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
Wangchunshu Zhou, Yuchen Eleanor Jiang, Peng Cui +5 authors
RecurrentGPT combines the long-term memory mechanism of RNNs with large language models to generate arbitrarily long, interpretable text with interactive capabilities.
Soham Deshmukh, Benjamin Elizalde, Rita Singh +1 authors
Pengi, an audio language model using transfer learning, achieves state-of-the-art performance across various audio tasks by framing them as text-generation tasks.
Andrea Schioppa, Xavier Garcia, Orhan Firat
Combining self-supervised and supervised objectives during pre-training of Large Language Models improves their in-context learning by incorporating cross-lingual parallel data.
Mustafa Safa Ozdayi, Charith Peris, Jack FitzGerald +5 authors
Prompt-tuning can control the extraction rates of memorized content in large language models, offering both attack and defense mechanisms to manage privacy risks.
Liting Chen, Lu Wang, Hang Dong +9 authors
Introspective Tips improve LLM decision-making in few-shot and zero-shot learning without parameter fine-tuning by enhancing prompt design.
Emily Reif, Minsuk Kahng, Savvas Petridis
LinguisticLens is an interactive tool for visualizing and analyzing syntactic diversity in datasets generated by large language models.
Yue Wu, So Yeon Min, Shrimai Prabhumoye +5 authors
SPRING framework uses LLMs and DAGs to reason and play open-world games by processing game prompts, outperforming RL baselines without training.
Vikas Raunak, Amr Sharaf, Hany Hassan Awadallah +1 authors
GPT-4 demonstrates superior performance in automatic post-editing of NMT outputs across multiple language pairs, achieving state-of-the-art results.
Simeng Sun, Yang Liu, Shuohang Wang +2 authors
PEARL is a prompting framework that decomposes complex questions into a sequence of actions and executes them using LLMs to reason over long documents, outperforming existing methods.
Viorica Pătrăucean, Lucas Smaira, Ankush Gupta +21 authors
The Perception Test evaluates multimodal models' perceptual and reasoning skills across video, audio, and text through real-world videos and diverse annotations.
Sungdong Kim, Sanghwan Bae, Jamin Shin +4 authors
A novel framework achieves LLM alignment with minimal human labor by using synthetic feedback for reward modeling and reinforcement learning, outperforming models trained on human-annotated data.
Cédric Colas, Laetitia Teodorescu, Pierre-Yves Oudeyer +2 authors
A language model augmented autotelic agent learns diverse, abstract goals through cultural transmission without predefined representations or curricula.
Badr AlKhamissi, Siddharth Verma, Ping Yu +3 authors
Fine-tuned OPT models with and without explanations were evaluated on a wide range of reasoning tasks, revealing that explanations enhance performance in non-finetuned models and specific reasoning skills.
Sebastian Ruder, Jonathan H. Clark, Alexander Gutkin +24 authors
XTREME-UP is a benchmark evaluating language models across 88 under-represented languages on user-centric tasks with limited data, focusing on ASR, OCR, MT, and information access technologies.
Chaitanya Malaviya, Peter Shaw, Ming-Wei Chang +2 authors
A dataset named QUEST challenges retrieval systems by testing their ability to interpret natural language queries with implicit set operations and cross-reference them with Wikipedia documents.
Richard Antonello, Aditya Vaidya, Alexander G. Huth
Larger transformer-based language models like OPT and LLaMA improve brain response prediction accuracy compared to smaller models, scaling log-linearly with both model size and fMRI training set size.
Akshita Jha, Aida Davani, Chandan K. Reddy +3 authors
SeeGULL is a globally diverse stereotype dataset built using large language models, spanning 178 countries and including offensiveness scores and regional comparisons.
Zezhong Wang, Fangkai Yang, Pu Zhao +5 authors
A benchmark QA dataset for Microsoft products and IT issues is introduced, along with a model fusion framework enhancing domain-specific LLM performance.
Lyle Regenwetter, Yazan Abu Obaideh, Faez Ahmed
MCD is a multi-objective counterfactual optimization method for design problems that enhances efficiency and provides recommendations for better functional performance.
Eliza Kosoy, Emily Rose Reagan, Leslie Lai +2 authors
Classical developmental psychology experiments adapted to evaluate AI models reveal that while language models can mimic social understanding, they struggle with early object and action understanding, theory of mind, and causal reasoning, suggesting the need for real-world experience.
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