# Oct 13, 2025

**Breakthrough in photonics**: Researchers at USC have developed the first optical device utilizing **optical thermodynamics**, allowing light to route itself without switches or external control, guided by thermodynamic principles.

**ATLAS achieves up to 4x faster LLM inference** by utilizing adaptive learning to optimize speculative decoding in real-time, enhancing performance as usage increases.

**LaTeXpOsEd** reveals significant **security risks** in preprint archives, exposing sensitive information through a systematic analysis of over **1.2 TB** of data from **100,000 arXiv submissions**.

**Aidlab** offers developers access to **gold-standard physiological data** through a free SDK, enabling real-time streaming of 13 data types, including raw ECG and motion, across multiple platforms like Unity and Flutter.

**oavif** introduces a **target quality encoding** framework for image compression, optimizing speed and consistency through advanced metrics and convergence algorithms, significantly enhancing user experience in image quality management.

**StreamingVLM** introduces a novel approach for real-time understanding of **infinite video streams**, effectively addressing the challenges of **latency** and **memory usage** through a unified framework that aligns training with streaming inference.

**NVIDIA's Vera Rubin NVL144** is set to revolutionize AI factories with **800 VDC data centers**, supported by over 50 partners, enhancing scalability and energy efficiency for future demands in AI processing.

**GraphMERT** is a novel graphical encoder-only model that distills **high-quality knowledge graphs (KGs)** from unstructured text, achieving state-of-the-art accuracy and superior symbolic representations compared to existing frameworks.

The **Sandwiched Policy Gradient (SPG)** method enhances the alignment of **diffusion large language models (dLLMs)** with human preferences by utilizing both upper and lower bounds of the true log-likelihood, addressing the limitations of traditional policy gradient methods.

**Eigenvalues** play a crucial role in understanding the performance of sequence models, revealing how they influence **memory** and **long-range dependency** modeling across various architectures.

The **locality dial** is a groundbreaking feature that allows large language models to adjust their internal representations dynamically, balancing between **localist** and **distributed** encodings without the need for retraining.

The **Fine-grained Low-Rank Compressor (FLRC)** optimally allocates ranks per layer and employs **progressive low-rank decoding**, enhancing text generation quality in large language models (LLMs).

**CLARity** is a novel RL framework that enhances reasoning quality in expert LLMs using a **cost-effective** approach, relying solely on a small, general-purpose LLM to improve logical consistency without the need for expensive Process Reward Models.

**DICE** enhances **LLM** outputs by utilizing **SLM-guided chain-of-thought correction**, effectively refining responses to meet specific user requirements without the need for extensive fine-tuning.