Minions: Cost-efficient Collaboration Between On-device and Cloud Language Models This paper explores a cost-efficient collaboration between small, on-device language models (LMs) and large, cloud-hosted LMs for data-intensive tasks. The MinionS protocol improves task
@askalphaxiv
-
HVI Color Space for Low-Light Image Enhancement
By
–
HVI: A New color space for Low-light Image Enhancement This paper introduces HVI, a new color space designed for Low-Light Image Enhancement (LLIE), addressing the color bias and brightness artifacts found in traditional RGB and HSV spaces. The authors also propose CIDNet, a
-
Fractal Generative Models: New Self-Similar Architecture Paradigm
By
–
Fractal Generative Models Overview: This paper introduces fractal generative models, a new paradigm where generative models are recursively structured using atomic generative modules, resulting in self-similar architectures inspired by fractals in mathematics. The framework is
-
Bi’an: Bilingual Benchmark for RAG Hallucination Detection
By
–
Bi'an: A Bilingual Benchmark and Model for Hallucination Detection in Retrieval-Augmented Generation Bi’an introduces a bilingual benchmark dataset (Bi’anBench) and lightweight judge models for hallucination detection in Retrieval-Augmented Generation (RAG). The dataset spans
-
Avat3r: High-Fidelity 3D Head Avatar Creation from Images
By
–
Avat3r: Large Animatable Gaussian Reconstruction Model for High-fidelity 3D Head Avatars Avat3r is a model for creating high-quality, animatable 3D head avatars from just a few images, eliminating the need for expensive multi-view capture setups. It leverages large
-
FFTNet: Efficient FFT-Based Alternative to Self-Attention
By
–
The FFT Strikes Back: An Efficient Alternative to Self-Attention FFTNet replaces self-attention with an efficient O(n log n) spectral filtering approach using the Fast Fourier Transform (FFT). By operating in the frequency domain, it captures long-range dependencies while a
-
SWE-RL: Enhancing LLM Reasoning Through Software Evolution
By
–
SWE-RL: Advancing LLM Reasoning via Reinforcement Learning on Open Software Evolution SWE-RL is a reinforcement learning (RL) approach that enhances LLM reasoning for software engineering by learning from open-source software evolution data. It trains Llama3-SWE-RL-70B on GitHub
-
AI Co-Scientist: Multi-Agent System for Scientific Hypothesis Generation
By
–
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
-
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