About me
- I have been an Applied Scientist II at Amazon since May 2024. Before joining Amazon, I obtained my CS PhD from Georgia Institute of Technology specializing in Machine Learning, with publications at ICLR, NeurIPS, ACL, EMNLP, EACL, and more, co-advised by Prof. Srijan Kumar and Prof. Mustaque Ahamad.
- At Amazon, I post-train LLMs (from 35B to 1.6T MoE LLMs) and build self-improving coding and ads agents on the Amazon Ads Foundation Model & AWS CodeLLM teams. My current focus areas are:
- LLM Post-Training: SFT, preference alignment (DPO and KTO), RL from AI/human feedback and verifiable rewards (e.g., PPO, GRPO, DAPO, GSPO) including agentic long-horizon RL and reasoning RL, and model consolidation (data & weight mixing, OPD/MT-OPD/OPSD)
- Recursive Self-Improvement (RSI): Self-evolving agents that update their own harnesses (skills, memory, prompts, and workflows), and autonomous AI research systems that improve and post-train LLMs
- Agents, Datasets, and Benchmarks: Tool-use and coding agents, LLM reasoning, synthetic datasets, data curation pipelines and RL environments, LLM-as-a-judge, and benchmarks
- I also serve on the program committees (e.g., as a reviewer) of top AI/ML/LLM conferences, such as ICLR, IJCAI, KDD, and WWW. Besides, I am always available for academic collaborations from academia and industry. Please feel free to contact me if you have research ideas or industry collaborations.
News
One paper, How Do Latent Reasoning Methods Perform Under Weak and Strong Supervision?, accepted to EACL 2027, Main Findings!
Adaptive Auto-Harness: Sustained Self-Improvement on Open-Ended Task Streams accepted as a poster at the NeurIPS 2026 Workshop on Continual Learning for Enterprise AI Agents (CLEA)!
Two papers on self-evolving LLM agents — Harness Updating Is Not Harness Benefit and Evo-Harness — accepted to the EMNLP 2026 main conference!
Updated our benchmark blog with Kimi-K3 results — a three-family open-source comparison across Qwen3.5, GLM5.2, and Kimi-K3.
One paper, A Reward-Guided Dual-Phase Framework for Adaptive Inference-Time Reasoning, accepted to ACL 2026, Main Findings!
Position paper Agentic Evolution is the Path to Evolving LLMs accepted as a poster at the First Workshop on Agent Skills, co-located with ACM CAIS 2026 (the ACM Conference on AI and Agentic Systems).
New blog post: Designing Self-evolving Coding Agents for Terminal-Bench 2.0 — reaching 76.47% and top #7 on the TB2 leaderboard.
New blog post: Revisiting DeepSeek’s 13 Research Papers — the complete technical evolution of DeepSeek base and reasoning models.
TRAJECT-Bench: A Trajectory-Aware Benchmark for Evaluating Agentic Tool Use accepted to ICLR 2026!
Two new blog posts: Building and Aligning LLM using User Data and RL and Awesome Cursor Training Tutorial.