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。
Shuai Bai, Keqin Chen, Xuejing Liu +24 authors
Qwen2.5-VL, the latest vision-language model, advances visual recognition, document parsing, and video comprehension through dynamic resolution processing, Window Attention, and a native Vision Transformer.
50 篇论文 · 按点赞排序
Deepak Nathani, Lovish Madaan, Nicholas Roberts +14 authors
A new framework and benchmark, MLGym and MLGym-Bench, is introduced for evaluating and developing LLM agents across diverse AI research tasks using reinforcement learning.
Michael Tschannen, Alexey Gritsenko, Xiao Wang +11 authors
SigLIP 2, a multilingual vision-language encoder, improves upon SigLIP with unified training techniques, enhancing performance in zero-shot classification, image-text retrieval, localization, and dense prediction across various model sizes and data diversity.
Jingyang Yuan, Huazuo Gao, Damai Dai +12 authors
NSA, a trainable sparse attention mechanism, enhances long-context modeling efficiency without sacrificing performance, achieving improvements in speed and accuracy over full attention models.
Shen Nie, Fengqi Zhu, Zebin You +7 authors
LLaDA, a diffusion model trained from scratch, outperforms autoregressive models in benchmarks and demonstrates strong instruction-following capabilities, challenging the dominance of ARMs in LLMs.
M-A-P Team, Xinrun Du, Yifan Yao +92 authors
SuperGPQA, a benchmark evaluating LLMs across 285 disciplines, reveals performance gaps and offers insights into the collaborative filtering process with human experts.
Sergey Pletenev, Maria Marina, Daniil Moskovskiy +4 authors
Low-rank adaptation (LoRA) effectively updates LLMs with new knowledge but risks degrading performance on external benchmarks due to data bias and overconfidence.
Yuhao Zhang, Zhiheng Liu, Fan Bu +3 authors
Soundwave addresses the representation space gap and sequence length inconsistency in end-to-end speech large language models using an efficient training strategy and novel architecture, outperforming Qwen2-Audio with significantly less data.
Yuri Kuratov, Mikhail Arkhipov, Aydar Bulatov +1 authors
Using per-sample optimization, compression ratios of up to x1500 are achieved for sequence-to-vector compression in language models, highlighting a large gap between theoretical and practical limits.
Dacheng Li, Shiyi Cao, Chengkun Cao +6 authors
A hybrid test-time scaling framework improves code generation coverage and accuracy across various models and domains.
Jianwei Yang, Reuben Tan, Qianhui Wu +10 authors
Magma is a multimodal foundation model with both verbal intelligence and spatial-temporal intelligence, trained on diverse datasets to perform agentic tasks like UI navigation and robotic manipulation, outperforming specialized models.
Lijie Liu, Tianxiang Ma, Bingchuan Li +4 authors
A unified video generation framework called Phantom integrates text and image inputs to produce subject-consistent videos, focusing on human generation and leveraging cross-modal alignment.
Guoqing Ma, Haoyang Huang, Kun Yan +112 authors
Step-Video-T2V is a state-of-the-art text-to-video model with 30B parameters, utilizing a deep compression Variational Autoencoder and a DiT with 3D full attention for high-quality video generation.
Jaehyeong Jo, Sung Ju Hwang
A continuous diffusion model for language modeling that leverages the geometry of discrete distributions outperforms existing discrete models and matches autoregressive models in performance.
Ziming Liu, Yifan Yang, Chengruidong Zhang +4 authors
RAS, a novel sampling strategy for diffusion transformers, dynamically adjusts sampling ratios based on regions of focus, achieving speedups in diffusion models with minimal quality loss.
Kenneth Enevoldsen, Isaac Chung, Imene Kerboua +83 authors
The MMTEB benchmark expands MTEB with over 500 tasks across 250+ languages to comprehensively evaluate text embeddings, finding that smaller multilingual models can outperform large LLMs in many cases.
Tian Xie, Zitian Gao, Qingnan Ren +7 authors
A rule-based reinforcement learning system for large reasoning models uses synthetic logic puzzles to develop advanced reasoning skills and demonstrates generalization on complex math benchmarks.
Samuel Miserendino, Michele Wang, Tejal Patwardhan +1 authors
We introduce SWE-Lancer, a benchmark of over 1,400 freelance software engineering tasks from Upwork, valued at \1 million USD total in real-world payouts. SWE-Lancer encompasses both independent engineering tasks--ranging from 50 bug fixes to \$32,000 feature implementations--and managerial tasks, where models choose between technical implementation proposals. Independent tasks are graded with end-to-end tests triple-verified by experienced software engineers, while managerial decisions are assessed against the choices of the original hired engineering managers. We evaluate model performance and find that frontier models are still unable to solve the majority of tasks. To facilitate future research, we open-source a unified Docker image and a public evaluation split, SWE-Lancer Diamond (https://github.com/openai/SWELancer-Benchmark). By mapping model performance to monetary value, we hope SWE-Lancer enables greater research into the economic impact of AI model development.
Yue Huang, Chujie Gao, Siyuan Wu +63 authors
A framework is provided for enhancing the trustworthiness of Generative Foundation Models via comprehensive governance principles and the introduction of TrustGen, a benchmarking platform that evaluates trustworthiness across diverse GenFMs.
Xialin He, Runpei Dong, Zixuan Chen +1 authors
A learning framework is developed to enable humanoid robots to automatically recover from falls on varied terrains by discovering and refining motion trajectories.
Alejandro Cuadron, Dacheng Li, Wenjie Ma +13 authors
Analysis of overthinking in Large Reasoning Models shows it correlates with decreased performance, and selective mitigation techniques can improve performance and reduce costs.
Jonathan Roberts, Mohammad Reza Taesiri, Ansh Sharma +20 authors
ZeroBench, a new visual reasoning benchmark, demonstrates the limitations of current large multimodal models in understanding images through a set of questions they cannot solve.
Zihan Liu, Shuangrui Ding, Zhixiong Zhang +6 authors
SongGen is a unified, single-stage, auto-regressive transformer model for text-to-song generation, offering controllable attributes and flexible output modes.
Yuetai Li, Xiang Yue, Zhangchen Xu +5 authors
Mix Distillation improves small model reasoning by balancing long and short chain-of-thought examples, addressing the Small Model Learnability Gap.
Hao Gao, Shaoyu Chen, Bo Jiang +11 authors
3DGS-based closed-loop reinforcement learning for autonomous driving integrates photorealistic simulation and specialized rewards, reducing collision rates compared to imitation learning.
Bencheng Liao, Hongyuan Tao, Qian Zhang +5 authors
A framework converts existing decoder-only Multimodal Large Language Models to linear-complexity models using distillation, reducing computational complexity and memory usage while maintaining performance.
Gleb Gerasimov, Yaroslav Aksenov, Nikita Balagansky +2 authors
Layer-Integrated Memory (LIMe) addresses representation collapse in Transformers by allowing access to hidden states from earlier layers, leading to consistent performance improvements across tasks.
Feng Luo, Rui Yang, Hao Sun +5 authors
Decomposed Reward Models (DRMs) use PCA to extract interpretable preference dimensions from binary comparisons, providing a scalable alternative to traditional reward models for personalized AI systems.
Austin Yubo He, Zi-Wen Liu
Reinforcement learning is used to develop low-weight quantum error-correcting codes, significantly reducing physical qubit overhead compared to existing methods.
Jusen Du, Weigao Sun, Disen Lan +2 authors
Mixture-of-Memories (MoM) uses independent memory states to enhance recall in sequence modeling while maintaining computational efficiency, outperforming linear models and achieving performance comparable to Transformers.
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