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
Yuanyi Wang, Yifan Yang, Su Lu +9 authors
Research investigates how task geometry influences continual post-training of large language models, identifying geometry conflict as both a cause of forgetting and a control mechanism for update integration.
George Wu, Nan Jing, Qing Yi +7 authors
TMAS is a multi-agent framework for test-time scaling that enhances large language model reasoning through structured collaboration and hierarchical memory systems.
Shihao Qi, Jie Ma, Rui Xing +15 authors
Multi-agent systems face challenges in sustained coordination and error propagation, requiring integrated approaches that enable continuous diagnosis, reorganization, and behavioral refinement across structured collaboration stages.
Eilam Shapira, Moshe Tennenholtz, Roi Reichart
AI agents can predict counterpart decisions in negotiation games by combining tabular features with LLM-based text representations and hidden states from a frozen observer model, outperforming direct prompting methods.
Shuangrui Ding, Xuanlang Dai, Long Xing +14 authors
WildClawBench evaluates language and vision-language models on realistic long-horizon tasks using actual CLI environments with real tools instead of synthetic sandboxes.
Bowen Peng, Théo Gigant, Jeffrey Quesnelle
Token-Superposition Training (TST) improves pre-training efficiency by combining contiguous tokens into bags during a superposition phase with multi-hot cross-entropy objective, achieving faster training times without architectural changes.
Yuanyi Wang, Yanggan Gu, Yiming Zhang +6 authors
Empirical scaling laws for language model merging reveal power-law relationships between model size, expert count, and cross-entropy performance, enabling predictive planning for optimal model composition.
Junfeng Fang, Zhepei Hong, Mao Zheng +7 authors
Rubric-based on-policy distillation achieves improved sample efficiency over traditional logit-based methods by using structured semantic rubrics instead of teacher logits.
Yifan Wang, Tong He
A novel approach called Warp-as-History enables camera-controlled video generation by transforming camera-induced warps into pseudo-history representations, achieving zero-shot capability without training or test-time optimization.
Jisu Nam, Jahyeok Koo, Soowon Son +4 authors
TrackCraft3R enables efficient dense 3D tracking from monocular video by adapting video diffusion transformers to follow physical points across frames using dual-latent representation and temporal RoPE alignment.
Zhennan Chen, Junwei Zhu, Xu Chen +7 authors
Latent-to-Pixel transfer paradigm efficiently leverages pre-trained latent diffusion models to create pixel-space models with minimal training overhead and high-resolution generation capabilities.
Runhui Huang, Jie Wu, Rui Yang +2 authors
AlphaGRPO enhances multimodal generation by applying Group Relative Policy Optimization to AR-Diffusion Unified Multimodal Models through self-reflective refinement and decompositional verifiable reward mechanisms.
Dong-Yang Li, Wang Zhao, Yuxin Chen +5 authors
Pixal3D introduces a pixel-aligned 3D generation approach that addresses fidelity issues in 3D asset creation by establishing direct pixel-to-3D correspondences through back-projection conditioning.
Bihui Yu, Xinglong Xu, Junjie Jiang +6 authors
Visual typesetting optimization addresses the gap between compilable LaTeX documents and publication-ready PDFs through vision-in-the-loop agents that iteratively diagnose and repair layout defects.
Tsz Ting Chung, Lemao Liu, Mo Yu +1 authors
Many-shot in-context learning for reasoning tasks exhibits different scaling behaviors than non-reasoning tasks, with demonstration ordering and selection significantly impacting performance.
Xuehai Bai, Yang Shi, Yi-Fan Zhang +7 authors
Recent image editing models have achieved remarkable progress in instruction following, multimodal understanding, and complex visual editing. However, existing benchmarks often fail to faithfully reflect human judgment, especially for strong frontier models, due to limited task difficulty and coarse-grained evaluation protocols. In parallel, reward models have become increasingly important for RL-based image editing optimization, yet existing reward model benchmarks still rely on unrealistic evaluation settings that deviate from practical RL scenarios. These limitations hinder reliable assessment of both image editing models and reward models. To address these challenges, we introduce Edit-Compass and EditReward-Compass, a unified evaluation suite for image editing and reward modeling. Edit-Compass contains 2,388 carefully annotated instances spanning six progressively challenging task categories, covering capabilities such as world knowledge reasoning, visual reasoning, and multi-image editing. Beyond broad task coverage, Edit-Compass adopts a fine-grained multidimensional evaluation framework based on structured reasoning and carefully designed scoring rubrics. In parallel, EditReward-Compass contains 2,251 preference pairs that simulate realistic reward modeling scenarios during RL optimization.
Xuanyu Zhu, Yan Bai, Yang Shi +4 authors
DRoRAE enhances visual representation by fusing multi-layer features from pretrained vision encoders through adaptive routing and incremental correction, improving reconstruction and generation quality.
Yihao Meng, Zichen Liu, Hao Ouyang +11 authors
CausalCine enables interactive, multi-shot video generation by addressing limitations of autoregressive models through causal modeling, dynamic memory routing, and real-time distillation techniques.
Hyeon Hwang, Jiwoo Lee, Jaewoo Kang
ThinC framework enables mathematical problem solving where code serves as the primary reasoning mechanism instead of a verification tool, demonstrating superior performance on math benchmarks.
Bowen Peng, Subho Ghosh, Jeffrey Quesnelle
Lighthouse Attention enables efficient training of causal transformers at long sequences by using hierarchical selection-based attention that reduces computational complexity while maintaining model performance.
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