# Gabriele Serussi

> Gabriele Serussi is an AI researcher working on efficient world models and video AI, at Ben-Gurion University's INSIGHT Lab and at Decart.

- Name: Gabriele Serussi
- Occupation: AI researcher
- Current positions: PhD student, INSIGHT Lab, Ben-Gurion University of the Negev; Research Intern, Decart
- Research topics: World models, Video generation, Video understanding, Diffusion models, Vision-language models, Video retrieval, Generative modeling, Computer vision, Machine learning, Robotics, Efficient deep learning, Representation learning, Hardware-aware optimization, Multimodal fusion
- Homepage: https://gabrieleserussi.github.io/
- Bio for introductions: https://gabrieleserussi.github.io/bio/
- Email: serussig@post.bgu.ac.il

Mission, in Gabriele's words: "I build world models that spend only the compute they need."

## About

Gabriele Serussi is an AI researcher working on efficient world models: models that understand video, generate it and act on it while spending only the compute they need. He is a PhD student in Electrical and Computer Engineering at Ben-Gurion University of the Negev, advised by Prof. Chaim Baskin in the INSIGHT Lab, and a Research Intern at Decart, where he works on real-time video generation. Before that, he was a Research Intern at the Bosch Center for Artificial Intelligence, supervised by Prof. Chaim Baskin and Dotan Di Castro, working on video retrieval and on fusing vision with radar.

His research covers long-video understanding, generative world models and robotics. It includes HERBench (CVPR 2026), a video question answering benchmark in which every question needs at least three separate cues from the video, with more than 11,000 dataset downloads; Structured Diffusion Bridges (ICML 2026), which translate between modalities with less paired data; HiMu (WACV 2027), a training-free frame selector for long videos; VEDJE, a cached reranker for scalable video-text retrieval presented at an ICML 2026 workshop; and a drone that localizes itself from its own rotor noise (IROS 2024, oral pitch). He also co-wrote a research grant on the adversarial evaluation of generative AI, funded by Israel's Ministry of Innovation, Science and Technology.

He holds an MSc in Computer Science from the Technion, advised by Prof. Alex Bronstein and Prof. Chaim Baskin, and a BSc in Computer Engineering from Sapienza University of Rome.

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.

Long-video understanding. Benchmarks that show where Video-LLMs fail, and methods that let them look at less or train less. Papers: HERBench (CVPR 2026), HiMu (WACV 2027), VEDJE (ICML 2026 Workshop (AdaptFM)), PREGEN (arXiv 2026).

Generative world models. Real-time video generation at Decart, and diffusion bridges that translate between modalities with less paired data, or none. Papers: Structured Diffusion Bridges (ICML 2026).

Robotics. Drones that shape their own rotor noise to localize themselves, and a growing interest in world models for robots. Papers: Rotor-noise localization (IROS 2024).

In progress: Autoregressive video generation. Generating long videos one chunk at a time while keeping them consistent.

In progress: Diffusion distillation. Distilling diffusion models into faster generators.

In progress: Evidence aggregation in Transformers. Helping Transformers combine many pieces of evidence, the failure HERBench (https://gabrieleserussi.github.io/research/herbench/) exposed in Video-LLMs.

In progress: 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.

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

## Experience and education

- Research Intern, real-time video generation, Decart (2026 to present)
- Research Intern, video retrieval and multimodal fusion of vision and radar, supervised by Prof. Chaim Baskin and Dotan Di Castro, Bosch Center for Artificial Intelligence (2024 to 2026)
- PhD, Electrical and Computer Engineering, INSIGHT Lab, advised by Prof. Chaim Baskin, Ben-Gurion University of the Negev (2024 to present)
- MSc, Computer Science, advised by Prof. Alex Bronstein and Prof. Chaim Baskin, Technion, Israel Institute of Technology (2021 to 2024)
- BSc, Computer Engineering, Sapienza University of Rome (2018 to 2021)

## Grants and awards

- Adversarial Evaluation and Hardening of Generative AI Trustworthiness, Ministry of Innovation, Science and Technology (Israel) (2026 to 2029). 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.

## Research interests

World models, Video generation, Video understanding, Diffusion models, Vision-language models, Video retrieval, Generative modeling, Computer vision, Machine learning, Robotics, Efficient deep learning, Representation learning, Hardware-aware optimization, Multimodal fusion

## Publications

"*" marks equal contribution.

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

Authors: Dan Ben-Ami*, Gabriele Serussi*, Kobi Cohen, Chaim Baskin
Page: https://gabrieleserussi.github.io/research/herbench/
Venue: CVPR 2026, co-first author.
Result: 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%).
Keywords: video question answering, VideoQA benchmark, multi-evidence reasoning, long video understanding, Video-LLMs, frame selection

Abstract: Video Large Language Models (Video-LLMs) are improving rapidly, yet current Video Question Answering (VideoQA) benchmarks often admit single-cue shortcuts, under-testing reasoning that must integrate evidence across time. We introduce HERBench, a benchmark designed to make multi-evidence integration unavoidable: each question requires at least three non-overlapping cues drawn from distinct video segments. HERBench contains 26,806 five-way multiple-choice questions across 12 compositional tasks. To make evidential demand measurable, we introduce the Minimum Required Frame-Set (MRFS), the smallest number of frames a model must fuse to answer correctly, and show that HERBench imposes higher evidential demand than prior benchmarks. Evaluating 13 state-of-the-art Video-LLMs yields only 31-42% accuracy, only modestly above the 20% random-guess baseline. We disentangle this failure into two critical bottlenecks: (1) a retrieval deficit, where frame selectors overlook key evidence, and (2) a fusion deficit, where models fail to integrate information even when all necessary evidence is provided. HERBench thus provides a principled benchmark for studying robust multi-evidence video understanding.

- Paper: https://openaccess.thecvf.com/content/CVPR2026/papers/Ami_HERBench_A_Benchmark_for_Multi-Evidence_Integration_in_Video_Question_Answering_CVPR_2026_paper.pdf
- arXiv: https://arxiv.org/abs/2512.14870
- Project: https://gabrieleserussi.github.io/HERBench/
- Code: https://github.com/DanBenAmi/HERBench
- Dataset: https://huggingface.co/datasets/DanBenAmi/HERBench

```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}
}
```

### Structured Diffusion Bridges: Inductive Bias for Denoising Diffusion Bridges

Authors: Eitan Kosman, Gabriele Serussi, Chaim Baskin
Page: https://gabrieleserussi.github.io/research/sdb/
Venue: ICML 2026.
Result: 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.
Keywords: diffusion bridges, modality translation, unpaired learning, semi-paired learning, cycle consistency, super-resolution, multi-view to 3D

Abstract: Modality translation is inherently under-constrained, as multiple cross-modal mappings may yield the same marginals. Recent work has shown that diffusion bridges are effective for this task. However, most existing approaches rely on fully paired datasets, thereby imposing a single data-driven constraint. We propose a diffusion-bridge framework that characterizes the space of admissible solutions and restricts it via alignment constraints, treating paired supervision as an optional heuristic rather than a prerequisite. We validate our method on synthetic and real modality translation benchmarks across unpaired, semi-paired, and paired regimes, showing consistent performance across supervision levels. Notably, it achieves near fully-paired quality with a substantial relaxation in pairing requirements, and remaining applicable in the unpaired regime. These results highlight diffusion bridges as a flexible foundation for modality translation beyond fully paired data.

- arXiv: https://arxiv.org/abs/2605.02973
- OpenReview: https://openreview.net/forum?id=nWI6lV7BvA

```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}
}
```

### HiMu: Hierarchical Multimodal Frame Selection for Long Video Question Answering

Authors: Dan Ben-Ami, Gabriele Serussi, Kobi Cohen, Chaim Baskin
Page: https://gabrieleserussi.github.io/research/himu/
Venue: WACV 2027.
Result: 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.
Keywords: frame selection, long video question answering, multimodal large language models, training-free, neuro-symbolic reasoning, audio-visual

Abstract: Long-form video question answering requires reasoning over extended temporal contexts, making frame selection a critical bottleneck for multi-modal large language models (MLLMs) bound by finite context windows. Within the controlled frame-budget regime that governs practical deployment, prior selectors score frames against a single global query embedding; as a result, compositional multimodal questions that involve temporal ordering or cross-modal cues such as “what happens on screen right after the narrator mentions the reaction?” are flattened into a representation that loses sub-event ordering and modality bindings. We introduce HiMu, a training-free framework for compositional multimodal frame selection. A single text-only LLM call decomposes the query into a hierarchical logic tree whose leaves are atomic predicates, each routed to a lightweight expert spanning vision (CLIP, open-vocabulary detection, OCR) and audio (speech recognition and non-speech sound matching). Expert signals are normalized, smoothed to align across modalities, and composed bottom-up through fuzzy-logic operators that enforce temporal sequencing and adjacency, yielding a continuous per-frame satisfaction curve. Under the standard 16-frame budget on Video-MME, LongVideoBench, and HERBench-Lite, HiMu achieves state-of-the-art accuracy among frame selection methods and improves over uniform sampling across seven diverse MLLMs as a drop-in module, matching the accuracy of uniform sampling at 4× its frame budget, without retraining and without multiple iterative MLLM calls during selection.

- arXiv: https://arxiv.org/abs/2603.18558

```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},
}
```

### Active propulsion noise shaping for multi-rotor aircraft localization

Authors: Gabriele Serussi*, Tamir Shor*, Tom Hirshberg, Chaim Baskin, Alex M. Bronstein
Page: https://gabrieleserussi.github.io/research/ears/
Venue: IROS 2024, oral pitch, co-first author. Also presented at ICML 2024 Workshop (AI4Science).
Result: 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.
Keywords: drone localization, acoustic localization, multi-rotor aircraft, propulsion noise, rotor phase modulation, robotics

Abstract: Multi-rotor aerial autonomous vehicles (MAVs) primarily rely on vision for navigation purposes. However, visual localization and odometry techniques suffer from poor performance in low or direct sunlight, a limited field of view, and vulnerability to occlusions. Acoustic sensing can serve as a complementary or even alternative modality for vision in many situations, and it also has the added benefits of lower system cost and energy footprint, which is especially important for micro aircraft. This paper proposes actively controlling and shaping the aircraft propulsion noise generated by the rotors to benefit localization tasks, rather than considering it a harmful nuisance. We present a neural network architecture for selfnoise-based localization in a known environment. We show that training it simultaneously with learning time-varying rotor phase modulation achieves accurate and robust localization. The proposed methods are evaluated using a computationally affordable simulation of MAV rotor noise in 2D acoustic environments that is fitted to real recordings of rotor pressure fields.

- Paper: https://doi.org/10.1109/IROS58592.2024.10802115
- arXiv: https://arxiv.org/abs/2402.17289
- Code: https://github.com/tamirshor7/EARS_code

```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}
}
```

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

Authors: Shahaf Wagner*, Gabriele Serussi*, Dan Ben Ami, Tomer Galanti, Chaim Baskin
Page: https://gabrieleserussi.github.io/research/vedje/
Venue: ICML 2026 Workshop (AdaptFM), co-first author.
Result: 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.
Keywords: video-text retrieval, reranking, cross-encoder, video caching, efficient retrieval, MSR-VTT

Abstract: Finding the right video often requires distinguishing similar scenes in which different events occur. Joint matching improves retrieval, but processing rich video representations for each query is costly. VEDJE compresses features within sampled frames while keeping their representations separate in a reusable cache. Feature-change prediction supplies an auxiliary training signal that improves retrieval from the compressed cache without adding work at query time. On MSR-VTT, MSVD, DiDeMo, and ActivityNet, VEDJE improves R@1 over matched first-stage retrievers in both retrieval directions. On MSR-VTT, it reaches 59.8 text-to-video R@1 with a fine-tuned VideoCLIP-XL first stage. In the VideoPrism configuration, shrinking the per-video cache fourfold to 12 KiB preserves text-to-video recall within 0.2 points. These results show that accurate video search can operate on compact evidence, encoded once and reused as new queries arrive.

- arXiv: https://arxiv.org/abs/2610.11850
- OpenReview: https://openreview.net/forum?id=ZqB0Ab4juk
- Project: https://gabrieleserussi.github.io/vedje/
- Code: https://github.com/GabrieleSerussi/vedje

```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: Uncovering Latent Thoughts in Composed Video Retrieval

Authors: Gabriele Serussi*, David Vainshtein*, Jonathan Kouchly*, Dotan Di Castro, Chaim Baskin
Page: https://gabrieleserussi.github.io/research/pregen/
Venue: arXiv 2026, co-first author.
Result: PREGEN reaches 99.73 Recall@1 on WebVid-CoVR, 27.23 points above the previous best, by reading a frozen VLM's hidden states with a lightweight encoder for composed video retrieval.
Keywords: composed video retrieval, vision-language models, frozen VLM, hidden states, WebVid-CoVR

Abstract: Composed Video Retrieval (CoVR) aims to retrieve a video based on a query video and a modifying text. Current CoVR methods fail to fully exploit modern Vision-Language Models (VLMs), either using outdated architectures or requiring computationally expensive fine-tuning and slow caption generation. We introduce PREGEN (PRE GENeration extraction), an efficient and powerful CoVR framework that overcomes these limitations. Our approach uniquely pairs a frozen, pre-trained VLM with a lightweight encoding model, eliminating the need for any VLM fine-tuning. We feed the query video and modifying text into the VLM and extract the hidden state of the final token from each layer. A simple encoder is then trained on these pooled representations, creating a semantically rich and compact embedding for retrieval. PREGEN significantly advances the state of the art, surpassing all prior methods on standard CoVR benchmarks with substantial gains in Recall@1 of +27.23 and +69.59. Our method demonstrates robustness across different VLM backbones and exhibits strong zero-shot generalization to more complex textual modifications, highlighting its effectiveness and semantic capabilities.

- arXiv: https://arxiv.org/abs/2601.13797

```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},
}
```

## News

- 2026-10-25: Our proposal (https://gabrieleserussi.github.io/#grants) 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.
- 2026-10-14: HiMu (https://gabrieleserussi.github.io/research/himu/), a training-free frame selector for long videos, was accepted at WACV 2027.
- 2026-10-08: VEDJE (https://gabrieleserussi.github.io/research/vedje/), a cached reranker for scalable video-text retrieval, was posted to arXiv.
- 2026-07-11: Two of our papers were presented at ICML 2026 in Seoul: Structured Diffusion Bridges (https://gabrieleserussi.github.io/research/sdb/) in the main conference and VEDJE (https://gabrieleserussi.github.io/research/vedje/) at the AdaptFM workshop.
- 2026-06-05: HERBench (https://gabrieleserussi.github.io/research/herbench/) was presented at CVPR 2026 in Denver.
- 2026-05-03: Structured Diffusion Bridges (https://gabrieleserussi.github.io/research/sdb/), accepted at ICML 2026, was posted to arXiv.

## Profiles

- Google Scholar: https://scholar.google.com/citations?user=GJ19YUEAAAAJ&hl=en
- GitHub: https://github.com/GabrieleSerussi
- X: https://x.com/GSerussi
- LinkedIn: https://www.linkedin.com/in/gabriele-serussi
- ORCID: https://orcid.org/0009-0001-4280-4918
- OpenReview: https://openreview.net/profile?id=~Gabriele_Serussi1
- DBLP: https://dblp.org/pid/371/3989.html
- Semantic Scholar: https://www.semanticscholar.org/author/2288184194
- Hugging Face: https://huggingface.co/GSerussi
- alphaXiv: https://www.alphaxiv.org/@gabriele-serussi