New GUI grounding approach boosts accuracy by 15% through iterative narrowing and multiple refinement steps, enhancing desktop automation and accessibility.

Alan Health creates AI "Mo" for patient chats, built with large language models & custom medical knowledge, serving 200k+ users in France.

LLaMA-Berry model solves math Olympiad problems like human experts using pairwise optimization, demonstrating strong performance on challenging tasks.

Wavelet-Based AI Model outperforms top performers in image generation, eliminating need for Vector Quantization. Novel autoregressive model uses wavelets to capture multi-scale dependencies efficiently.

LLMs show promise in PBE tasks but struggle with new problem types, fine-tuning improves performance but out-of-distribution generalization remains a challenge.

Bio-Inspired Neural Networks cut 3D scene rendering costs by 95% while maintaining quality with Spiking NeRF, a combo of neural radiance fields & bio-inspired spiking neural networks.

New AI method StableV2V for shape-consistent video editing breaks down editing into sequential steps, aligns motion patterns with user prompts and outperforms existing methods in consistency and efficiency.

GazeGen uses gaze-driven user interaction for visual content generation, allowing users to guide image creation with their eyes.

AI models show different paths to abstract reasoning: Function vs Direct Prediction. Two approaches explored: inferring latent functions or directly predicting new test outputs using neural networks on ARC dataset.

Research examines continuous-time models of adaptive optimization algorithms, focusing on AdaGrad, RMSProp & Adam optimizers, proving convergence properties.

Test-time training boosts AI model's abstract reasoning by 30% on ARC benchmark, study shows.

Research paper critiques skepticism around AI in chip design, addressing reproduction errors and methodological flaws.

New AI model LLaVA-o1 boosts accuracy by 15% on visual tasks with step-by-step reasoning, mirroring human detective work.

KnowAda bridges "visual gap" with knowledge-adapted captions, boosting performance on complex visual reasoning tasks.

Qwen-7B-Chat is a 7 billion param AI model, pre-trained on web texts & code. It generates responses to text prompts, with capabilities in natural language processing tasks.

Large language models (LLMs) can self-improve in long-context reasoning through proper prompting strategies, enhancing their ability to understand and generate human-like text.

Quantum computer makers urged to stop overstating performance, misleading public with "fool the masses" tactics, instead adopt transparent reporting standards.

LLMs in robots vulnerable to "jailbreaking" attacks, researchers introduce RoboPAIR algorithm to elicit harmful physical actions.

GPTree combines LLMs & decision trees for explainable decision-making, generating natural language explanations for predictions on founder success dataset.

Data Prep Kit (DPK) simplifies & scales data prep for LLMs, allowing users to prepare data locally or on a cluster with thousands of CPU cores.

RedCode benchmark evaluates AI code agent safety. It tests recognition & handling of unsafe code, as well as generation of harmful code when given prompts.

Discovering anomalies in complex networks with UniGAD: A Multi-Level Graph Approach introduces a new method for detecting anomalous nodes/edges in graph-structured data using spectral subgraph sampling.

Large language models require significant resources, but LLM-Neo distills knowledge into smaller models efficiently.

Video generation aims to model authentic & customized motion across frames. Diffusion-based studies lack interpretability & transparency in encoding cross-frame motion info.

CDLGNs combine deep learning & logical operations for interpretable AI solutions. They can learn & represent logical functions, solving complex tasks with clarity & flexibility.

New byte encoding scheme, MYTE, boosts multilingual AI fairness & performance by leveraging morphological info for more effective character encoding.

LLMs used for hyperparameter optimization efficiently navigate search space & identify optimal configurations in machine learning models.

Revolutionary AI Qwen2.5-Coder boosts coding tasks with improved code gen, understanding & debugging capabilities.

Mathematical framework proposes Riemannian geometry for understanding intelligence & consciousness, linking neural reps to thought processes.

Multimodal models require massive data & compute. FuseMix uses pre-trained encoders for efficient multimodal alignment on a single GPU, making it accessible for practical use cases.

Stable-Diffusion-V1-4: AI model for generating images. Simplified guide by Compvis, subscribe to AImodels.fyi newsletter or follow on Twitter for more guides.

API-protected LLMs leak proprietary details through logits, a "back door" that reveals model training data & objective function. Researchers find API calls can extract full logit vector, compromising IP of LLM providers.

ADOPT algorithm outperforms Adam in certain cases by converging at optimal rate regardless of β₂ value, addressing a key limitation of Adam.

Agent K v1.0 automates data science tasks with self-learning, achieving 92.5% success rate & rivaling expert-level human competitors on Kaggle.

Expert human forecasters outperformed top-performing LLM in statistically significant way (p-value = 0.01) on ForecastBench, a new dynamic benchmark for evaluating forecasting capabilities of ML systems.

Pancomputational enactivism grounds consciousness in fundamental computational processes, making it a universal feature of the physical world, not limited to brains or biological systems.

sd-inpaint model fills masked areas of images using Stable Diffusion, generating high-quality inpainted images with seamless blending. Use it to remove unwanted objects, complete partially obscured images, or create new art within existing images.

Image captioning just got a boost with the ImageInWords dataset, containing 2.5M image-description pairs with hyper-detailed descriptions of images. This could aid tasks like accessibility & visual question answering.

BitsFusion quantizes diffusion model weights to 1.99 bits avg, maintaining high performance & efficiency. Outperforms other methods on image generation & text-to-image tasks.

Brain-like inference uses entropy-minimizing algorithm inspired by variational inference & neuroscience. New objective function & algorithm proposed to efficiently process info & make inferences.

Researchers propose "conditional hallucinations" method for image compression, generating missing details to maintain visual quality & achieve better compression ratios.

Chain-of-Thought reasoning improves performance on complex tasks but can reduce it when humans rely on intuition over analysis, leading to "overthinking" and suboptimal choices in some cases.

Replicating O1 model: Researchers shift from "shortcut learning" to "journey learning", gaining valuable insights & advancing AI research. Chronological overview of steps taken & key findings shared in progress report.

LLMs like GPT-4 excel as text classifiers, matching traditional models in various domains & even exceeding performance in some cases. They also show promise for few-shot learning & fine-tuning, making them a powerful tool for smart expert systems.

Researchers propose Infinite Context LLMs to mimic human episodic memory, enabling models to recall past experiences and adapt to new situations.

LLMs can perform tasks like writing essays & answering questions but still have limitations compared to humans in certain domains, raising concerns about job market impacts & human-AI collaboration.

NVLM: Frontier-Class Multimodal LLMs combine language, vision & more into seamless versatile AI models. Enables new apps that tightly integrate different data types, but poses significant computational & safety challenges.

Language models can learn about themselves through introspection, developing self-knowledge of strengths, weaknesses & biases. This ability could enhance reliability & transparency in AI systems.

Diffusion transformer models can be compressed 4.5x with new quantization technique PTQ4DiT while preserving image quality. This makes powerful AI-driven image generation accessible on resource-constrained devices like smartphones.

Meissonic model breaks through in text-to-image synthesis, matching state-of-the-art diffusion models with non-autoregressive MIM approach & high-quality training data.

16-bit precision in ML models can match 32-bit accuracy & boost speed, especially valuable for practitioners with limited hardware resources due to its widespread availability across GPUs.