Systematic psycholinguistic evaluation of language models reveals partial sensitivity to argument roles, with models distinguishing plausible from implausible verbs but lacking human-like selective patterns.

Eun-Kyoung Rosa Lee, Sathvik Nair, Naomi Feldman

Enhance Dense Passage Retrieval with Entailment Tuning, achieving 3% improvement on NQ dataset.

Lu Dai, Hao Liu, Hui Xiong

MiCEval introduces a fine-grained, multi-metric framework for automated evaluation of multimodal chain-of-thought, outperforming cosine similarity and fine-tuning metrics.

Xiongtao Zhou, Jie He, Lanyu Chen et al.

GenEOL leverages LLMs to generate diverse sentence variations, improving embeddings by averaging, with an average gain of 2.85 points on STS benchmarks.

Raghuveer Thirukovalluru, Bhuwan Dhingra

A unified pizza–sandwich–lasagna framework shows that 2× KV-cache reduction usually preserves quality, while aggressive compression favors sandwich-middle.

You Wu, Haoyi Wu, Kewei Tu

MemTree employs a dynamic tree-structured memory, significantly improving long-term memory management for LLMs, excelling in multi-turn dialogue and document reasoning.

Alireza Rezazadeh, Zichao Li, Wei Wei et al.

This study evaluates 24 multilingual LLMs within a real-world Indian healthcare chatbot setting using a unified RAG framework, focusing on factual correctness and cultural relevance.

Varun Gumma, Ananditha Raghunath, Mohit Jain et al.

OpenAI's o1 model excels in reasoning tasks like math and coding using Test-time Compute methods.

Siwei Wu, Zhongyuan Peng, Xinrun Du et al.

DeCo dynamically integrates preceding-layer knowledge into decoding, reducing hallucination rates by 10.8% in multimodal large language models.

Chenxi Wang, Xiang Chen, Ningyu Zhang et al.

Li and Zhou reveal that MoE models' routing weights (RW) serve as effective embeddings without fine-tuning, outperforming hidden states (HS) when combined.

Ziyue Li, Tianyi Zhou

FlatQuant employs learnable affine transformations with Kronecker decomposition to achieve sub-1% accuracy loss in W4A4 quantization of LLaMA-3-70B, boosting speed by 2.3x.

Yuxuan Sun, Ruikang Liu, Haoli Bai et al.

EGAD enhances Longformer with keyword-based global attention, boosting long-range dependency modeling for abstractive summarization.

Evan Lucas, Dylan Kangas, Timothy C Havens

DRAFT employs self-driven feedback to iteratively refine tool documentation, enhancing LLM understanding and utilization.

Changle Qu, Sunhao Dai, Xiaochi Wei et al.

Herald: Translates Mathlib4 to natural language using dual augmentation, enhancing LLM performance in mathematical reasoning.

Guoxiong Gao, Yutong Wang, Jiedong Jiang et al.

TorchTitan supports 4D parallelism, boosting Llama 3.1 training by 65.08% with modular design and hardware co-optimization.

Wanchao Liang, Tianyu Liu, Less Wright et al.

Proposes MoLAS and AgentSquare, using module evolution and recombination to optimize LLM agents with a 17.2% performance boost.

Yu Shang, Yu Li, Keyu Zhao et al.

Casablanca created a 48-hour multi-dialect Arabic speech dataset with manual annotations, evaluating multilingual models’ adaptability and performance.

Bashar Talafha, Karima Kadaoui, Samar Mohamed Magdy et al.

MetricX-24 employs multi-stage training and synthetic data, significantly outperforming MetricX-23 in translation quality assessment.

Juraj Juraska, Daniel Deutsch, Mara Finkelstein et al.

X-ALMA employs plug-and-play modules and ARPO optimization to achieve top translation quality across 50 languages, outperforming Aya-101 and Aya-23 on FLORES-200 and WMT'23 datasets.

Haoran Xu, Kenton Murray, Philipp Koehn et al.

HELMET offers seven task categories to evaluate long-context language models, improving existing benchmarks.

Howard Yen, Tianyu Gao, Minmin Hou et al.