# DeepGaze3.5-VL: Modeling Scanpaths via Autoregressive Token Prediction

arXiv: https://arxiv.org/abs/2607.02083 · HTML: https://mlsanity.com/paper/2607.02083/

## Key facts

- Cited by 1 papers in the map (data through 14 Oct 2026).
- Builds on 16 of its 43 references that are in the map.
- First posted on arXiv 2 Jul 2026.

## Builds on (most cited first, 16 of 16)

- [LoRA: Low-Rank Adaptation of Large Language Models](https://mlsanity.com/paper/2106.09685/) (2021)
- [Kosmos-2: Grounding Multimodal Large Language Models to the World](https://mlsanity.com/paper/2306.14824/) (2023)
- [Shikra: Unleashing Multimodal LLM's Referential Dialogue Magic](https://mlsanity.com/paper/2306.15195/) (2023)
- [Pix2seq: A Language Modeling Framework for Object Detection](https://mlsanity.com/paper/2109.10852/) (2021)
- [Predicting the Driver's Focus of Attention: the DR(eye)VE Project](https://mlsanity.com/paper/1705.03854/) (2017)
- [Deep Gaze I: Boosting Saliency Prediction with Feature Maps Trained on ImageNet](https://mlsanity.com/paper/1411.1045/) (2014)
- [DeepGaze IIE: Calibrated prediction in and out-of-domain for state-of-the-art saliency modeling](https://mlsanity.com/paper/2105.12441/) (2021)
- [CAT2000: A Large Scale Fixation Dataset for Boosting Saliency Research](https://mlsanity.com/paper/1505.03581/) (2015)
- [EML-NET:An Expandable Multi-Layer NETwork for Saliency Prediction](https://mlsanity.com/paper/1805.01047/) (2018)
- [Gazeformer: Scalable, Effective and Fast Prediction of Goal-Directed Human Attention](https://mlsanity.com/paper/2303.15274/) (2023)
- [TempSAL -- Uncovering Temporal Information for Deep Saliency Prediction](https://mlsanity.com/paper/2301.02315/) (2023)
- [State-of-the-Art in Human Scanpath Prediction](https://mlsanity.com/paper/2102.12239/) (2021)
- [GazeXplain: Learning to Predict Natural Language Explanations of Visual Scanpaths](https://mlsanity.com/paper/2408.02788/) (2024)
- [Modeling Human Gaze Behavior with Diffusion Models for Unified Scanpath Prediction](https://mlsanity.com/paper/2507.23021/) (2025)
- [Few-shot Personalized Scanpath Prediction](https://mlsanity.com/paper/2504.05499/) (2025)
- [Modeling Saliency Dataset Bias](https://mlsanity.com/paper/2505.10169/) (2025)

## Built on by (most cited first, 1 of 1)

- [Moving Forward with Video Saliency: A New Dataset and Benchmark where Motion Matters](https://mlsanity.com/paper/2610.03276/) (2026)

## Abstract (arXiv)

Understanding human visual attention on a scene over time has applications in domains such as interface design and inferring cognitive states. Modeling visual scanpaths has historically relied on specialized architectures with hand-crafted priors. While these architectures can model fixation sequences, their rigid structural biases restrict easy extendability and flexible conditioning. For instance, integrating task-specific instructions or adapting to distinct viewer identities requires custom, disjoint architectural additions. We frame scanpath prediction purely as a discrete sequence modeling task. By mapping coordinates into a text vocabulary, we leverage the pretrained representations of Vision-Language Models. This framing absorbs diverse factors of variation: simple prompting allows for global conditioning, such as providing viewer identities to capture personalized biases, or task-specific objectives like visual search. The framework can also integrate per-fixation attributes, such as individual fixation durations, alongside spatial locations. The autoregressive alignment enables the scalable, exact computation of per-fixation log-likelihoods, directly equivalent to the commonly used Information Gain (IG) metric. Our model, DeepGaze3.5-VL, establishes a new state-of-the-art across multiple datasets, achieving 2.18 bits of IG on MIT1003, a 46% improvement over DeepGaze III. This advantage persists even when baselines use identical high-capacity vision encoders. Beyond predictive performance, our generative framework serves as a powerful computational tool for direct behavioral interventions, allowing for controlled in-silico simulations that would be experimentally difficult or impossible to conduct in vivo. We demonstrate this ability by performing controlled interventions on the durations of pre-saccadic fixations, recovering known oculomotor phenomena purely from data.