KV cache and long-context inference
Work on KV cache selection, eviction, and long-context serving under memory constraints.
Examples: query-oriented KV selection (QuoKA), KV cache eviction for long contexts (KeyDiff).
Junyoung Park
Senior AI Researcher at Qualcomm AI Research
I work on efficient inference systems for large language models, especially KV cache selection, speculative decoding, and long-context serving under resource constraints.
Efficient LLM inference
KV cache management
Speculative decoding
Long-context inference
RL for combinatorial optimization
Research Focus
I work mostly on efficient LLM inference, with continuing interests in speculative decoding and reinforcement learning for combinatorial optimization.
Speculative decoding
Systems work on draft-model alignment and recursive decoding for faster language model serving.
Examples: recursive speculative decoding, draft-model alignment for speculative decoding.
RL for combinatorial optimization
Earlier work on reinforcement-learning methods for routing, scheduling, and optimization libraries.
Examples: parallel autoregressive policies for multi-agent optimization (PARCO), an RL library for combinatorial optimization (RL4CO).
Publications
NeurIPS 2025
NeurIPS 2025
KDD 2025
ICLR Workshop 2026
Transportation Science 2025
Genetic Algorithms with Neural Cost Predictor for Solving Hierarchical Vehicle Routing Problems
ICLR Workshop 2024
Earlier publications and selected preprints 14 papers
Expand to view work from 2023 back to 2018.
MSEC 2023
ICLR 2023
AAMAS 2023
Learn to Solve the Min-Max Multiple Traveling Salesmen Problem with Reinforcement Learning
arXiv 2023
FOCA: First-Order Context-Based Adaptation for Generalizing to New Dynamical Systems
arXiv 2023
arXiv 2022
Continuous-Depth Neural Models for Dynamic Graph Prediction
arXiv 2022
ScheduleNet: Learn to Solve Multi-Agent Scheduling Problems with Reinforcement Learning
IJPR 2021
IEEE Trans. Sustainable Energy 2019
Wind Field-Based Short-Term Turbine Response Forecasting by Stacked Dilated Convolutional LSTMs
IF = 9
Energy 2019
IEEE Power & Energy Society 2018
Predicting Wind Turbine Power and Load Outputs by Multi-Task Convolutional LSTM Model
Latest News
Recent paper acceptances, workshop activity, and research updates.
- Mar 2026 CAOTE accepted to ICLR 2026 Workshop MemAgents.
- Mar 2026 QuoKA (query-oriented KV selection for efficient LLM prefill) accepted to ICLR 2026.
- Sep 2025 Two papers accepted to NeurIPS 2025: KeyDiff (KV cache eviction for long-context LLM inference) and PARCO (parallel autoregressive models for multi-agent combinatorial optimization).
- Aug 2025 Routefinder accepted to TMLR 2025.
- Jul 2025 RL4CO accepted to KDD 2025. Check out rl4co.ai4co.org.
- Mar 2024 Recursive Speculative Decoding accepted to the LLM Agents Workshop at ICLR 2024.
- Oct 2023 Started as Senior AI Researcher at Qualcomm AI Research to work on efficient LLM.
Education
Training in industrial engineering, optimization, and machine learning at KAIST.