Ruochen Jiao
Applied Scientist, Amazon · Ph.D., Northwestern University
I am a core contributor to pre-training and post-training of large language models at Amazon (Store Foundational AI). My current research centers on making LLMs stronger and more efficient through reinforcement learning, knowledge distillation, and adaptive compute. I received my Ph.D. from Northwestern University, advised by Prof. Qi Zhu, where my work focused on trustworthy embodied agents.
News
- Jan 2026 Two papers accepted to ICLR 2026: SFT Doesn't Always Hurt General Capabilities: Revisiting Domain-Specific Fine-Tuning in LLMs and Shop-R1: Enhancing E-commerce with Reinforcement Learning.
- Jan 2025 Can We Trust Embodied Agents? Exploring Backdoor Attacks against Embodied LLM-based Decision-Making Systems accepted to ICLR 2025.
- Jun 2024 Kinematics-aware Trajectory Generation and Prediction with Latent Stochastic Differential Modeling accepted to IROS 2024.
- Mar 2024 Empowering Autonomous Driving with Large Language Models: A Safety Perspective accepted to LLMAgents @ ICLR 2024.
- Jul 2023 Semi-supervised Semantics-guided Adversarial Training for Robust Trajectory Prediction accepted to ICCV 2023.
- Apr 2023 Enforcing Hard Constraints with Soft Barriers: Safe Reinforcement Learning in Unknown Stochastic Environments accepted to ICML 2023.
Recent Posts
- Apr 2026 Adaptive Think: Training LLMs to Choose When to Think — Teaching models to autonomously route between think/no-think modes via RL with paired rollouts and curriculum scheduling.
Selected Publications
- Can We Trust Embodied Agents? Exploring Backdoor Attacks against Embodied LLM-based Decision-Making SystemsICLR 2025First systematic study of backdoor vulnerabilities in embodied LLM agents. We show that adversarial triggers injected into environment observations can hijack LLM-based planners, causing unsafe actions in robotic and autonomous driving tasks.
- SFT Doesn't Always Hurt General Capabilities: Revisiting Domain-Specific Fine-Tuning in LLMsICLR 2026Challenges the common belief that supervised fine-tuning degrades general capabilities. Through controlled experiments across diverse domains, we identify conditions under which SFT preserves or even improves general performance.
- Kinematics-aware Trajectory Generation and Prediction with Latent Stochastic Differential ModelingIROS 2024A latent stochastic differential equation framework that incorporates vehicle kinematics constraints for generating physically plausible and diverse future trajectories in autonomous driving scenarios.
- Semi-supervised Semantics-guided Adversarial Training for Robust Trajectory PredictionICCV 2023Proposes a semi-supervised adversarial training method that leverages semantic scene understanding to improve trajectory prediction robustness against distribution shifts and adversarial perturbations.
- Enforcing Hard Constraints with Soft Barriers: Safety-driven Reinforcement Learning in Unknown Stochastic EnvironmentsICML 2023Introduces soft barrier functions to enforce hard safety constraints in RL without requiring prior knowledge of environment dynamics, enabling provably safe exploration in stochastic settings.
Experience
- Applied Scientist · Amazon, Seattle
Store Foundational AI · 2024 – present - Applied Scientist Intern · Amazon, Seattle
Jun 2023 – Sep 2023 - Research Scientist Intern · Toyota InfoTech Labs, Mountain View
Jun 2021 – Sep 2021
Education
- Ph.D. in Computer Engineering · Northwestern University
Advisor: Prof. Qi Zhu · 2019 – 2024 - B.S. in Electrical Engineering · Zhejiang University
2015 – 2019
Service
Conference reviewer: NeurIPS, ICLR, ICML, ECCV, ICCV, CVPR, IROS, ICRA, AAAI, AISTATS
Journal reviewer: TMLR, RA-L, TMM, TNNLS, TCAD, TCPS, TIV, IEEE JSAC