Studies what transfers when models are trained with RL and which factors control that transfer. The paper shows that RL generalization depends on pre-training knowledge, but RL without strong prior knowledge is still useful: hard-task RL-trained models perform better than base models when enough information is provided in context, improve on domains where the model is already competent, and provide a better initialization for further staged RL training.

Introduced a multi-turn benchmark to evaluate state-tracking and cooperation in frontier models. RL-trained a 4B model on curated data, outperforming all non-reasoning baselines and performing comparably to reasoning models such as o4-mini. Released 1500+ (~90K data points) game trajectories for SFT and move-level ratings for RLVR.

Released a new dataset of 300 cultural artefacts spanning 6 cultural axes and 64 countries to measure long-tail bias in T2I models. Conducted 2,700 artifact-level surveys with culturally aligned annotators. Introduced Marginal Information Attribution, achieving a 2× improvement in Spearman correlation with human preferences over prior baselines.

Integrated Gated Linear Units into the attention module for parallel MLP-attention computation.

Designed a deep neural network to augment scarce flight telemetry for improved anomaly detection.