CV

Doctor of Philosophy

North Carolina State University

Computer Science

• Advisor: Jung-Eun Kim
• Research Interests: Efficient Inference-time LLM Reasoning and Trustworthy AI
• Award: 2026-2027 Provost’s Doctoral Fellowship

Master of Engineering

Duke University

Electrical and Computer Engineering

• GPA: 3.83 / 4.0
• Research Experience: Collaborated with Duke CEI on research projects in trustworthy AI.
• Award: Duke Kunshan University Scholarship Award (162K RMB)

Bachelor of Engineering

Shenzhen University

Computer Science and Technology

• GPA: 3.75 / 4.0
• Awards/Honor: Two-time winner of The Second Award of Studying Star in 2020 and 2021 (ranked 4 and 6); Outstanding Graduate of the Year 2022

StepBack: Step-Level Error-Localized Resampling for Efficient Test-Time Reasoning

Yang Ouyang, Jung-Eun Kim

In Proceedings of NeurIPS 2026

DenseSteer: Steering Small Language Models towards Dense Math Reasoning

Yang Ouyang, Shuhang Lin, Jung-Eun Kim

In Proceedings of ICML 2026

Layer-Level Self-Exposure and Patch: Affirmative Token Mitigation for Jailbreak Attack Defense

Yang Ouyang, Hengrui Gu, Shuhang Lin, Wenyue Hua, Jie Peng, Bhavya Kailkhura, Meijun Gao, Tianlong Chen, Kaixiong Zhou

In Proceedings of NAACL 2025

Min-K%++: Improved Baseline for Detecting Pre-Training Data of LLMs

Jingyang Zhang, Jingwei Sun, Eric Yeats, Yang Ouyang, Martin Kuo, Jianyi Zhang, Hao Frank Yang, Hai Li

In Proceedings of ICLR 2025

StepBack: Step-Level Error-Localized Resampling for Efficient LLM Reasoning

In Proceedings of NeurIPS 2026

• Proposed StepBack, a training-free inference-time decoding method that improves self-consistency by localizing likely first-error positions in chain-of-thought trajectories and resampling only uncertain suffixes instead of regenerating full reasoning paths.
• Designed lightweight step-level rollback signals using NLL increases and process reward model score drops to allocate test-time compute more efficiently within reasoning trajectories.
• Evaluated across math reasoning and multi-hop QA benchmarks; the PRM-based variant improved accuracy on 10 / 14 model-dataset pairs while reducing tokens in all settings, with up to 39.8% token savings over SC@32.

Large-Scale Elastic Continued Pretraining for MoE Models

Project

• Adapted elastic training to a closed-source 15B-A2B MoE model using 1,024 GPUs, enabling three nested, weight-sharing subnetworks (3B, 9B, 15B) for different resource budgets.
• Explored five architectural dimensions (MoE depth, expert width, expert count, attention heads, and top-k routing) through budget-constrained sampling and importance-guided pruning, constructing nested subnetworks under joint total-parameter and active-parameter constraints.
• Optimized shared-weight continued pretraining across 11 benchmarks, outperforming size- and training-budget-matched standalone models by 1.6 and 3.9 points for small and large subnetworks, respectively.

DenseSteer: Steering Small Language Models towards Dense Math Reasoning

In Proceedings of ICML 2026

• Proposed Dense Reasoning as a structural property of strong mathematical reasoning, empirically showing that higher reasoning density (fewer steps, more tokens per step) correlates with higher accuracy on GSM8K and OOD benchmarks.
• Developed a training-free, inference-time activation steering framework using only 50 rewrite-based contrastive pairs, avoiding distribution shift and preserving low token-level NLL.
• Achieved up to +5% GSM8K, +17.5% AMC, and +10% AIME accuracy gains on Qwen-2.5-3B-Instruct and LLaMA-3.2-3B-Instruct compared with heavy training-based distillation methods.

Software Engineer Intern, DeepJavaLibrary Department

Amazon Web Services | AI Infrastructure

• Developed core integration components between DJL-Serving and KServe, implementing 3 Java-based APIs for health monitoring, model introspection, and inference data processing within the model serving stack.
• Supported cloud-native model serving on Kubernetes by configuring deployment manifests, service endpoints, and runtime parameters for containerized inference workloads.