Yutaka Matsuo

Steering latent dynamics with readout feedback can unlock performance improvements in recurrent models that traditional inference methods fail to achieve.

SCSE transforms recurrent computation in Looped Transformers by ensuring that deviations from a learned anchor enhance performance without compromising stability.

Temporally centered SIGReg boosts multi-task learning performance by 1.7x, revealing critical incompatibilities in latent representation structures.

Discontinuous dynamics aren't the biggest problem for policy gradients in differentiable simulators; variance control often matters more.

Ditch the energy functions: C-voting unlocks better test-time reasoning in recurrent models by simply picking the most confident trajectory.

Forget Transformers; this new recurrent architecture learns more stable representations and generalizes better out-of-distribution by interleaving fast latent updates with slower, self-organizing observation processing.

Current multimodal LLMs struggle to count objects and ground evidence in videos longer than 30 minutes, achieving only ~25% accuracy compared to human performance on a new benchmark.

LLMs' chain-of-thought reasoning often falls apart due to factual incompleteness, with errors compounding across multiple hops, as revealed by a new multi-hop QA dataset.

Predict transformer training failures *before* you even start training, with 99.5% accuracy, using just a single forward pass.

LLMs that ace medical exams still fumble basic clinical judgment, prematurely deciding cases or abstaining unnecessarily when information is incomplete, revealing a critical gap in their real-world applicability.