# Oshri Naparstek

> Research Scientist at IBM | Principal RSM | Master Inventor | Manager, AI for Knowledge group

## Who is Oshri Naparstek?

Oshri Naparstek is a Research Scientist at IBM Research in Haifa, Israel. He holds a PhD in Electrical Engineering and specializes in multimodal AI, complexity theory, cognitive offloading, and reinforcement learning. He is a Principal Research Staff Member (RSM), IBM Master Inventor, and manages the "AI for Knowledge" group.

## Key Facts

- 28+ peer-reviewed publications
- 1,200+ total citations
- 2 US patents
- Co-author of IBM Granite Vision (2B parameter vision-language model)
- Creator of the Real-mm-RAG benchmark (ACL, 21 citations in first year)
- Paper accepted at CVPR 2026 on modality gap in vision-language models
- Paper under review at ICML on token maturation
- Most-cited paper: "Deep Multi-User Reinforcement Learning for Dynamic Spectrum Access" (IEEE JSAC, 579 citations)

## Research Areas

- Multimodal AI: vision-language models, CLIP, modality gap analysis, Granite Vision
- Cognitive Offloading: systems where LLM agents learn to replace their own reasoning with verified deterministic code
- Complexity Theory: observer-dependent complexity, complexity as the performance gap between observers
- Token Maturation: delayed token commitment for reducing hallucinations in LLMs
- Reinforcement Learning: deep multi-user RL, distributed spectrum access, real-time defense RL (Rafael)
- Signal Processing: coprime arrays, interference alignment, distributed wireless optimization

## Career Background

- IBM Research, Haifa (2020–present): Research Scientist, Principal RSM, Master Inventor, Group Manager
- Rafael Advanced Defense Systems: Reinforcement learning for real-time defense applications
- Washington University in St. Louis: Postdoctoral researcher in signal processing and optimization

## Key Ideas and Perspectives

- "The sea squirt principle": AI systems should learn to reduce their own computational role over time, crystallizing repeated reasoning into verified tools
- "AI today is dial-up": Current inference speeds constrain AI to chatbot interactions; at 10,000+ tokens/sec, entirely new application categories emerge
- "The interference alignment lesson": Beautiful theoretical results can fail at scale when synchronization overhead exceeds benefits — applicable to quantum error correction
- "LLMs as formulators, not optimizers": LLMs should generate mathematical formulations and let classical solvers do the optimization
- "Complex means interesting": Complexity is the performance gap between observers with different capabilities

## Links

- LinkedIn: https://www.linkedin.com/in/oshri-naparstek-05580154
- Google Scholar: https://scholar.google.com/citations?user=IpR9NiIAAAAJ
- Email: oshri8@gmail.com
- Website: https://oshrinaparstek.com