This study introduces a similarity-based pruning method to remove 50% of Attention layers in Llama-2-70B, achieving 48.4% speedup with only 2.4% performance loss.

Shwai He, Guoheng Sun, Zheyu Shen et al.

FIRST method improves reranking efficiency by 50% using single token decoding on the BEIR benchmark.

Revanth Gangi Reddy, JaeHyeok Doo, Yifei Xu et al.

PE-Rank leverages passage embeddings for efficient listwise reranking, reducing latency by 4.5x while maintaining performance.

Qi Liu, Bo Wang, Nan Wang et al.

Constructed PKU-SafeRLHF dataset with decoupled helpfulness and harmlessness annotations, covering 19 harm categories and 3 severity levels, to improve LLM safety alignment.

Jiaming Ji, Donghai Hong, Borong Zhang et al.

SEC-QA framework generates financial multi-document QA datasets, enhancing model evaluation accuracy.

Viet Dac Lai, Michael Krumdick, Charles Lovering et al.

EVOAGENT employs evolutionary algorithms to automatically generate diverse multi-agent systems, significantly enhancing LLM task-solving capabilities.

Siyu Yuan, Kaitao Song, Jiangjie Chen et al.

ReaL system employs parameter reallocation and MCMC-based execution plan search to accelerate RLHF training up to 3.58×.

Zhiyu Mei, Wei Fu, Kaiwei Li et al.

VGA reduces hallucinations in GUI understanding via image-centric fine-tuning using a 63.8k dataset and FAC method.

Ziyang Meng, Yu Dai, Zezheng Gong et al.

EAGER introduces a two-stream generative recommender integrating behavior and semantic info, achieving superior performance.

Ye Wang, Jiahao Xun, Minjie Hong et al.

Introduces a lexically grounded subword segmentation method, improving morphological plausibility.

Jindřich Libovický, Jindřich Helcl

HiFiAlbedo estimates high-fidelity facial albedo from a single image via texture quantization.

Zimin Ran, Xingyu Ren, Xiang An et al.

LayerMerge jointly prunes and merges convolutional and activation layers, achieving 1.36× speed-up with minimal accuracy loss.

Jinuk Kim, Marwa El Halabi, Mingi Ji et al.

ChatGLM-4, based on Transformer, trained on 10 trillion tokens, rivals GPT-4, supports multi-tool autonomous decision-making.

Team GLM, :, Aohan Zeng et al.

Proposes RecBLR with behavior-dependent linear recurrent units, achieving high performance, training speed, and low inference cost.

Chengkai Liu, Jianghao Lin, Hanzhou Liu et al.

MOYU study reveals potential of dynamic activation in LLMs for inference speedup.

Chi Ma, Mincong Huang, Chao Wang et al.

DART-Math significantly improves math problem-solving via difficulty-aware rejection tuning using small open-source models.

Yuxuan Tong, Xiwen Zhang, Rui Wang et al.

Joint channel, layer, and block pruning via MINLP with accurate latency modeling achieves high pruning ratios and improved speed/accuracy trade-offs.

Xinglong Sun, Barath Lakshmanan, Maying Shen et al.

Satyrn combines structured data analysis with LLMs, achieving 86% factual accuracy in report generation.

Marko Sterbentz, Cameron Barrie, Shubham Shahi et al.

Proposes diffusion-based autoregressive image generation without vector quantization, achieving FID<3.5 on ImageNet 256×256.

Tianhong Li, Yonglong Tian, He Li et al.

Introduces DataComp-LM, a benchmark with 240T tokens, using model-based filtering to improve training data quality, achieving 64% MMLU 5-shot accuracy with a 7B model.

Jeffrey Li, Alex Fang, Georgios Smyrnis et al.