Towards an AI co-scientist This paper introduces an AI co-scientist, a multi-agent system built on Gemini 2.0, designed to assist researchers by generating and refining novel scientific hypotheses. The system employs a "generate, debate, and evolve" framework to iteratively
@askalphaxiv
-
AI Co-Scientists and LLM Reasoning Advances This Week
By
–
This week, AI is stepping into new dimensions—becoming co-scientists, sculpting 3D avatars, and blending cloud and on-device models into a seamless dance of creativity and efficiency. – Towards an AI co-scientist
– SWE-RL: Advancing LLM Reasoning via Reinforcement Learning -
Meta’s Llama3-SWE-RL Outperforms GPT-4o on Code Tasks
By
–
DeepSeek-R1 proved RL boosts LLM reasoning. Meta now scales it for real-world SWE tasks Llama3-SWE-RL-70B, trained with RL on GitHub issues, PRs, & code fixes Outperforms SFT on math & code reasoning Best medium-sized LLM on SWE-Bench Verified
Comparable to GPT-4o -
Fractal Generative Model Uses Divide-and-Conquer Approach
By
–
Yep, however rather than using one large autoregressive model over the entire image, the fractal generative model takes a divide-and-conquer approach. In other words, the model recursively divides the image into patches, and embeds them to form the input
sequence for a -
Fractal Generative Models: Recursive Architecture for Image Generation
By
–
What if generative models could build themselves recursively like fractals? Presenting Fractal Generative Models High quality pixel-by-pixel image generation Efficient modeling of high-dimensional data
Scalable fractal-like architecture Check out the images below -
VideoJAM: Joint Appearance-Motion Representations for Video Generation
By
–
JAM: Joint Appearance-Motion Representations for Enhanced Motion Generation in Models VideoJAM introduces a framework that improves motion coherence in video generation by learning a joint appearance-motion representation. This framework uses a modified training
-
Radiant Foam: Real-Time Differentiable Ray Tracing Technique
By
–
Radiant Foam: Real-Time Differentiable Ray Tracing Radiant Foam introduces a real-time differentiable ray tracing technique using a volumetric mesh to model light transport, offering speed and quality similar to 3D Gaussian Splatting (3DGS) without requiring special hardware.
-
LIMO: Mathematical Reasoning Emerges From Few Examples
By
–
LIMO: Less is More for Reasoning LIMO challenges the belief that complex reasoning requires vast datasets, showing that mathematical reasoning abilities can emerge with just a few examples. It achieves high accuracy on benchmarks with only 817 training samples. Problem:
-
Flow Q-Learning: Offline RL with Flow-Matching Policies
By
–
Flow Q-Learning Flow Q-Learning (FQL) is an offline reinforcement learning (RL) method that uses flow-matching policies to model complex action distributions. It avoids the challenges of iterative action generation by training an expressive one-step policy with RL, which
-
Inference-Time Compute Improves Adversarial Robustness
By
–
Trading Inference-Time Compute for Adversarial Robustness This paper explores how increasing inference-time compute improves the adversarial robustness of reasoning models (OpenAI o1-preview and o1-mini) without adversarial training. Problem: Adversarial attacks remain a major