The paper introduces Voxtral Realtime, a novel automatic speech recognition (ASR) model designed for native streaming with sub-second latency. Unlike chunking-based approaches, Voxtral Realtime is trained end-to-end for streaming with explicit audio-text alignment, leveraging the Delayed Streams Modeling framework. The model incorporates a new causal audio encoder and Ada RMS-Norm for improved delay conditioning, and achieves performance comparable to Whisper at a 480ms delay after large-scale pretraining across 13 languages.

This paper generalizes the connection between Direct Preference Optimization (DPO) and human choice theory, extending the normative framework underlying DPO. By reworking the standard human choice theory, the authors demonstrate that any compliant machine learning analytical choice model can be embedded within any human choice model. This generalization supports non-convex losses and provides a unifying framework for various DPO extensions like margins and length correction.

This paper introduces semi-nonnegative matrix factorization (SNMF) to decompose MLP activations in LLMs into interpretable features, addressing limitations of sparse autoencoders (SAEs) in mechanistic interpretability. SNMF learns sparse linear combinations of co-activated neurons and maps them to activating inputs, enhancing interpretability and enabling causal steering. Experiments on Llama 3.1, Gemma 2, and GPT-2 demonstrate that SNMF-derived features outperform SAEs and a supervised baseline in causal steering while aligning with human-interpretable concepts, revealing a hierarchical structure in MLP activation space.