Researcher positions at INSIGHT Lab

We're Hiring — INSIGHT Lab, Ben-Gurion University

M.Sc. · Ph.D. · Postdoc

Led by Prof. Chaim Baskin · School of Electrical and Computer Engineering

If you believe the next generation of AI will need to understand the physical world — not just describe it — you should know what we're building toward. For over a decade, we've focused on making deep learning efficient enough to actually deploy: quantization, compression, architecture search. Recently, we've been turning that into something bigger. We showed that vision-language models can run up to 80% cheaper by predicting the minimum resolution a query actually needs — before tokenization even starts [1]. We built a benchmark that revealed how current Video-LLMs fail when reasoning requires integrating evidence scattered across time [2] — the kind of failure that matters when a model needs to understand what's happening in the real world, not just caption a single frame. We introduced a principled way to adapt Vision Transformers to multiple tasks simultaneously without interference [3], and we rethought how graph neural networks mix information across neighbors [4], laying the groundwork for architectures that generalize across structured domains. None of these are isolated results. They're the building blocks for where we're heading next.

[1] https://arxiv.org/abs/2510.19496

[2] https://arxiv.org/abs/2512.14870

[3] https://arxiv.org/abs/2502.06029

[4] https://arxiv.org/abs/2409.19414

What we work on

🌍 Efficient World Models — building predictive environment models that are compact and fast enough for real-time generative interaction, bridging the gap between world simulation and practical deployment

🎬 Efficient Video & Multimodal Reasoning — developing vision-language systems that reason over long temporal horizons under real compute budgets, from frame selection to multi-evidence integration

🔗 Graph Foundation Models — scaling graph neural networks toward general-purpose pre-trained models for heterogeneous, dynamic, and structured data

⚡ Efficient Diffusion Models — compressing and accelerating diffusion architectures for scalable generation without sacrificing quality

We also maintain active research in quantization, compression, adversarial robustness, LLM safety, and discourse-level language understanding.

📈 We publish regularly at NeurIPS, CVPR, ICLR, ICCV, ECCV, EMNLP, and TMLR.

What you'd actually do here

You won't just run baselines. Depending on the role, you might design efficient architectures for world models that need to run in real time, build graph foundation models that generalize across domains, develop lightweight inference pipelines that make video-language reasoning practical at scale, or push diffusion models toward deployment-ready efficiency. You'd be joining at a moment where the foundational pieces are in place and the most exciting work is just starting.

💻 Infrastructure: Multiple terabytes of VRAM across 14 dedicated multi-GPU nodes — including Blackwell-generation hardware. You won't wait in a queue to test your ideas at scale.

We also have active collaborations with industry partners and top academic groups internationally.

Who we're looking for

M.Sc. — B.Sc. from a top Israeli university (honors preferred), strong hands-on deep learning experience

Ph.D. — Master's in a related field, at least one top-tier publication

Postdoc — 4+ top-tier publications, independent research direction