Quant LetterNo. 21
19 items across 1 sections, as sent to readers on 16 October 2023. Paper titles open their ML-Quant page; ↗ goes to the source.
The general machine-learning papers the letter carried in 2023-25.
19 items
Superior Language Model: Mistral 7B v0.1 is a language model with 7 billion parameters that excels in reasoning, mathematics, and code generation, and has a version specifically designed to follow instructions.
200 shares3,920 citations todaySource ↗
Spatial Referring in Images: Ferret is a Multimodal Large Language Model that can understand and locate spatial references in images, outperforming other models in region-based and localization-required multimodal chatting.
52 shares604 citations todaySource ↗
Extended Context in LLMs: MemGPT is a system that manages different memory levels, providing extended context within large language models' limited context windows, enhancing document analysis and multi-session chat performance.
33 shares1,373 citations todaySource ↗
Real-Time View Synthesis: Im4D is a hybrid scene representation that combines grid-based geometry and multi-view image-based appearance for dynamic view synthesis from multi-view videos, providing high-quality rendering and efficient training.
19 shares76 citations todaySource ↗
Vision-Language Programmer: Octopus is a vision-language model that can interpret an agent's vision and textual task objectives to generate complex action sequences and executable code, showing improved decision-making in various tasks.
14 shares104 citations todaySource ↗
The article presents a theoretical framework for training large transformer models for in-context reinforcement learning, offering the first quantitative analysis of their capabilities.
10 shares90 citations todaySource ↗
The study uses a brain-inspired modular training method in machine learning to improve network performance and neuron clustering in compositional cognitive tasks.
10 shares12 citations todaySource ↗
The research proposes new reinforcement learning algorithms for matrix estimation problems with low-rank structure, offering improved performance guarantees.
8 shares9 citations todaySource ↗
The study presents a new self-supervised image pretraining method using continuous videos, showing that a single video can compete with ImageNet for various tasks.
3 shares49 citations todaySource ↗
The paper provides an analysis framework for rendering scenes from novel viewpoints in a monocular video, suggesting improvements over some scene-specific methods.
2 shares30 citations todaySource ↗
Text-based Video Editing: The article introduces FateZero, a new method for editing real-world videos using text, which outperforms previous models in maintaining video structure, motion, and frame consistency.
822 shares574 citations todaySource ↗
Enhancing Text-to-Image Models: The study suggests a two-stage process using a pretrained language model to improve image generation accuracy in diffusion models, enabling multi-round scene specification in various languages.
179 shares276 citations todaySource ↗
Hallucination Detection for LLMs: The paper presents SelfCheckGPT, a new approach for fact-checking black-box model responses without an external database, proving its superior ability to detect and rank factual and non-factual sentences.
99 shares1,236 citations todaySource ↗
The research analyzes the 'centroid' property in the context of Markov Chain Monte Carlo methods, exploring the relationship between parametric families, quasi-arithmetic means, and divergence functions.
92 shares3 citations todaySource ↗
Text-to-Video Model Benchmark: StoryBench is a new benchmark for assessing text-to-video models, offering tasks of different levels of difficulty and guidelines for human evaluation of video narratives.
62 shares28 citations todaySource ↗
Deep Residual Network Feature Learning: The research explores depthwise parametrizations in deep residual networks, pinpointing Depth-$\mu$P as the best parametrization for maximizing feature learning and diversity, but notes its limitations in deeper networks.
53 shares103 citations todaySource ↗
Sound-Video Matching: Soundify is a new system that aids video editors in synchronizing sounds with videos, proven to lessen workload and enhance usability in a human evaluation study.
48 shares29 citations todaySource ↗
Molecular Property Prediction Language Model: GPT-MolBERTa, a self-supervised language model that uses textual descriptions of molecules to predict their properties, is introduced, demonstrating high performance on various molecule property benchmarks.
38 shares28 citations todaySource ↗
Minimal Human Supervision Language Model Alignment: The paper introduces SALMON, a new method for aligning base language models with minimal human supervision using principle-following reward models, showing its superior performance on multiple benchmark datasets.
36 shares67 citations todaySource ↗