Trends · 研究趋势
数据来自 Hugging Face 论文的 AI 提取关键词,按月统计研究方向的增长与热度。
Jingqi Tong, Yurong Mou, Hangcheng Li +11 authors
The "Thinking with Video" paradigm enhances multimodal reasoning by integrating video generation models, demonstrated through the Video Thinking Benchmark and improved performance on both vision and text tasks.
Wei Huang, Yi Ge, Shuai Yang +11 authors
QeRL, a quantization-enhanced reinforcement learning framework, accelerates RL training for large language models by combining NVFP4 quantization with Low-Rank Adaptation and an Adaptive Quantization Noise mechanism, achieving significant speedups and improved performance.
An Yang, Baosong Yang, Binyuan Hui +55 authors
The Qwen2 series, comprising 0.5 to 72 billion parameter models, surpasses prior open models across language understanding, generation, multilingualism, coding, math, and reasoning, with exceptional performance in benchmarks like MMLU, GPQA, HumanEval, GSM8K, BBH, MT-Bench, Arena-Hard, and LiveCodeBench.
Jin Xu, Zhifang Guo, Jinzheng He +11 authors
Qwen2.5-Omni is a multimodal model that processes text, images, audio, and video in a streaming fashion and generates text and speech using a dual-track architecture, achieving state-of-the-art performance on multimodal benchmarks.
Mohamad Zbib, Mohamad Bazzi, Ammar Mohanna +2 authors
Speculative decoding effectiveness depends on draft model training data alignment with downstream tasks, with specialized drafters performing better when combined through confidence-based routing rather than simple averaging.
Eric Zelikman, Georges Harik, Yijia Shao +3 authors
Quiet-STaR is an extension of the Self-Taught Reasoner that teaches language models to infer and generate rationales for each token, improving their ability to predict difficult tokens and answer questions without task-specific fine-tuning.
Chengrun Yang, Xuezhi Wang, Yifeng Lu +4 authors
OPRO, a method using large language models to optimize tasks described in natural language, outperforms human-designed prompts on various benchmark datasets.
Alexander Amini, Anna Banaszak, Harold Benoit +30 authors
LFM2, a family of compact foundation models, achieves high efficiency and performance on-device through hardware-in-the-loop architecture search and advanced training techniques, supporting various tasks including multimodal applications.
Zicheng Lin, Tian Liang, Jiahao Xu +6 authors
The proposed cDPO approach identifies and rewards critical tokens to improve the reasoning accuracy of LLMs by utilizing contrastive estimation between positive and negative models.
Arindam Mitra, Luciano Del Corro, Guoqing Zheng +11 authors
AgentInstruct, an agentic framework, generates high-quality synthetic data for post-training language models, leading to significant performance improvements across various benchmarks.
Richard Yuanzhe Pang, Weizhe Yuan, Kyunghyun Cho +3 authors
An iterative preference optimization method using a modified DPO loss improves reasoning accuracy on various datasets by optimizing winning and losing reasoning steps in Chain-of-Thought candidates.
Sumeet Ramesh Motwani, Chandler Smith, Rocktim Jyoti Das +6 authors
Multi-agent LLM training improves performance on reasoning tasks by assigning specialized roles and utilizing joint outcome-based rewards to enhance collaboration among models.
Xin Lai, Zhuotao Tian, Yukang Chen +3 authors
A novel method called Step-DPO enhances LLMs' mathematical reasoning by optimizing individual reasoning steps, achieving significant accuracy improvements with minimal data and training.
Aitor Ormazabal, Che Zheng, Cyprien de Masson d'Autume +22 authors
Reka Core, Flash, and Edge are multimodal language models that achieve state-of-the-art performance across text, image, video, and audio tasks, offering high efficiency and competitive results relative to their compute class.
Sean O'Brien, Mike Lewis
Contrastive Decoding, a simple and training-free text generation method, outperforms other decoding techniques on various reasoning tasks and logical benchmarks by improving perceived quality and reducing reasoning errors.
Shubham Toshniwal, Ivan Moshkov, Sean Narenthiran +3 authors
A new math instruction tuning dataset, OpenMathInstruct-1, leverages the Mixtral model to achieve competitive performance on math reasoning benchmarks using open-source LLMs.
Haolin Chen, Yihao Feng, Zuxin Liu +8 authors
LaTent Reasoning Optimization (LaTRO) enhances LLMs' reasoning capabilities through variational optimization of latent distributions, improving zero-shot accuracy on complex reasoning tasks.
Hugh Zhang, Jeff Da, Dean Lee +12 authors
Evaluation of large language models on a new math benchmark, GSM1k, reveals that many models exhibit signs of overfitting to the existing GSM8k benchmark, leading to performance drops on GSM1k.
Shengnan An, Zexiong Ma, Zeqi Lin +3 authors
LeMa, a learning-from-mistakes approach, enhances LLMs' mathematical reasoning by learning from inaccurate reasoning paths corrected by GPT-4, surpassing SOTA performance on math problems.
Ning Miao, Yee Whye Teh, Tom Rainforth
A zero-shot verification scheme improves LLMs' question-answering performance by identifying and correcting errors in step-by-step reasoning.
北京市昌平区探索星信息技术及软件开发工作室
京ICP备2026059466号