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。
Chengpeng Li, Mingfeng Xue, Zhenru Zhang +7 authors
START integrates external tools into large reasoning models to enhance capabilities, using techniques like Hint-infer and Hint Rejection Sampling Fine-Tuning, achieving high performance across various benchmarks.
50 篇论文 · 按点赞排序
Jindong Jiang, Xiuyu Li, Zhijian Liu +13 authors
STORM integrates a temporal encoder using the Mamba State Space Model to enhance video understanding and reduce computational costs in multimodal large language models.
Abdelrahman Abouelenin, Atabak Ashfaq, Adam Atkinson +70 authors
Phi-4-Mini and Phi-4-Multimodal are compact yet highly capable language and multimodal models, featuring advanced attention mechanisms and LoRA adapters for efficient and powerful reasoning across various tasks.
Ziyu Liu, Zeyi Sun, Yuhang Zang +5 authors
Visual Reinforcement Fine-Tuning (Visual-RFT) enhances Large Vision-Language Models (LVLMs) through reinforcement learning with visual perception verifiable rewards, achieving competitive performance in various visual and reasoning tasks with limited data.
Sambal Shikhar, Mohammed Irfan Kurpath, Sahal Shaji Mullappilly +5 authors
LLMVoX is a lightweight, LLM-agnostic TTS system that decouples speech synthesis from model processing, ensuring high-quality speech with low latency and supporting seamless multimodal dialogue.
Yiran Zhao, Chaoqun Liu, Yue Deng +8 authors
Babel is an open multilingual LLM that expands its parameter count through layer extension, covering numerous languages and achieving superior performance in multilingual tasks compared to other open LLMs.
Kashun Shum, Yuzhen Huang, Hongjian Zou +5 authors
A lightweight data selection method called PreSelect effectively reduces computational requirements while enhancing language model performance by focusing on data predictive of downstream abilities.
Jingkang Yang, Shuai Liu, Hongming Guo +19 authors
EgoLife introduces EgoButler, combining EgoGPT and EgoRAG, to provide AI-powered assistance through wearable glasses using a comprehensive egocentric dataset.
Silei Xu, Wenhao Xie, Lingxiao Zhao +1 authors
Chain of Draft (CoD) improves the efficiency of Large Language Models (LLMs) in reasoning tasks by generating concise intermediate thoughts, enhancing accuracy while reducing token usage, cost, and latency.
Jay Zhangjie Wu, Yuxuan Zhang, Haithem Turki +6 authors
Difix3D+, a single-step diffusion model pipeline, enhances 3D reconstruction and novel-view synthesis by improving photorealism and reducing artifacts from extreme viewpoints.
Tin Nguyen, Logan Bolton, Mohammad Reza Taesiri +1 authors
Highlighted Chain-of-Thought Prompting improves human verification of LLM responses but can lead to overconfidence in incorrect answers.
Shimao Zhang, Xiao Liu, Xin Zhang +4 authors
A Process-based Self-Rewarding pipeline enhances LLMs' mathematical reasoning by iteratively evaluating and optimizing their own outputs.
Kanishk Gandhi, Ayush Chakravarthy, Anikait Singh +2 authors
Research reveals that effective self-improvement in language models is linked to their ability to perform specific reasoning behaviors, such as verification and backward chaining, which can be primed through appropriate training.
Zhuoqun Li, Haiyang Yu, Xuanang Chen +6 authors
SolutionRAG leverages tree-based exploration and bi-point thinking to achieve state-of-the-art performance on SolutionBench, a benchmark for generating complete and feasible engineering solutions.
Zhangchen Xu, Yang Liu, Yueqin Yin +2 authors
KodCode is a synthetic coding dataset that ensures broad coverage and correctness through systematic validation, enabling state-of-the-art performance in coding benchmarks.
Arnav Kumar Jain, Gonzalo Gonzalez-Pumariega, Wayne Chen +3 authors
We address the problem of code generation from multi-turn execution feedback. Existing methods either generate code without feedback or use complex, hierarchical reinforcement learning to optimize multi-turn rewards. We propose a simple yet scalable approach, muCode, that solves multi-turn code generation using only single-step rewards. Our key insight is that code generation is a one-step recoverable MDP, where the correct code can be recovered from any intermediate code state in a single turn. muCode iteratively trains both a generator to provide code solutions conditioned on multi-turn execution feedback and a verifier to score the newly generated code. Experimental evaluations show that our approach achieves significant improvements over the state-of-the-art baselines. We provide analysis of the design choices of the reward models and policy, and show the efficacy of muCode at utilizing the execution feedback. Our code is available at https://github.com/portal-cornell/muCode.
Sreyan Ghosh, Zhifeng Kong, Sonal Kumar +6 authors
AF2, an advanced Audio-Language Model, demonstrates state-of-the-art performance on over 20 benchmarks and extends audio understanding to long segments using the LongAudio dataset and LongAudioBench.
Kunlun Zhu, Hongyi Du, Zhaochen Hong +8 authors
A new benchmark evaluates LLM-based multi-agent systems in diverse scenarios, measuring task completion and quality of collaboration using coordination protocols and innovative strategies.
Weimin Xiong, Yifan Song, Qingxiu Dong +4 authors
MPO framework enhances LLM-based agent planning with explicit high-level guidance, improving task completion and generalization through meta plans.
Tong Wu, Junzhe Shen, Zixia Jia +2 authors
TOKENSWIFT addresses challenges in ultra-long sequence generation by introducing a framework that achieves significant speedup without compromising model quality.
Nam V. Nguyen, Dien X. Tran, Thanh T. Tran +4 authors
SemViQA combines Semantic-based Evidence Retrieval and Two-step Verdict Classification to enhance Vietnamese fact-checking, achieving top performance with improved speed.
Ziqian Ning, Huakang Chen, Yuepeng Jiang +5 authors
DiffRhythm, a latent diffusion-based model, synthesizes full-length songs with both vocals and accompaniment efficiently, using simple and scalable methods.
Jiaxin Deng, Shiyao Wang, Kuo Cai +5 authors
OneRec, an end-to-end generative model with encoder-decoder architecture, sparse Mixture-of-Experts, session-wise generation, and Iterative Preference Alignment using Direct Preference Optimization, significantly improves recommendation quality and user watch-time.
Amr Mohamed, Mingmeng Geng, Michalis Vazirgiannis +1 authors
Large language models distort information through iterative generation, influenced by language choice and chain complexity, with potential implications for AI-mediated information reliability.
L. Degeorge, A. Ghosh, N. Dufour +2 authors
Strategic data augmentation of small, well-curated datasets can achieve competitive performance in text-to-image generation with significantly fewer parameters and training images compared to large-scale models.
Jude Khouja, Karolina Korgul, Simi Hellsten +6 authors
A new evaluation benchmark LINGOLY-TOO is introduced to assess the reasoning capabilities of LLMs by reducing the impact of memorization and prior data exposure.
Xuanchi Ren, Tianchang Shen, Jiahui Huang +7 authors
GEN3C, a generative video model, uses a 3D cache to achieve precise camera control and temporal consistency in video generation, outperforming previous methods in sparse-view novel view synthesis.
Tingyu Song, Guo Gan, Mingsheng Shang +1 authors
A benchmark and evaluation method for instruction-following information retrieval in expert domains reveal challenges for current models in handling complex, domain-specific instructions.
Zhijian Zhuo, Yutao Zeng, Ya Wang +5 authors
HybridNorm, combining pre-normalization and post-normalization, enhances training stability and performance of deep transformer models, especially in LLMs.
Siming Huang, Yuliang Xu, Mingmeng Geng +2 authors
The analysis of Large Language Models' impact on Wikipedia shows a modest influence on article content and NLP tasks like machine translation and RAG, highlighting potential future risks.
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