Research projects and open-source teaching materials.
A controlled benchmark for multimodal rule inference and sequential action: can a model infer an unfamiliar symbolic world from examples, keep its rules straight, and use them to act? Even capable models struggle to integrate skills they can perform individually.
635 benchmark problems · 41.89% best model accuracy*
*As reported in the current public benchmark card; results depend on evaluation configuration.
An official NeurIPS tutorial connecting mathematical foundations and modern methods in Transformer positional encoding, with reference code for learning and teaching position methods.
Investigates how culturally informed design and chronemics can better support diverse crowdworker work styles.
A generative AI tool for gig workers, connecting practical system design with real workflows.