Publications
-
SQuID: A Benchmark for Quantitative Spatial Reasoning on Satellite Imagery
SQuID, a benchmark of 2,000 quantitative questions on satellite images with tolerance ranges informed by human annotations. The best zero-shot end-to-end VLM reaches 28.1%, and the QASR baseline reaches 42.0%.
Preprint, arXiv:2601.13401, 2026 · arXiv → · Dataset → · Code →
BibTeX
@misc{massih2026squid, title = {{SQuID}: A Benchmark for Quantitative Spatial Reasoning on Satellite Imagery}, author = {Massih, Peter A. and Cosatto, Eric and Min, Martin Renqiang}, year = {2026}, eprint = {2601.13401}, archivePrefix = {arXiv} } -
What Is a Visual Thought Worth? When Continuous Visual States Help Vision-Language Models Reason
Removes the image and keeps the latent visual states of CoVT and LVR fixed, to test whether the states can be decoded, are used, and add value beyond image features.
Under review at ICLR 2027
-
Test-Time Reasoning with Computational Evidence for Spatial Question Answering
A training-free framework that plans question-specific geometric computations and gives the VLM explicit computational evidence for multi-hop spatial questions.
Under review at ICLR 2027
-
Three Modalities of Production AI in Financial Services: Retrieval, Generation, and Detection
Three AI systems deployed under FINMA rules and used daily by bankers, fund screeners and tax specialists.
Master's thesis, EPFL, 2026 · DOI → · PDF →
BibTeX
@mastersthesis{massih2026thesis, title = {Three Modalities of Production AI in Financial Services: Retrieval, Generation, and Detection}, author = {Massih, Peter A.}, school = {EPFL}, year = {2026}, doi = {10.5281/zenodo.21074981} }