Hello! I am Jisen Li, an AI Researcher at Together AI. I recently graduated with a B.S. in Mathematics and Computer Science from the University of Illinois Urbana-Champaign (UIUC). My research interests focus on reinforcement learning (RL), inference, and AI agents. At Together AI, I work on RL algorithms and platforms in collaboration with Ivan Provilkov and Max Ryabinin, and previously worked on speculative decoding and quantization with Dr. Shirley Wu and Dr. Ben Athiwaratkun.

In industry experience, I have interned at Nexa AI, where I worked on on-device models and AI agents; at Snowflake, focusing on embedding models and hybrid retrieval systems; and at TikTok, where I developed multimodal recommendation models.

In academic research, I have worked at UIUC U-Lab under the supervision of Prof. Jiaxuan You, researching AI agent communication protocols and multi-agent benchmarking. I have also conducted research at the University of Chicago under the supervision of Dr. Eliu A. Huerta, focusing on AI deployment, and at Tsinghua University under the guidance of Prof. Yongqiang Lyu, focusing on computer vision and digital signal processing.

Outside of research, I enjoy fitness and strategy games such as Go, and I’m always up for a game of billiards. Whether it’s a Go board or a pool table, I love games that reward careful planning and precise execution.

Research Interests

• Post-Training: Reinforcement learning (RL) algorithms, agentic RL with environment rewards, and scalable RL training platforms.
• Inference: Speculative decoding, KV cache quantization, and efficient model serving.
• Agentic AI: Agents with memory, planning, and decision-making for collaboration.

News 🎉

Selected Publications

^* Equal contribution, ^† Core contributor. Full list on Google Scholar.

Industry Experience

Together AI – AI Researcher Jul 2026 – Present

• RL Post-Training Platform (Together Post-Training): Contributed to Together’s post-training service, which lets users write their own SFT/RL training loops while the platform handles trainer–generator orchestration, asynchronous RL, and multi-LoRA serving.
• Diffusion LLM Post-Training (ORL): On-policy residual learning that trains diffusion LLMs on their own denoising trajectories, achieving ~6.9× end-to-end throughput.
• Quantization-Aware Training (QATFactory): Quantization-aware distillation and RL framework for LLMs that exports deployment-ready NVFP4, MXFP4, and Q4_K models.

Together AI – Research Intern Jan 2026 – May 2026

• Speculative Decoding (Aurora): Built a training-serving system that continuously learns speculators from live inference traces, achieving 1.45× speedup on frontier models.
• KV-Cache Quantization (SAW-INT4): Token-wise INT4 KV-cache quantization with block-diagonal Hadamard rotation, near-lossless at zero serving overhead.
• KV-Cache Quantization (OSCAR): Offline covariance-aware rotation for 2-bit KV-cache quantization, with ~8× memory reduction and up to 7× throughput.

Nexa AI (acquired by Qualcomm) – AI Research Engineer Intern Oct 2025 – Dec 2025

• On-Device AI Agent: Built an on-device tool-use agent with Qwen3-VL and contributed to nexa-sdk for efficient NPU inference.

Snowflake – Software Engineer Intern (AI/ML) May 2025 – Aug 2025

• Cortex Search: Fine-tuned Arctic Embed v2.0 for retrieval (+15% nDCG) and built Gaussian-process auto-tuning for hybrid search, deployed to production.

TikTok – Machine Learning Engineer Intern Aug 2024 – Dec 2024

• Multimodal Recommendation: Developed Rec-Qwen, a LLaVA-style recommender aligning user and content embeddings to a frozen Qwen2.5 backbone, with a 28% CTR accuracy gain.

Academic Research

Together AI Sep 2025 – Dec 2025

• KV Cache Quantization (Kitty): Mixed-precision 2-bit KV caching with dynamic channel-wise precision boost, achieving 8× memory reduction at near-zero accuracy loss. Supervised by Dr. Shirley Wu.

University of Illinois Urbana-Champaign Jul 2025 – Present

• Scientific Literature Search Agent (PaperPilot): Multi-turn search agent that induces executable DAG workflows over paper-search tools, improving Hit@5 from 58.0 to 77.0. Supervised by Prof. Bingxin Zhao and Prof. Jiaxuan You.
• AI Agent Protocol Benchmarking (ProtocolBench): Built a unified benchmark for multi-agent communication protocols across document QA, collaborative coding, and multi-agent path finding. Supervised by Prof. Jiaxuan You.

Earlier Research 2020 – 2024

• University of Chicago (Globus Lab): AI model deployment on HPC. Supervised by Dr. Eliu A. Huerta.
• Tsinghua University: Non-contact HRV monitoring with computer vision. Supervised by Prof. Yongqiang Lyu.

Education

University of Illinois Urbana-Champaign (UIUC) Aug 2023 – May 2026
B.S. in Mathematics + Computer Science, GPA: 3.98/4.0

🌱 Gap Year Jan 2023 – Aug 2023

Huazhong University of Science and Technology (HUST) Sep 2021 – Jan 2023
Undergraduate Studies (Transferred)

Academic Service

• Reviewer, ICLR 2027
• Program Committee, ACL-SELVA 2026