Research Scientist · IBM Research
Multimodal AI · Language Generation · Learning & Optimization
I study how AI systems represent information, generate language, and turn reasoning into useful computation. My work connects mathematical ideas with models, experiments, and tools people can explore.
At IBM Research Haifa, I work on multimodal learning and document understanding. I co-authored Granite Vision and REAL-MM-RAG. My background spans a PhD in Electrical Engineering, distributed optimization, and reinforcement learning.
What if a language model could refine a token before committing to it?
Autoregressive language generation through continuous prediction in embedding space, with discrete tokens produced by projection at commitment time. An optional “liquid tail” allows local refinement before commitment. Originally introduced as Token Maturation.
Paper · Project & generation demos
| Project | What it does |
|---|---|
| Graph Map | Collaborative knowledge graphs for AI conversations, with an MCP server and an interactive browser interface. |
| Skill Editor | A visual editor for agent workflows, linking a node graph to the underlying Markdown source. |
| Evolution Agent | An experimental code optimization framework combining genetic programming with a multi-tier LLM strategy. |
- Granite Vision — A lightweight, open-source vision-language model for enterprise document understanding. Co-author.
- REAL-MM-RAG — A benchmark for realistic multimodal document retrieval, including table-heavy documents and rephrased queries. ACL 2025 · Co-author.
- Complexity as Advantage — A regret-based perspective on emergent structure and observer-dependent complexity. Sole author.
- Continuous language generation: separating the evolution of a representation from the decision to emit a token.
- Cognitive offloading: turning repeated AI reasoning into reusable, verifiable computation.
- Representation and complexity: understanding what makes a pattern useful to one observer and invisible to another.
I write about these questions and build interactive explanations at oshrinaparstek.com.
Outside research: music, macro photography, and freediving. For research conversations and collaboration, find me on LinkedIn.