APIGen-MT employs a two-phase blueprint validation and simulated interaction to generate high-quality multi-turn data, surpassing GPT-4o and Claude 3.5.

Akshara Prabhakar, Zuxin Liu, Ming Zhu et al.

Introduces KnowSelf for LLMs' situational self-awareness, using two-stage training to enhance planning with minimal external knowledge.

Shuofei Qiao, Zhisong Qiu, Baochang Ren et al.

Diverse in-context example selection after program decomposition boosts semantic parsing accuracy.

Mayank Kothyari, Sunita Sarawagi, Soumen Chakrabarti et al.

AD-GPT combines Llama3 and BERT to enhance Alzheimer's disease information retrieval accuracy.

Ziyu Liu, Lintao Tang, Zeliang Sun et al.

HyperRAG leverages KV-cache reuse to boost RAG reranker efficiency by 2-3x, maintaining high quality.

Yuwei An, Yihua Cheng, Seo Jin Park et al.

MultiBLiMP 1.0 employs an automated pipeline leveraging UD and UniMorph to generate over 128,000 minimal pairs across 101 languages, evaluating large language models' formal syntactic competence.

Jaap Jumelet, Leonie Weissweiler, Joakim Nivre et al.

Proposes multi-layered memory combining text, KV cache, parameters, and hidden states to enhance long-term memory in LLMs.

Lianlei Shan, Shixian Luo, Zezhou Zhu et al.

Leveraging LLMs with prompting and fine-tuning enhances machine translation, especially for low-resource languages and document-level tasks.

Baban Gain, Dibyanayan Bandyopadhyay, Asif Ekbal et al.

InfiniteICL breaks context window limits via long short-term memory transformation, reducing context length by 90% and achieving 103% performance.

Bowen Cao, Deng Cai, Wai Lam

Introduces Causal Intervened Likelihood (CIL) for assessing inferability in future forecasting, building the PROPHET benchmark.

Zhengwei Tao, Pu Wu, Zhi Jin et al.

BoundlessBPE breaks pre-tokenization limits, boosting bytes/token by up to 15%.

Craig W. Schmidt, Varshini Reddy, Chris Tanner et al.

Proposes a four-dimensional framework for TTS, including when, how, where, and how well, advancing LLM reasoning capabilities.

Qiyuan Zhang, Fuyuan Lyu, Zexu Sun et al.

JudgeLRM employs reinforcement learning with outcome-driven rewards to surpass SFT models, with 7B/8B models exceeding GPT-4 in F1 scores.

Nuo Chen, Zhiyuan Hu, Qingyun Zou et al.

Proposes a methodology-centered taxonomy for LLM agents, covering construction, collaboration, and evolution, advancing towards artificial general intelligence.

Junyu Luo, Weizhi Zhang, Ye Yuan et al.

Proposes R-PRM, leveraging stronger models for seed data, preference optimization, and multi-trajectory inference, boosting step evaluation F1 by 11.9 points.

Shuaijie She, Junxiao Liu, Yifeng Liu et al.

Using Llama-3.1-8B, 73% of responses were highly relevant with 100% recall.

Julia Ive, Felix Jozsa, Nick Jackson et al.

Gemma 3 introduces multimodal, 128K tokens long context, and vision understanding, using architecture improvements and distillation, outperforming Gemma 2.

Gemma Team, Aishwarya Kamath, Johan Ferret et al.

AdaptiVocab reduces token usage by over 25% via lightweight vocabulary adaptation, boosting efficiency in low-resource domains.

Itay Nakash, Nitay Calderon, Eyal Ben David et al.

This paper introduces JudgeAnything, a benchmark leveraging multimodal large language models (MLLMs) as automated judges across 15 diverse tasks, revealing strong performance in understanding but limitations in generation.

Shu Pu, Yaochen Wang, Dongping Chen et al.

Proposes multi-strategy methods—model pruning, output control, prompt engineering—to reduce reasoning steps, cutting costs while maintaining accuracy.

Yang Sui, Yu-Neng Chuang, Guanchu Wang et al.