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。
48 篇论文 · 按点赞排序
André Storhaug, Jingyue Li
Research investigates the effectiveness of parameter-efficient fine-tuning methods compared to full fine-tuning for generating unit tests using large language models.
Yanting Chen, Yi Ren, Xiaoting Qin +7 authors
VLM-based methods, Direct Frame-Based Approach and Differential Frame-Based Approach, for extracting user actions from desktop recordings are proposed and evaluated.
Wenyue Hua, Ollie Liu, Lingyao Li +9 authors
Game-theoretic workflows enhance the rationality of large language models in strategic decision-making tasks and negotiations, improving their performance and reducing susceptibility to exploitation.
Harry Mayne, Yushi Yang, Adam Mahdi
Investigation into why sparse autoencoders yield misleading decompositions of steering vectors in large language models reveals two main limitations.
Yungang Yi, Weihua Li, Matthew Kuo +1 authors
PerceiverS, using Effective Segmentation and Multi-Scale attention, enhances long-structured and expressive symbolic music generation by combining cross-attention and self-attention.
Fadhel Ayed, Ali Maatouk, Nicola Piovesan +3 authors
Hermes, a chain of LLM agents using blueprints, facilitates automatic, reliable, and accurate network modeling through structured logical steps, advancing towards fully autonomous network operations.
Zitong Lan, Chenhao Zheng, Zhiwei Zheng +1 authors
Acoustic Volume Rendering (AVR) uses frequency-domain volume rendering and spherical integration to synthesize impulse responses more accurately and realistically than leading methods.
Qiang Zhou, Shaofeng Zhang, Nianzu Yang +2 authors
MVideo enhances text-to-video generation by incorporating mask sequences for precise action control, improving the depiction of complex motions in long-duration videos.
Yushi Yang, Filip Sondej, Harry Mayne +1 authors
Direct preference optimization reduces toxicity in language models by balancing neuron effects across multiple groups rather than solely dampening the most toxic neurons.
Zeyu Zhang, Hang Gao, Akide Liu +5 authors
KMM enhances Mamba's motion generation through key frame masking and contrastive learning, achieving state-of-the-art results with reduced parameters.
Shauli Ravfogel, Anej Svete, Vésteinn Snæbjarnarson +1 authors
A framework using Generalized Structural Equation Models and Gumbel-max trick allows generation of meaningful counterfactuals in language models, revealing side effects of common intervention techniques.
Maya Varma, Jean-Benoit Delbrouck, Zhihong Chen +2 authors
RaVL improves VLM robustness by identifying and mitigating spurious correlations at the local image feature level during fine-tuning.
Zhaofeng Wu, Xinyan Velocity Yu, Dani Yogatama +2 authors
Language models develop a shared representation space that organizes semantically similar inputs from diverse languages and modalities, facilitating cross-modal and cross-lingual understanding and processing.
Cheng Zhang, Hanna Foerster, Robert D. Mullins +2 authors
A method, HSPI, is introduced to identify the GPU architecture and software stack of a machine learning model based on its input-output behavior, providing a means to verify the authenticity of LLM inference services.
Yen-Ting Lin, Chao-Han Huck Yang, Zhehuai Chen +10 authors
A Multi-Task Correction MoE enhances error correction across different domains, achieving state-of-the-art performance with improved WER and BLEU scores.
Aayush Shah, Shankar Jayaratnam
Two small protein language models, based on Llama-3-8B and Phi-3-mini, achieve comparable performance to larger models in protein generation tasks using parameter-efficient adaptation techniques.
Xiaojun Wu, Junxi Liu, Huanyi Su +10 authors
Golden Touchstone is a comprehensive bilingual benchmark for financial LLMs, addressing limitations of existing benchmarks and providing a practical evaluation tool for models' performance across multiple tasks.
Nam Le Hai, Anh M. T. Bui, Phuong T. Nguyen +2 authors
A dataset of self-admitted technical debt comments coupled with source code enhances prediction accuracy in detecting technical debt types.
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