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.

Proposes a character prediction algorithm from subword LLMs, improving AAC accuracy; domain adaptation further enhances simple conversational text performance.

Dylan Gaines, Keith Vertanen

MiniMax-01 leverages Lightning Attention combined with MoE, reaching 4560B parameters, supporting up to 4 million tokens, matching GPT-4 performance.

MiniMax, Aonian Li, Bangwei Gong et al.

LongProc benchmarks long-context language models on long procedural generation, revealing current model limitations.

Xi Ye, Fangcong Yin, Yinghui He et al.

Introduces URSA framework with large-scale datasets MMathCoT-1M and DualMath-1.1M, combined with PS-GRPO algorithm, significantly enhancing multimodal mathematical reasoning, outperforming GPT-4o by 8.4%.

Ruilin Luo, Zhuofan Zheng, Yifan Wang et al.

Partition cover-based tokenization algorithm GREEDTOK outperforms BPE and Unigram, achieving ~3% better compression on real-world corpora and lower bits per byte in large-scale pretraining.

Jia Peng Lim, Shawn Tan, Davin Choo et al.

This study compares noise injection, multi-task learning, and layer swapping for Norwegian dialect slot and intent detection, achieving up to 97.6% accuracy.

Verena Blaschke, Felicia Körner, Barbara Plank

Proposes BeliN corpus and MultiGen model, integrating category, aspect, sentiment features, achieving BLEU 18.61 in Bengali religious news headline generation.

Md Osama, Ashim Dey, Kawsar Ahmed et al.

Introduces Sui Generis score to quantify plot diversity in LLM stories; experiments with GPT-4 and LLaMA-3 show high plot element repetition, indicating lack of creativity.

Weijia Xu, Nebojsa Jojic, Sudha Rao et al.

LE-MCTS uses PRM-guided MCTS to ensemble reasoning steps, reaching 45.2% on MATH and 71.1% on MQA.

Sungjin Park, Xiao Liu, Yeyun Gong et al.

Proposes Cache-Augmented Generation (CAG) leveraging long-context LLMs to preload knowledge, eliminating retrieval latency, and outperforming RAG in efficiency and accuracy.

Brian J Chan, Chao-Ting Chen, Jui-Hung Cheng et al.

LMFusion freezes language modules and trains dedicated image modules, achieving 20% better image understanding and 3.6% improved image generation with 50% FLOPs.

Weijia Shi, Xiaochuang Han, Chunting Zhou et al.

Efficient knowledge injection via self-distillation surpasses RAG and supervised fine-tuning.

Kalle Kujanpää, Pekka Marttinen, Harri Valpola et al.

PA-RAG introduces multi-perspective preference optimization, significantly improving RAG generators' informativeness, robustness, and citation accuracy, with an average gain of 13.97%.

Jiayi Wu, Hengyi Cai, Lingyong Yan et al.

Introduced Open Universal Arabic ASR Leaderboard to evaluate multi-dialect models and reveal generalization capabilities.

Yingzhi Wang, Anas Alhmoud, Muhammad Alqurishi

Introduces RAG-RewardBench, a benchmark evaluating 45 reward models across four challenging RAG scenarios, revealing significant limitations in current models' preference alignment.

Zhuoran Jin, Hongbang Yuan, Tianyi Men et al.

Proposed object, embedding, and VLM-based metrics achieve 0.55-0.87 correlation with human ratings for automatic image transcreation evaluation.

Simran Khanuja, Vivek Iyer, Claire He et al.

This survey reviews multimodal large language models' (MLLMs) in mathematical reasoning, focusing on benchmarks, methodologies, and challenges, with key results showing >85% accuracy.

Yibo Yan, Jiamin Su, Jianxiang He et al.

This study evaluates GPT-4 mini's in-context learning for music entity recognition, revealing significant impact of entity exposure on performance and robustness.

Simon Hachmeier, Robert Jäschke

Proposed a task-oriented dialogue system for Wolof using Rasa, cross-lingual transfer, and in-house machine translation, achieving intent classification F1 of 0.995.

Derguene Mbaye, Moussa Diallo