Firehose

This paper introduces AREX, a self-improving agent that recursively refines its answers by verifying intermediate results and using the partially verified state to guide subsequent refinement, aiming to improve the efficiency of deep research. Practitioners might care about AREX's approach to reducing the cost of discovery and verification in complex research tasks.

This paper introduces a new framework for building AI agents in Python, allowing developers to write agents that are both deterministic and model-agnostic. Practitioners might care because this framework could simplify the development of reliable AI agents for various applications.

This paper introduces Robostral Navigate, a vision-language model that enables robots to navigate using only a single monocular RGB camera, making it more scalable and cost-effective for deployment across various robotic platforms. Practitioners might care about this because it can simplify navigation tasks for robots in real-world environments.

This paper develops a new framework for robotic grasping in complex scenes that uses language to specify task requirements, allowing robots to grasp objects with more precision and flexibility. Practitioners working on robotic grasping and manipulation might care about this approach because it could improve the performance and efficiency of robots in real-world applications.

This paper introduces a new method, Surrogate Latent Policy Optimization (SLPO), to improve the performance of autoregressive latent reasoners in reinforcement learning by combining outcome-reward RL with latent reasoning. Practitioners might care because SLPO can help scale the performance of these models at test time, which is currently a major challenge.

This paper introduces a new environment called Trace, which allows vision-language models to reason across multiple domains and tasks, using a taxonomy-guided approach. Practitioners might care because this work could lead to more generalizable and transferable AI models.

Simulation for physical AI systems relies on generating large amounts of physically grounded data, which is challenging to collect in the real world due to safety, cost, and practicality concerns. Simulation engines like MuJoCo, Isaac Sim, and others can generate photorealistic data using GPU parallelism, enabling developers to train reinforcement learning policies and test policies against rare scenarios. The choice of simulation engine depends on factors such as scalability, sensor support, 3D asset formats, and environmental fidelity required for the specific use case. AI summary

This paper develops a new optimization framework called Isospectral Optimization (ISO) that improves the performance of language models trained using reinforcement learning with verifiable rewards (RLVR) by reusing the base model's weight spectra while adapting the input and output singular frames. Practitioners might care about this because it could lead to faster and more accurate training of large language models.

This paper introduces H^2SD, a hybrid hindsight self-distillation framework for reinforcement learning with verifiable rewards, which combines the strengths of different methods to improve large language models' reasoning capabilities. Practitioners may care about this work because it addresses limitations of existing methods and shows promising results on challenging reasoning benchmarks.

This paper proposes a new framework for video spatial reasoning that focuses on learning geometric consistency to improve the accuracy and stability of models in tasks like navigation and question answering. Practitioners working on multimodal models and spatial reasoning tasks may benefit from this approach.

This paper proposes a new way for large language models to learn from feedback, allowing them to retain more detailed information about the quality of their responses and learn from it in a more nuanced way. Practitioners might care because this approach could lead to better performance on tasks where the model doesn't have a clear way to evaluate its own output.

Researchers have developed a way to control the effort mode of large language models (LLMs), allowing them to switch between low-effort, medium-effort, and high-effort reasoning modes. This is achieved through training using reinforcement learning with verifiable rewards (RLVR), which provides a reward signal for correct and incorrect responses. AI summary

This episode features Andy Beam and Rafa Gómez-Bombarelli from Lila Sciences, discussing their vision for AI science factories as the next frontier for generating internet-scale datasets. They explain how their automated labs, leveraging AI…

Alistair Pullen, CEO of Cosine, discusses the UK's initiative to build a sovereign large language model (LLM) in response to US export controls on frontier AI like Fable. He explains Cosine's strategy to compete with larger labs by focusing…

David Dalrymple, known as Davidad, discusses his shift from formal verification approaches to an 'Alignment with Awakening' framework, emphasizing the formation of coalitions of aligned AIs that recognize shared moral truths. He shares empi…

Tim Scarfe interviews the Tufa Labs ARC-AGI-3 team to dissect their winning approach on the ARC-AGI-3 benchmark, focusing on how their system discovers goals and balances exploration with action efficiency. The episode explores the challeng…

In this episode of Latent Space, Evan Feinberg and Sergey Edunov of Genesis Molecular AI discuss their pioneering work in applying diffusion models to protein-small molecule interactions for drug discovery. They explain how their foundation…

This episode features Thomas von Tschammer of Neural Concept, discussing how physics-aware AI is revolutionizing product engineering. Neural Concept's models accelerate design evaluation from days to minutes, enabling companies like Jaguar …

This episode argues that current AI progress is primarily driven by an immense quantity of high-quality, task-specific data, rather than improvements in sample efficiency. The speaker highlights the vast data requirements of frontier models…

Carina Hong, CEO of Axiom Math, discusses the company's recent $200M Series A funding and their perfect Putnam exam score, highlighting their mission to scale "verified AI" through formal mathematics. She explains how formal verification, u…

Eric Jang explains how to build AlphaGo from scratch using modern AI tools, detailing the game of Go's rules and the core Monte Carlo Tree Search (MCTS) algorithm. He describes how deep neural networks, specifically value and policy network…

Kyle Corbitt, founder of OpenPipe and leader of CoreWeave's serverless training team, provides a master class on reinforcement learning (RL) and custom fine-tuning for AI models. He explains how RL differs from supervised fine-tuning (SFT) …

This episode features Qasar Younis and Peter Ludwig, co-founders of Applied Intuition, discussing their company's mission to build physical AI for various moving systems like cars, trucks, and mining equipment. They delve into the evolution…