MindAgent: Emergent Gaming Interaction
Ran Gong, Qiuyuan Huang, Xiaojian Ma +8 authors
MindAgent infrastructure evaluates multi-agent coordination and collaboration using LLMs in gaming scenarios, introducing CUISINEWORLD and CoS metric.
Explore · 每周精选
发现最受关注的研究论文,追踪研究趋势,订阅感兴趣的期刊与关键词。
50 篇论文 · 按点赞排序
Boolformer: Symbolic Regression of Logic Functions with Transformers
Stéphane d'Ascoli, Samy Bengio, Josh Susskind +1 authors
Boolformer, a Transformer architecture, achieves competitive symbolic regression of Boolean functions on real-world datasets and gene regulatory networks with significant speed improvements over genetic algorithms.
BTLM-3B-8K: 7B Parameter Performance in a 3B Parameter Model
Nolan Dey, Daria Soboleva, Faisal Al-Khateeb +11 authors
BTLM-3B-8K, a state-of-the-art 3 billion parameter open-source language model, outperforms existing 3B models and rivals 7B models, offering excellent long context performance and reduced memory and inference compute requirements.
Struc-Bench: Are Large Language Models Really Good at Generating Complex Structured Data?
Xiangru Tang, Yiming Zong, Yilun Zhao +2 authors
Structure-aware fine-tuning improves Large Language Models' ability to generate complex structured data by reducing formatting errors and enhancing adherence to natural language constraints.
SlimPajama-DC: Understanding Data Combinations for LLM Training
Zhiqiang Shen, Tianhua Tao, Liqun Ma +5 authors
Research on SlimPajama-DC reveals the impact of various data deduplication strategies and proportions on large language model training, showing improved performance with optimal data diversity.
Stabilizing RLHF through Advantage Model and Selective Rehearsal
Baolin Peng, Linfeng Song, Ye Tian +3 authors
Two methods, Advantage Model and Selective Rehearsal, are introduced to enhance the stability and performance of RLHF training for Large Language Models.
Compositional Foundation Models for Hierarchical Planning
Anurag Ajay, Seungwook Han, Yilun Du +7 authors
HiP, a compositional foundation model, integrates language, vision, and action models to plan and execute complex, long-horizon manipulation tasks through hierarchical reasoning and iterative refinement.
A Large-scale Dataset for Audio-Language Representation Learning
Luoyi Sun, Xuenan Xu, Mengyue Wu +1 authors
An automatic pipeline creates a large audio-text dataset, Auto-ACD, improving performance on audio-language retrieval, captioning, and classification tasks.
End-to-End Speech Recognition Contextualization with Large Language Models
Egor Lakomkin, Chunyang Wu, Yassir Fathullah +3 authors
A novel method using LLMs for contextualizing speech recognition models achieves significant performance improvements with minimal additional parameters.
FoleyGen: Visually-Guided Audio Generation
Xinhao Mei, Varun Nagaraja, Gael Le Lan +4 authors
FoleyGen, a V2A system employing a Transformer model with novel visual attention mechanisms, achieves superior performance by leveraging bidirectional neural audio codecs and various pretrained visual encoders.
Language Embedded Radiance Fields for Zero-Shot Task-Oriented Grasping
Adam Rashid, Satvik Sharma, Chung Min Kim +4 authors
LERF-TOGO uses vision-language models and DINO features for zero-shot task-oriented grasping by extracting object masks and ranking grasps over specific parts.
TextBind: Multi-turn Interleaved Multimodal Instruction-following
Huayang Li, Siheng Li, Deng Cai +5 authors
TextBind is an annotation-free framework that enhances large language models for multimodal instruction following by generating conversations from image-caption pairs.
Leveraging Contextual Information for Effective Entity Salience Detection
Rajarshi Bhowmik, Marco Ponza, Atharva Tendle +5 authors
Fine-tuning medium-sized language models with a cross-encoder style significantly improves salient entity detection compared to feature engineering approaches.
A Distributed Data-Parallel PyTorch Implementation of the Distributed Shampoo Optimizer for Training Neural Networks At-Scale
Hao-Jun Michael Shi, Tsung-Hsien Lee, Shintaro Iwasaki +5 authors
Shampoo, an AdaGrad-based method using block-diagonal preconditioners with Kronecker product approximations, enhances neural network training performance in PyTorch, especially distributed multi-GPU environments.
Retrieval-Augmented Text-to-Audio Generation
Yi Yuan, Haohe Liu, Xubo Liu +3 authors
A retrieval-augmented approach enhances text-to-audio generation by addressing class imbalance and improving generation performance for rare audio classes.
A Data Source for Reasoning Embodied Agents
Jack Lanchantin, Sainbayar Sukhbaatar, Gabriel Synnaeve +3 authors
A new data generator for machine reasoning is introduced, integrating with an embodied agent, and tests baseline models including pre-trained language models and graph-structured Transformers on a knowledge-graph representation of the database.
360^circ Reconstruction From a Single Image Using Space Carved Outpainting
Nuri Ryu, Minsu Gong, Geonung Kim +2 authors
POP3D is a framework that creates a 360-degree 3D model from a single image using a combination of depth prediction, space carving, generative modeling, and neural implicit surfaces.
The Languini Kitchen: Enabling Language Modelling Research at Different Scales of Compute
Aleksandar Stanić, Dylan Ashley, Oleg Serikov +5 authors
A protocol for comparing language models based on equivalent computational resources evaluates GPT-2 and a new high-throughput LSTM, showing predictable scaling and intersection in performance at 50,000 accelerator hours.
Stack-and-Delay: a new codebook pattern for music generation
Gael Le Lan, Varun Nagaraja, Ernie Chang +5 authors
A new stack-and-delay decoding strategy for hierarchical token stacks in music generation improves inference speed without significantly compromising quality.
S3-DST: Structured Open-Domain Dialogue Segmentation and State Tracking in the Era of LLMs
Sarkar Snigdha Sarathi Das, Chirag Shah, Mengting Wan +5 authors
S3-DST, a structured prompting technique with Pre-Analytical Recollection, enhances joint dialogue segmentation and state tracking in open-domain dialogs handled by LLM-based systems.