Goldfish: a suite of 1000+ monolingual models for 350 low-resource languages, trained with 125M parameters, outperforming large multilingual models in perplexity and grammaticality.

Tyler A. Chang, Catherine Arnett, Zhuowen Tu et al.

Anim-Director employs GPT-4-driven multimodal models for fully autonomous, controllable animation video generation from simple narratives.

Yunxin Li, Haoyuan Shi, Baotian Hu et al.

ConVerSum employs contrastive learning for data-efficient cross-lingual summarization, outperforming baselines in low-resource settings.

Sanzana Karim Lora, M. Sohel Rahman, Rifat Shahriyar

Math-PUMA uses 692K alignment pairs and three-stage training to reduce the text–vision gap in mathematical reasoning.

Wenwen Zhuang, Xin Huang, Xiantao Zhang et al.

Introduces OpenEP task and OpenEPBench dataset, combined with StkFEP framework, for open-ended future event prediction with multi-perspective, diverse outputs.

Yong Guan, Hao Peng, Xiaozhi Wang et al.

Proposes Speculative Diffusion Decoding (SpecDiff), achieving up to 7.2× speedup by integrating discrete diffusion models for parallel sequence generation.

Jacob K Christopher, Brian R Bartoldson, Tal Ben-Nun et al.

SWIFT supports 550+ LLMs and 200+ MLLMs, integrating multiple training techniques for scalable, efficient fine-tuning.

Yuze Zhao, Jintao Huang, Jinghan Hu et al.

Proposes multi-agent LLM framework for automated data storytelling, with 1494 real-world examples.

Mohammed Saidul Islam, Md Tahmid Rahman Laskar, Md Rizwan Parvez et al.

ToolSandbox is a stateful, conversational benchmark for evaluating LLM tool use, emphasizing implicit state dependencies and dynamic milestone tracking.

Jiarui Lu, Thomas Holleis, Yizhe Zhang et al.

Fine-tuning small models improves their Theory of Mind task performance, achieving 46% alignment with large models.

Nunzio Lore, Sepehr Ilami, Babak Heydari

Study finds multimodal LLMs in GUI environments are susceptible to environmental distractions, affecting their faithfulness.

Xinbei Ma, Yiting Wang, Yao Yao et al.

This study compares GPT-4, Llama 2, and Gemini in zero-shot tasks across English and low-resource South Asian languages, revealing significant performance gaps.

Md. Arid Hasan, Prerona Tarannum, Krishno Dey et al.

Negative Attention Score Alignment (NASA) reduces negative bias in large language models, enhancing precision and recall.

Sangwon Yu, Jongyoon Song, Bongkyu Hwang et al.

Meta-Reward introduces self-judgment and meta-judgment to improve LLMs, achieving 39.4% win rate on AlpacaEval 2, surpassing previous methods.

Tianhao Wu, Weizhe Yuan, Olga Golovneva et al.

This study compares RAG and long-context LLMs, proposing Self-Route for cost-effective hybrid routing, achieving performance close to long-context models.

Zhuowan Li, Cheng Li, Mingyang Zhang et al.

Proposes three methods—IsoFLOPs analysis, derivative estimation, and loss fitting—to predict optimal vocabulary size for LLMs, finding larger models need bigger vocabularies.

Chaofan Tao, Qian Liu, Longxu Dou et al.

Introduces LMMS-EVAL, LMMS-EVAL LITE, and LIVEBENCH, covering 50+ tasks, to address the evaluation trilemma of large multimodal models.

Kaichen Zhang, Bo Li, Peiyuan Zhang et al.

Case2Code task enhances code generation via large-scale synthetic data.

Yunfan Shao, Linyang Li, Yichuan Ma et al.

NeedleBench evaluates LLMs' retrieval and reasoning across sparse and dense contexts, revealing weaknesses in dense reasoning tasks.

Mo Li, Songyang Zhang, Taolin Zhang et al.

Using MidEast-TE-mini, the study compares graph, text, and mixed LLM forecasters; GPT-4 extraction reaches 72.60%, while tuning and RAG help most.

He Chang, Chenchen Ye, Zhulin Tao et al.