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
Xintong Hao, Ruijie Zhu, Ge Zhang +2 authors
The Massive Genre-Audience reformulation method augments training data to reduce repetition-related degradation and supports more efficient scaling of large language models.
Kaikai An, Li Sheng, Ganqu Cui +4 authors
UltraIF enhances open-source LLMs' instruction-following capabilities by decomposing complex prompts and using evaluation questions to filter responses, achieving competitive scores on benchmarks.
Jinyang Wu, Mingkuan Feng, Shuai Zhang +4 authors
AStar leverages MCTS to enhance multimodal large language models' visual reasoning from limited data, achieving superior accuracy while maintaining efficiency.
Yiren Song, Danze Chen, Mike Zheng Shou
LayerTracer, a diffusion transformer framework, generates high-quality, editable layered SVGs using text-conditioned DiT for blueprint generation and conditional diffusion for vectorization, outperforming existing methods in quality and editability.
Maohao Shen, Guangtao Zeng, Zhenting Qi +7 authors
Satori, a 7B LLM enhanced with COAT reasoning and two-stage training, achieves top performance in math benchmarks and generalizes well to unrelated tasks.
Huanqia Cai, Yijun Yang, Winston Hu
A multimodal IQ test framework (MM-IQ) is presented to evaluate reasoning capacities in AI, revealing significant performance gaps compared to human baselines.
Yiren Song, Cheng Liu, Mike Zheng Shou
A diffusion transformer-based framework, MakeAnything, generates coherent procedural tutorials across domains using a large dataset and asymmetric low-rank adaptation, outperforming existing methods.
Bill Yuchen Lin, Ronan Le Bras, Kyle Richardson +4 authors
LLMs exhibit decreased accuracy in complex non-monotonic reasoning, as demonstrated through the ZebraLogic framework assessing logic grid puzzles.
Ziyan Guo, Zeyu Hu, Na Zhao +1 authors
MotionLab is a unified deep learning framework that uses rectified flows, MotionFlow Transformer, Aligned Rotational Position Encoding, and Motion Curriculum Learning to enable efficient and versatile human motion generation and editing across various tasks.
Yinjie Wang, Ling Yang, Guohao Li +2 authors
ScoreFlow framework improves performance in large language model multi-agent systems through efficient gradient-based optimization and Score-DPO, enabling smaller models to outperform larger ones with reduced computational costs.
Daniil Tiapkin, Daniele Calandriello, Johan Ferret +4 authors
Experiments reveal that knowledge distillation can lead to teacher hacking, an issue akin to reward hacking, but it can be mitigated by using online data generation techniques that enhance data diversity.
Ahmed Heakl, Sara Ghaboura, Omkar Thawkar +4 authors
AIN, a bilingual English-Arabic multimodal model, achieves state-of-the-art performance in Arabic and strong English-language visual capabilities across diverse domains.
Jinbo Xing, Long Mai, Cusuh Ham +5 authors
MotionCanvas integrates user-driven controls into image-to-video generation to enable intuitive 3D-aware motion design without requiring 3D training data.
DiJia Su, Hanlin Zhu, Yingchen Xu +3 authors
A hybrid representation using latent tokens generated by VQ-VAE reduces the length of reasoning traces in large language models, improving performance on both reasoning and fine-tuning tasks.
Dongwon Jo, Jiwon Song, Yulhwa Kim +1 authors
FastKV is a KV cache compression framework that reduces prefill and decoding latency by decoupling context computation from cache budget through token-selective propagation and independent key-value entry selection.
Zongyu Lin, Yao Tang, Xingcheng Yao +4 authors
QLASS uses Q-value estimation for stepwise guidance in language agents, improving performance with limited annotated data.
Zhiheng Lyu, Xueguang Ma, Wenhu Chen
A unified perception framework treating all modalities as pixel inputs (PEAP) outperforms token-based input in some tasks but degrades reasoning and coding capabilities, requiring enhancements in foundational models’ perceptual abilities.
Xiang Liu, Zhenheng Tang, Hong Chen +6 authors
The study evaluates KV cache compression methods in LLMs and introduces ShotKV, which improves performance on long-context tasks under high compression ratios.
Gaoyue Zhou, Hengkai Pan, Yann LeCun +1 authors
DINO World Model learns from offline trajectories to predict future visual dynamics and optimize behavior for diverse tasks without expert demonstrations or reward modeling.
Yuanye Liu, Jiahang Xu, Li Lyna Zhang +6 authors
CFPO, an integrated optimization method for both content and formatting of prompts, enhances LLM performance across tasks.
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