SafeDialBench evaluates LLM safety across 22 scenarios and 7 attack strategies.

Hongye Cao, Sijia Jing, Yanming Wang et al.

CiteCheck introduces a large-scale Chinese citation faithfulness dataset using a two-stage manual annotation and LLM-based data augmentation, enabling efficient training of models with high robustness.

Ziyao Xu, Shaohang Wei, Zhuoheng Han et al.

UniRoute uses feature vectors for dynamic routing, enabling efficient LLM inference across 30+ unseen models.

Wittawat Jitkrittum, Harikrishna Narasimhan, Ankit Singh Rawat et al.

AMKOR fuses parametric and retrieved knowledge using probabilistic beam search, achieving state-of-the-art multi-hop QA performance.

Jackson Coleman, Isaiah Lawrence, Benjamin Turner

LIMO achieves complex reasoning with minimal data, scoring 63.3% on AIME24 and 95.6% on MATH500.

Yixin Ye, Zhen Huang, Yang Xiao et al.

Proposes Token Assorted, mixing latent and text tokens to improve reasoning, reducing trace length by 17% and boosting performance by up to 13.3%.

DiJia Su, Hanlin Zhu, Yingchen Xu et al.

iVISPAR uses a sliding geometric puzzle to evaluate VLM spatial reasoning across 3D, 2D, and text modalities, highlighting current limitations.

Julius Mayer, Mohamad Ballout, Serwan Jassim et al.

AStar uses retrieved thought cards to boost training-free multimodal reasoning, reaching 53.9% on MathVerse.

Jinyang Wu, Mingkuan Feng, Guocheng Zhai et al.

Proposes a dynamic KL-weighted diffusion model guided by large language models, achieving superior text-to-image synthesis with FID 30.5 and human score 4.6.

Julian Perry, Frank Sanders, Carter Scott

M+ extends MemoryLLM with a long-term memory mechanism, enabling over 160k tokens retention, significantly improving long-text understanding.

Yu Wang, Dmitry Krotov, Yuanzhe Hu et al.

Reward-Guided Speculative Decoding (RSD) combines lightweight draft models with reward signals to improve LLM inference efficiency, reducing up to 4.4× FLOPs while boosting accuracy.

Baohao Liao, Yuhui Xu, Hanze Dong et al.

Large-scale controlled study shows data contamination inflates BLEU scores up to 30 points, especially in 8B models, with contamination timing and format affecting results.

Muhammed Yusuf Kocyigit, Eleftheria Briakou, Daniel Deutsch et al.

LLM-AutoDiff introduces a graph-based automatic prompt optimization framework using textual gradients, supporting multi-component and cyclic LLM workflows, significantly improving accuracy and efficiency.

Li Yin, Zhangyang Wang

Docling integrates DocLayNet and TableFormer for fast, resource-efficient multi-format document parsing with high accuracy.

Nikolaos Livathinos, Christoph Auer, Maksym Lysak et al.

Proposes Parametric RAG, embedding external knowledge into model parameters for improved efficiency and performance.

Weihang Su, Yichen Tang, Qingyao Ai et al.

Analyzed 28 Indian language-supporting LLMs; found significant performance disparities, especially in low-resource languages.

Aatman Vaidya, Tarunima Prabhakar, Denny George et al.

GUI-Bee employs autonomous exploration with Q-ICRL to collect environment-specific data, improving GUI action grounding in novel environments.

Yue Fan, Handong Zhao, Ruiyi Zhang et al.

Proposed EHPC leverages evaluator heads for prompt compression, reducing input length and inference latency by up to 40%.

Weizhi Fei, Xueyan Niu, Guoqing Xie et al.

Proposes a prompt engineering + Balanced RAG framework for detecting disinformation campaigns, achieving 42.4% precision and 85.1% recall on election data.

Nikos Kanakaris, Heng Ping, Xiongye Xiao et al.

Zep employs Graphiti's time-aware knowledge graph, dynamically integrating conversational and business data, achieving 94.8% accuracy and 90% latency reduction, outperforming MemGPT.

Preston Rasmussen, Pavlo Paliychuk, Travis Beauvais et al.