Main Content

• Gemini 2.5 Computer Use model
  The Gemini 2.5 Computer Use model is now available via the Gemini API, enabling developers to create agents that interact with user interfaces, outperforming competitors in web and mobile tasks with lower latency.

• 2025 Nobel Prize in Physics
  The 2025 Nobel Prize in Physics was awarded to John Clarke, Michel H. Devoret, and John M. Martinis for their groundbreaking experiments demonstrating macroscopic quantum tunnelling and energy quantisation in superconducting circuits, revealing quantum properties on a human scale.

• Tiny Recursion Model (TRM)
  The Tiny Recursion Model (TRM) achieves 45% accuracy on ARC-AGI-1 and 8% on ARC-AGI-2 with only 7M parameters, demonstrating that smaller models can effectively tackle complex reasoning tasks without the need for extensive resources.

• Gradient descent
  Gradient descent operates effectively in deep learning by navigating regions of weight space where the sharpness (S(w)) remains below the critical threshold (2/\eta), despite oscillating dynamics that suggest instability.

• New probabilistic patterns in prime numbers
  Mathematicians have uncovered new probabilistic patterns in prime numbers, suggesting a deeper order within their seemingly chaotic distribution, linked to the Riemann zeta function and its zeros.

• Bharath Ramsundar's PhD journey
  Bharath Ramsundar's PhD journey illustrates the evolution of his research in machine learning and computational biology, culminating in the development of DeepChem, an open-source package aimed at enhancing drug discovery processes.

• AgentFlow
  AgentFlow introduces a trainable, in-the-flow agentic framework that optimizes planning through a coordinated system of four specialized modules, enhancing performance in multi-turn interactions.

• Reactive Transformer (RxT)
  The Reactive Transformer (RxT) introduces a novel architecture that shifts from a data-driven to an event-driven paradigm, enabling real-time processing of conversational turns while maintaining context in a fixed-size Short-Term Memory (STM) system.

• Generative models
  Generative models, particularly flow models, can be enhanced by integrating exogenous inputs through predictive control algorithms like MPC and MPPI to guide image generation from Gaussian noise.

• Lion Schedule-Free optimizer
  The Lion Schedule-Free optimizer eliminates the need for learning-rate schedulers, achieving 3x faster convergence by utilizing sign agreement to scale updates, enhancing training stability and allowing for hot restarts.

• MADPO
  MADPO offers a granular solution to the data quality issues faced by DPO and β-DPO, enhancing performance by assigning unique weights to samples based on difficulty, leading to improved outcomes in sentiment generation tasks.

• Distributional Semantics Tracing (DST)
  The proposed Distributional Semantics Tracing (DST) framework offers a novel approach to understanding hallucinations in Large Language Models (LLMs) by mapping internal reasoning processes and contextual meanings, enhancing interpretability.

• How AI-native 6G networks will revolutionize telecommunications
  AI-native 6G networks will revolutionize telecommunications by enabling real-time AI traffic management, supporting applications like autonomous vehicles and smart agriculture, thus positioning the U.S. as a leader in the global AI economy.

• Long-context models (LCMs)
  Long-context models (LCMs) excel in processing lengthy sequences but are vulnerable to contextual noise, which can mislead attention and predictions; our study introduces the Integrated Gradient (IG) score to effectively detect this noise.