Research Briefing — February 13, 2026

Introduces Voxtral Realtime, a natively streaming ASR model from Mistral that matches offline transcription quality (on par with Whisper) at sub-second latency (~480ms). Uses a new causal audio encoder and Ada RMS-Norm within a Delayed Streams Modeling framework, pretrained on 13 languages. Weights released under Apache 2.

Provides a mathematical framework for the linear representation hypothesis (LRH) in language models, proving that O(m^(4/3)) neurons suffice to linearly represent and access m features, with a near-matching lower bound.

This paper investigates test-time verification as a way to close the gap between intended instructions and generated actions in Vision-Language-Action (VLA) models for robotics. They characterize test-time scaling laws for embodied instruction following and show that jointly scaling rephrased instructions and generated actions greatly increases sample diversity. Notable authors include Chelsea Finn, Marco Pavone, and Azalia Mirhoseini.

Retrieval-aware distillation converts a pretrained Transformer into a hybrid Transformer-SSM model by preserving only 2% of attention heads (retrieval-critical 'Gather-and-Aggregate' heads) and distilling the rest into recurrent heads, recovering 95%+ performance on retrieval tasks.

Proposes Latent Forcing, which achieves efficiency of latent diffusion while operating on raw pixels by jointly processing latents and pixels with separate noise schedules. The latents serve as a scratchpad before high-frequency pixel features are generated. From Fei-Fei Li's group at Stanford.

Reveals that audio-LLMs like Gemini 2.0 Flash exhibit strong text dominance bias, following text 10x more often than audio when the two conflict, even when explicitly instructed to trust audio. Proposes that this reflects an asymmetry in arbitration accessibility rather than information quality.

Presents a global longitudinal analysis showing AI is creating a new linguistic hierarchy, with benefits concentrated in a small number of languages while most of the world's 7,000+ linguistic communities face digital marginalization.

Proposes C-JEPA, an object-centric world model extending JEPA to object-level representations with masking that induces latent interventions for causal reasoning. Shows gains in visual QA tasks. Authors include Yann LeCun and Randall Balestriero.

Introduces Gaia2, a benchmark for evaluating LLM agents in dynamic, asynchronous environments where environments evolve independently of agent actions. Includes write-action verifiers for RL training. Tests GPT-5 and other SOTA models, finding no model dominates across all capabilities.

Discovers a striking dissociation in transformer representations: angular (direction) perturbations damage language modeling while magnitude perturbations disproportionately damage syntactic processing, revealing distinct functional roles of direction and magnitude in Pythia models.