S. Lahlou

Most self-evolving LLM updates are useless, yet current methods inherit them all; new frameworks drastically cut harmful updates and preserve valuable experience.

A frozen video tokenizer's latent can be factored into controllable and reward-relevant components, unlocking significant downstream planning gains.

Compact retrievers can achieve nearly half the performance of larger models while being up to 9.7× faster, challenging the assumption that bigger is always better in retrieval tasks.

Arabic LLMs can speak the language of finance, but they often fail to reason about it, especially when it comes to causality and generation.

Forget retraining: HyperKKL uses a hypernetwork to dynamically condition KKL observers on exogenous inputs, enabling accurate state estimation for non-autonomous nonlinear systems.