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
Jiaqi Chen, Bang Zhang, Ruotian Ma +5 authors
Self-Play Critic (SPC) enhances LLM reasoning reliability by using adversarial self-play to improve error detection and reasoning performance without requiring manual step-level annotation.
Yu Li, Qizhi Pei, Mengyuan Sun +6 authors
CipherBank evaluates the reasoning capabilities of large language models in cryptographic decryption tasks, revealing significant gaps in their performance across classical and custom encryption methods.
Luca Moroni, Giovanni Puccetti, Pere-Lluis Huguet Cabot +6 authors
SAVA, a novel vocabulary adaptation technique using neural mapping, optimizes English LLMs for Italian, reducing token fertility and improving performance across various tasks with minimal additional training.
Xinyu Chen, Yunxin Li, Haoyuan Shi +4 authors
A novel video comprehension benchmark evaluates multimodal AI systems across diverse cultures and languages, revealing performance gaps in specific domains and linguistic contexts.
Hanlei Zhang, Zhuohang Li, Yeshuang Zhu +5 authors
MMLA benchmark assesses multimodal large language models' understanding of human language semantics across various core dimensions, highlighting limitations and offering a foundation for future research.
Pierre-Carl Langlais, Pavel Chizhov, Mattia Nee +6 authors
Two new mid-sized reasoning models, Pleias-RAG-350m and Pleias-RAG-1B, perform competitively on RAG benchmarks and support multilingual citation and grounding.
Haotian Luo, Haiying He, Yibo Wang +6 authors
A two-stage adaptive reasoning framework combining long and short chain-of-thought models, trained with bi-level preference learning, reduces inference costs while maintaining performance across mathematical datasets.
Piotr Nawrot, Robert Li, Renjie Huang +3 authors
Sparse attention methods enable extending long-context capabilities in Transformer LLMs with varying efficiency-accuracy trade-offs across different tasks and scales.
Toby Simonds
Reinforcement learning enhances large language models to outperform humans in complex rocket design optimization tasks, suggesting broader applications in engineering.
Qihao Liu, Ju He, Qihang Yu +2 authors
ReVision enhances video generation by integrating 3D physical priors into video diffusion models, improving motion fidelity and coherence with fewer parameters.
Sicheng Mo, Thao Nguyen, Xun Huang +9 authors
X-Fusion enhances pretrained LLMs for multimodal tasks by integrating vision-specific information through a dual-tower design and freezes the LLM's parameters for consistency in understanding and generation.
Jingyu Zhang, Jiacan Yu, Marc Marone +2 authors
BloomScrub, an inference-time method using Bloom filters, effectively reduces copyright infringement risks in large language models by detecting and rewriting infringing quotes.
Daneul Kim, Jingxu Zhang, Wonjoon Jin +4 authors
A method for zero-shot video customization achieves strong performance by decoupling subject-specific learning from temporal dynamics using image customization datasets and randomization techniques.
Mihai Nadas, Laura Diosan, Andrei Piscoran +1 authors
A new dataset, TF1-EN-3M, uses instruction-tuned models to generate three million English fables following a structured format, evaluated using a combination of automated metrics and human judgments, and released under a permissive license.
Thao Nguyen, Krishna Kumar Singh, Jing Shi +3 authors
Yo'Chameleon integrates soft-prompt tuning to personalize large multimodal models for answering questions and generating images of specific concepts from a few examples.
Zhikai Wang, Jiashuo Sun, Wenqi Zhang +4 authors
VCBENCH evaluates multimodal mathematical reasoning by assessing LVLMs on a benchmark with explicit visual dependencies and multi-image reasoning tasks.
Ranran Zhen, Juntao Li, Yixin Ji +7 authors
A survey explores methods to enhance low latency and high throughput in Large Language Model inference by addressing memory overhead and computational demands of the attention mechanism.
Jie Zhu, Qian Chen, Huaixia Dou +4 authors
A reasoning-enhanced framework, DianJin-R1, is proposed to improve financial domain reasoning in LLMs through augmented supervision and reinforcement learning, achieving superior performance on financial benchmarks.
Haoran Geng, Feishi Wang, Songlin Wei +34 authors
RoboVerse provides a comprehensive framework with high-quality synthetic data, simulation platform, and unified benchmarks to address data scaling and evaluation challenges in robotics.
Md Ashiqur Rahman, Raymond A. Yeh
The proposed method enhances downsampling in group equivariant architectures by incorporating anti-aliasing and subgroup selection, improving classification accuracy and model efficiency.
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