Portrait of Gabriele Serussi

Gabriele Serussi

I build world models that spend only the compute they need.

I'm a PhD student at Ben-Gurion University, advised by Prof. Chaim Baskin in the INSIGHT Lab, and a Research Intern at Decart, where I work on real-time video generation. Before my PhD, I earned an MSc in Computer Science at the Technion and a BSc in Computer Engineering at Sapienza University of Rome.

I'm always happy to talk about world models, efficient video models, robotics and research collaborations.

Highlights

  • CVPR 2026

    HERBench

    Each question forces a model to combine at least three cues from different parts of a video, and 13 Video-LLMs score only 31 to 42%.

  • Forward and reverse diffusion bridges between two modalities, with the marginal matching and cycle consistency constraints drawn over their trajectories.
    ICML 2026

    Structured Diffusion Bridges

    Diffusion bridges that reach near fully paired quality with half the pairs and still learn with none.

  • Score curves from several experts over a basketball broadcast, combined into one satisfaction curve that picks four frames.
    WACV 2027

    HiMu

    Picks 16 frames that beat 64 uniformly sampled ones on Video-MME, without any training.

  • Simulated sound pressure around a four-rotor drone in a square room, with positive pressure in red and negative in blue.
    IROS 2024 · Oral pitch

    Rotor-noise localization

    A drone that localizes itself by listening to its own rotor noise, which it learns to shape.

Research

A useful world model has to understand, generate and act on a budget. My work covers all three, and efficiency runs through it, from what a model looks at to the hardware it runs on.

Generate

Generative world models

Real-time video generation at Decart, and diffusion bridges that translate between modalities with less paired data, or none.

In progress

  • Autoregressive video generation

    Generating long videos one chunk at a time while keeping them consistent.

  • Diffusion distillation

    Distilling diffusion models into faster generators.

  • Evidence aggregation in Transformers

    Helping Transformers combine many pieces of evidence, the failure HERBench exposed in Video-LLMs.

  • Disaggregated inference

    Serving large models efficiently across different kinds of hardware.

Newest direction: memory for world models, deciding what a generator or a robot should keep when it cannot keep everything.

News

  • Our proposal on finding and fixing the failures of generative AI, which I co-wrote with Prof. Chaim Baskin, won a research grant from Israel's Ministry of Innovation, Science and Technology in a competitive national call.
  • HiMu, a training-free frame selector for long videos, was accepted at WACV 2027.
  • VEDJE, a cached reranker for scalable video-text retrieval, was posted to arXiv.
  • Two of our papers were presented at ICML 2026 in Seoul: Structured Diffusion Bridges in the main conference and VEDJE at the AdaptFM workshop.
  • HERBench was presented at CVPR 2026 in Denver.

Selected publications

All publications

  • HERBench: A Benchmark for Multi-Evidence Integration in Video Question Answering

    Dan Ben-Ami*, Gabriele Serussi*, Kobi Cohen, Chaim Baskin

    CVPR 2026Co-first author11k+ dataset downloads

    HERBench forces video models to aggregate evidence: each of its 26,806 questions needs at least three separate cues from distinct parts of the video, and 13 state-of-the-art Video-LLMs reach only 31 to 42% (random is 20%).

    BibTeX
    @InProceedings{Ben_Ami_2026_CVPR,
    author    = {Ben Ami, Dan and Serussi, Gabriele and Cohen, Kobi and Baskin, Chaim},
    title     = {HERBench: A Benchmark for Multi-Evidence Integration in Video Question Answering},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2026},
    pages     = {4505-4514}
}
  • Forward and reverse diffusion bridges between two modalities, with the marginal matching and cycle consistency constraints drawn over their trajectories.

    Structured Diffusion Bridges: Inductive Bias for Denoising Diffusion Bridges

    Eitan Kosman, Gabriele Serussi, Chaim Baskin

    ICML 2026

    Structured Diffusion Bridges reach near fully paired quality in modality translation with half the paired data, 25.2 PSNR against 25.6 for a fully paired baseline on zero-shot super-resolution, and still learn with no pairs at all.

    BibTeX
    @inproceedings{kosman2026structured,
title={Structured Diffusion Bridges: Inductive Bias for Denoising Diffusion Bridges},
author={Eitan Kosman and Gabriele Serussi and Chaim Baskin},
booktitle={Forty-third International Conference on Machine Learning},
year={2026},
url={https://openreview.net/forum?id=nWI6lV7BvA}
}
  • Score curves from several experts over a basketball broadcast, combined into one satisfaction curve that picks four frames.

    HiMu: Hierarchical Multimodal Frame Selection for Long Video Question Answering

    Dan Ben-Ami, Gabriele Serussi, Kobi Cohen, Chaim Baskin

    WACV 2027

    With 16 frames chosen by HiMu, a training-free frame selector for long-video question answering, Qwen3-VL-8B scores 73.2% on Video-MME, above its 71.7% with 64 uniformly sampled frames.

    BibTeX
    @misc{benami2026himuhierarchicalmultimodalframe,
      title={HiMu: Hierarchical Multimodal Frame Selection for Long Video Question Answering},
      author={Dan Ben-Ami and Gabriele Serussi and Kobi Cohen and Chaim Baskin},
      year={2026},
      eprint={2603.18558},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2603.18558},
}
  • Simulated sound pressure around a four-rotor drone in a square room, with positive pressure in red and negative in blue.

    Active propulsion noise shaping for multi-rotor aircraft localization

    Gabriele Serussi*, Tamir Shor*, Tom Hirshberg, Chaim Baskin, Alex M. Bronstein

    IROS 2024Oral pitchCo-first author

    Also presented at ICML 2024 Workshop (AI4Science).

    Learning a drone's rotor phases together with a neural localizer lets it localize itself from its own propulsion noise, more than 2.7 times better in simulation than the best handcrafted modulation.

    BibTeX
    @inproceedings{Serussi_2024,
  title={Active propulsion noise shaping for multi-rotor aircraft localization},
  url={http://dx.doi.org/10.1109/IROS58592.2024.10802115},
  DOI={10.1109/iros58592.2024.10802115},
  booktitle={2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  publisher={IEEE},
  author={Serussi, Gabriele and Shor, Tamir and Hirshberg, Tom and Baskin, Chaim and Bronstein, Alex M.},
  year={2024},
  month=oct,
  pages={472--479}
}
  • Plot of MSR-VTT text-to-video R@1 against online model parameters, with VEDJE marked at 59.8 and about 157M parameters among published systems from 63M to 7.6B.

    VEDJE: Video-Efficient Discriminative Joint Encoder for Scalable Video-Text Retrieval

    Shahaf Wagner*, Gabriele Serussi*, Dan Ben Ami, Tomer Galanti, Chaim Baskin

    ICML 2026 Workshop (AdaptFM)Co-first author

    VEDJE matches a 7.6B-parameter multimodal LLM retriever on MSR-VTT text-to-video retrieval (59.8 against 59.7 R@1) with about 157M online parameters, roughly 48 times fewer, by encoding each video once into a compact cache read by a 33M-parameter reranker.

    BibTeX
    @misc{wagner2026vedjevideoefficientdiscriminativejoint,
      title={VEDJE: Video-Efficient Discriminative Joint Encoder for Scalable Video-Text Retrieval},
      author={Shahaf Wagner and Gabriele Serussi and Dan Ben Ami and Tomer Galanti and Chaim Baskin},
      year={2026},
      eprint={2610.11850},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2610.11850},
}
  • PREGEN pipeline: video and text tokens pass through a frozen LLM, and the final-token hidden state of every layer is pooled by a small encoder into one embedding.

    PREGEN: Uncovering Latent Thoughts in Composed Video Retrieval

    Gabriele Serussi*, David Vainshtein*, Jonathan Kouchly*, Dotan Di Castro, Chaim Baskin

    arXiv 2026Co-first author

    Reads a frozen VLM's hidden states to reach 99.73 Recall@1 on WebVid-CoVR, 27.23 points above the previous best.

    BibTeX
    @misc{serussi2026pregenuncoveringlatentthoughts,
      title={PREGEN: Uncovering Latent Thoughts in Composed Video Retrieval},
      author={Gabriele Serussi and David Vainshtein and Jonathan Kouchly and Dotan Di Castro and Chaim Baskin},
      year={2026},
      eprint={2601.13797},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2601.13797},
}

* Equal contribution.

Experience and education

Grants and awards

  • Adversarial Evaluation and Hardening of Generative AI Trustworthiness

    Research grant from Israel's Ministry of Innovation, Science and Technology, won in a competitive national call with Prof. Chaim Baskin as PI, to build methods and an open-source toolkit that find and fix the failures of generative models. I helped shape the research plan and aims, wrote major sections of the proposal and now work on the project.

    2026 to 2029

Get in touch

I'm always happy to talk about world models, efficient video models, robotics and research collaborations.