Nikita Balagansky

Aligning DLM generation with reference distributions via feature-space MMD significantly boosts performance on perplexity and reasoning tasks without retraining the entire model.

Continuous reasoning in embedding space can outperform traditional token-based methods while slashing computational costs.

Stable features in sparse autoencoders dominate predictive power, while unstable features reveal a surprising low-dimensional structure that complicates reproducibility.

Causal manipulation of TTS features enables unprecedented control over speech synthesis, transforming how we can steer voice characteristics in real-time.

Bootstrapping student policies with teacher-aligned rollouts within a KL-constrained trust region significantly boosts on-policy distillation performance in math reasoning tasks.