Sjoerd van Steenkiste
sjoerdvansteenkiste at gmail dot com

I am a Research Scientist at Google Research and currently based in Mountain View. My current research focuses on:

• Methods for interacting with 4D visual scenes: learning representations that capture meaningful structure (objects, geometry, etc.) [cf. 1,2]; Controllable generation for scene editing [cf. 3,4].
• Foundations of LLMs: How data mixtures and architecture affect (pre-)training [cf. 5,6]; Understanding and improving (probabilistic) reasoning [cf. 7,8].

CV  /  Google Scholar  /  Twitter  /  Thesis

• New work on evaluating world models in LMs: Can Language Models Perform Implicit Bayesian Inference Over User Preference States? To appear at the NeurIPS System 2 Reasoning At Scale Workshop.
• New work investigating How Does Code Pretraining Affect Language Model Task Performance? To appear at the BlackboxNLP Workshop co-located with EMNLP.
• Two accepted papers at NeurIPS 2024 proposing new methods for 3D-Aware Multi-Object Scene Synthesis with Image Diffusion Models (spotlight) and for learning Scene-Grounded Video Representations by "moving off-the-grid" (spotlight).
• One EMNLP 2024 paper where we Benchmark Vision Language Models for Cultural Understanding.
• I will serve as a Senior Area Chair for ICLR 2025.
• Two accepted papers at NAACL 2024 comparing Syllogistic Reasoning in Humans and Language Models and evaluating The Impact of Depth on Compositional Generalization in Transformer Language Models.
• I will serve as an Area Chair for ICML 2024
• Two accepted papers at ICLR 2024 on Diffusion for Object-centric Representations of Scenes and on learning Dynamic Neural Scene Representations on Real-World Videos (spotlight).

For an up to date list of publications, please see my Google Scholar page.