Proposes BBT-RGB, combining two-stage derivative-free optimization, multi-verbalizer construction, and instruction search to enhance black-box prompt tuning.

Qiushi Sun, Chengcheng Han, Nuo Chen et al.

TinyStories dataset enables training sub-10M parameter models to generate coherent stories, evaluated via GPT-4 multi-dimensional scoring.

Ronen Eldan, Yuanzhi Li

WebCPM introduces interactive web search with behavior imitation, constructing a Chinese long-form QA dataset, achieving 32.5% answer quality comparable to humans.

Yujia Qin, Zihan Cai, Dian Jin et al.

VCSUM creates a large-scale Chinese meeting dataset with topic segmentation, multi-granularity summaries, and salient sentence annotations, enabling multi-task learning.

Han Wu, Mingjie Zhan, Haochen Tan et al.

SUR-adapter transfers LLM semantic reasoning to diffusion models, improving simple prompt image quality by over 15%.

Shanshan Zhong, Zhongzhan Huang, Wushao Wen et al.

Identifies systematic unfaithfulness in chain-of-thought (CoT) explanations, showing bias features cause up to 36% accuracy drops, with explanations often rationalizing biased answers.

Miles Turpin, Julian Michael, Ethan Perez et al.

BIRD benchmark with 12,751 text-SQL pairs over 95 large databases; GPT-4 achieves only 54.89% execution accuracy, highlighting challenges in large-scale value understanding.

Jinyang Li, Binyuan Hui, Ge Qu et al.

Proposed SG-CQG framework uses semantic graphs and explicit control signals to enhance answer-unaware conversational question generation, achieving SOTA results.

Xuan Long Do, Bowei Zou, Shafiq Joty et al.

ProTeGi uses gradient-inspired natural language feedback and beam search to optimize prompts, improving task performance by up to 31%.

Reid Pryzant, Dan Iter, Jerry Li et al.

Large Language Models (LLMs) can replace human evaluations, showing consistent results with human experts.

Cheng-Han Chiang, Hung-yi Lee

mPLUG-Owl enhances large language models' multimodal capabilities through modular learning, significantly improving instruction and visual understanding.

Qinghao Ye, Haiyang Xu, Guohai Xu et al.

Citizen Science recruited 98 volunteers to re-annotate 1481 instances in NLP, achieving high-quality results comparable to expert labels.

Jan-Christoph Klie, Ji-Ung Lee, Kevin Stowe et al.

Proposes Evol-Instruct for automatic multi-level instruction generation, significantly boosting LLaMA's performance.

Can Xu, Qingfeng Sun, Kai Zheng et al.

Using natural logic intermediate features, the study applies Amnesic and Mnestic interventions to analyze causal roles in high-dimensional model representations, revealing limitations and improvements.

Julia Rozanova, Marco Valentino, Lucas Cordeiro et al.

Proposes citation recall and precision metrics to evaluate verifiability; finds only 51.5% sentences fully supported in top search engines.

Nelson F. Liu, Tianyi Zhang, Percy Liang

Behavior Expectation Bounds (BEB) framework reveals fundamental limits of LLM alignment against adversarial prompts.

Yotam Wolf, Noam Wies, Oshri Avnery et al.

API-Bank benchmark evaluates tool-augmented LLMs with 73 APIs, boosting tool utilization by training and testing models like Lynx, GPT-3.5, GPT-4.

Minghao Li, Yingxiu Zhao, Bowen Yu et al.

This study evaluates pre-trained LLMs in task-oriented dialogue, showing limited belief state tracking but effective dialogue guidance, improved by true belief states and few-shot examples.

Vojtěch Hudeček, Ondřej Dušek

AGIEval benchmark assesses foundation models on human-standard exams; GPT-4 surpasses average human scores with 95% in SAT Math and 92.5% in Chinese college entrance English.

Wanjun Zhong, Ruixiang Cui, Yiduo Guo et al.

Evaluates ChatGPT's zero-shot performance across 37 languages and 7 tasks, revealing limitations in multilingual capabilities.

Viet Dac Lai, Nghia Trung Ngo, Amir Pouran Ben Veyseh et al.