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
Zilong Chen, Chaorui Deng, Kunchang Li +2 authors
Structured language in prompts drives diffusion loss scaling, enabling improved visual generation through annotated prompts and a trained prompter.
Zelong Sun, Jun Wang, Kaicheng Yang +3 authors
UniME-R1 improves multimodal retrieval by generating retrieval-centric reasoning guided by initial candidate feedback rather than query-only explanations.
Yifan Shen, Jian Xu, Boyi Li +6 authors
ChronoVision improves visual temporal reasoning by aligning latent imagery with logic through reconstructive prediction, ROI attention, and reinforcement learning, achieving strong results on video reasoning benchmarks.
Liangyang Ouyang, Ruicong Liu, Xuangeng Chu +2 authors
HelloWorld is a video world model that enables users to trigger natural social responses from on-screen characters via self-distilled training and inference-time cross-attention modulation.
DiffusionGemma Team, Adrien Ali Taïga, James Assiene +41 authors
DiffusionGemma is a fine-tuned mixture-of-experts language model that uses discrete diffusion to generate text blocks in parallel, achieving high speed while preserving capabilities like multimodal inputs and reasoning.
Chunji Lv, Yangguang Wei, Junlin Liu +6 authors
Persistent Consistency Self-Distillation improves reinforcement learning for language model agents by weighting token-level teacher guidance according to sustained local support, yielding stronger results on interactive tasks.
Yu Yang, Xuemeng Yang, Licheng Wen +17 authors
This work proposes agent-centric interactive world proxies that provide diverse feedback for continual agent improvement across inference, training, and co-evolution stages.
Jingsheng Zheng, Xinyuan Fang, Jintian Zhang +3 authors
OneDayAgent is a general long-horizon harness that decomposes open-ended tasks, manages execution memory, and verifies outputs across diverse LLM backends.
Xiangning Lin, Shenzhe Zhu, Shu Yang +23 authors
System prompts are instructions configured by developers to govern the behaviors of foundation models in AI applications. They are used throughout commercial AI products, but are rarely disclosed to the public or regulators, creating a serious trust and accountability gap in the wide deployment of AI systems. In this paper, we introduce Artificial Intelligence System Prompt Assurance (AISPA), a user-centric framework for systematically auditing system prompts in AI systems. AISPA examines specific parts of a system prompt and evaluates them along eight dimensions that matter to users. We then use this framework to review 3,249 instructions from system prompts in 88 commercial AI products, classifying each instruction as either protective (of users) or problematic. Our audit surfaces four core findings. First, system prompt design varies substantially across products and developers, with some organizations averaging over 60 protective instructions per product while others average fewer than 5. Second, protective instructions are widely adopted but shallow in scope: 98.9% of products contain at least one, yet only 24% cover all eight dimensions of the AISPA taxonomy. Third, system prompts have grown steadily longer and more protective of users, suggesting that user protection is becoming a more visible concern in commercial prompt design. Fourth, despite this progress, problematic instructions remain pervasive: roughly 40% of products contain at least one instruction that works against user interests, and protective and problematic instructions frequently coexist within the same prompt. Our findings highlight the need for greater transparency, standardization, and independent oversight for system prompts in commercial AI products.
Varun Ursekar, Apaar Shanker, Yash Maurya +4 authors
HarnessOpt-Bench measures how well frontier LLMs iteratively improve agent harnesses under constrained evaluation budgets, revealing substantial variation across models and tasks.
Jiale Han, Xiang Li, Jing Qian +7 authors
Economic world models are generative agent-based simulations organized by a six-level capability ladder to build high-fidelity economic sandboxes and AI training substrates.
Yuxue Yang, Shuyao Shang, Jiahe Wang +13 authors
WorldExam is a hierarchical benchmark that evaluates controllable video generation models on visual quality, control adherence, spatial consistency, and inherent world reactivity, revealing that current models lack consistent reactive capabilities despite high visual fidelity.
Shuhan Xue, Zixin Ding, Yichen Shen +6 authors
PAST-Bench evaluates whether persistent AI agents improve from retained experience across tasks, and Hermes+ enhances this via targeted loop interventions.
Zijian Xu, Wenshuo Zhang, Zisen Qin +4 authors
Researchers introduce a benchmark and task for resolving recurring personalized coding ambiguities across sessions using past user history, proposing a lightweight gating method to improve long-term assistant alignment.
Byeongho Heo, Jaehui Hwang, Sangdoo Yun +1 authors
On-policy delta distillation improves multilingual mathematical reasoning and reduces cross-language performance gaps, though multilingual data is needed to preserve target-language outputs.
Ayoub Kirouane, Christos Petrocheilos
Researchers adapted NVIDIA's Nemotron retrieval and generation stack for Modern Greek, showing that fine-tuned dense retrieval and reranking outperform baselines on specialist corpora, while a LoRA-tuned mixture-of-experts reader improves answer correctness and citation quality, and they released the HERA benchmark.
Zelin Tan, Yiqun Zhang, Hao Li +11 authors
Verified data synthesis improves language-model agents' ability to identify, apply, and coordinate reusable skills through high-quality supervised training.
Zhoushun Yu, Xiaoyu Hu, Xiangyu Xu
QQWorld improves latent world models by replacing the Epps-Pulley objective with quantile-quantile matching to maintain gradient control in heavy tails, boosting planning success.
Wanshun Su, Yang Shi, Feihu Liu +10 authors
OmniPack is a training-free framework that combines structural pre-LLM token compression with query-guided semantic refinement inside the LLM to reduce computational overhead in omni-modal models while preserving performance.
Senyu Fei, Xiaopeng Yu, Siyin Wang +3 authors
A world critic model that jointly predicts future latent states and estimates values improves temporal reasoning and state-of-the-art performance in vision-language-action reinforcement learning.
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