along Training Trajectories

Hengyu Jin, Shu Yang, Di Wang

Tracks how latent reasoning faithfulness evolves across training checkpoints, revealing dynamics that final-checkpoint evaluation alone can miss.

arXiv PDF
2026
Findings of ACLFindings of ACL 2026

RISER: Orchestrating Latent Reasoning Skills for Adaptive Activation Steering

Wencheng Ye, Xiaoyang Yuan, Yi Bin, Hengyu Jin, Liang Peng, Pengpeng Zeng, Heng Tao Shen

A plug-and-play activation-steering framework that adaptively composes reusable latent reasoning skills for more accurate and token-efficient reasoning.

ACL Anthology PDF
2026
IEEE TACAccepted · IEEE Transactions on Affective Computing

CORA: A Cognitive Reframing Dialogue Agent Powered by Large Language Models

T. Chen, X. Chen, Y. Shen, L. Zhang, Hengyu Jin, C. Cheng

A large-language-model-powered dialogue agent for multi-turn cognitive reframing grounded in cognitive behavioral therapy.

DOI
2026
CHI 2026CHI 2026

Kairotask: Probing the Bridge Between Vague Intents and Spatiotemporal Contexts

Jiachen Du, Yuanhao Zhang, Hengyu Jin, Xinkai Zou, Junwei Zhu, Xinyi Fu

A context-aware task assistant that uses language models to turn vague intentions into actionable spatiotemporal cues.

DOI

03 / Experience

Research experience

Jul 2026 — PresentThuwal, Saudi Arabia

PRADA Lab, KAUST

Visiting Student · Advisor: Prof. Di Wang

  • Mechanistic analysis of latent reasoning faithfulness across training trajectories.
  • MUENV, a verifiable interactive environment for multi-user agent training and evaluation.
Jun 2026 — PresentRemote

Aether AI

Research Intern · Advisor: Prof. Biwei Huang (UC San Diego)

  • Execution-time failure verification for vision-language-action agents, including natural failure-data generation in Genie Sim.
Jun 2025 — Oct 2025Remote

Tsinghua University

Research Intern · Advisor: Prof. Xinyi Fu

  • Kairotask: context-aware scheduling from vague intentions, combining Qwen-Plus reasoning with an iOS geofencing field study.
Oct 2024 — Dec 2025Shanghai, China

Tongji University

Research Intern · Advisor: Prof. Ying Shen

  • Hidden-state steering for latent reasoning through reusable cognitive primitive vectors.
  • CORA, including LoRA-based supervised fine-tuning and blinded evaluation for CBT-guided multi-turn dialogue.

04 / Projects

Selected projects

Co-founder · Full-stack leadNov 2025 — Feb 2026

Follo AI

A context-aware scheduling agent extended from the Kairotask framework.

Designed a response–perception–memory architecture combining LLM-based intent interpretation, environmental context, and structured user memory; led mobile and backend development and delivered a functional MVP.

AI agentsiOSContext awareness
Team lead · Core developerJun 2025 — Sep 2025

Scrollscape

An immersive Chinese-garden education experience for Apple Vision Pro.

Implemented visionOS interaction and rendering components, including gesture-based navigation, virtual–physical scene transitions, and an automated aesthetic-evaluation module.

Vision ProSpatial computingHuman–computer interaction

05 / Background

Education & recognition

Education

Tongji University

B.Eng. in Software Engineering · Sep 2023 — Present

GPA 4.60 / 5.00 · 91.01 / 100

Recognition

Selected awards

First Prize, China Collegiate Computing Contest — East China Region

First-grade Ethnic Class Special Scholarship, 2023–2025

Toolbox

Building research systems

Python · C++ · Swift · PyTorch · Hugging Face Transformers

SwiftUI · React · FastAPI · Git · Linux · Slurm