Fernando Acero

Foundation models trained on synthetic order flow perform amortized Bayesian inference, decoding latent market regimes with surprising accuracy and outperforming traditional methods for simulator calibration.

LLMs aren't always necessary: simple text embeddings in numerical bandits can match or beat LLM-based contextual bandits at a fraction of the cost.

Smaller language models can now achieve state-of-the-art claim verification performance by jointly optimizing for decomposition quality and verification accuracy using a novel reinforcement learning approach.