TAP4LLM enhances large-scale table reasoning by sampling, augmenting, and serializing data, achieving a 7.93% performance boost.

Yuan Sui, Jiaru Zou, Mengyu Zhou et al.

Proposes LLM360 framework, fully open-sourcing 7B models, training code, data, checkpoints, and analysis for transparency.

Zhengzhong Liu, Aurick Qiao, Willie Neiswanger et al.

EQ-Bench assesses LLMs' emotional understanding via dialogue-based emotion intensity ratings, achieving a correlation of 0.97 with broad intelligence benchmarks.

Samuel J. Paech

Llama Guard employs LLM-based multi-class classification with a safety taxonomy, achieving high performance (AUPRC 0.945) in content moderation tasks.

Hakan Inan, Kartikeya Upasani, Jianfeng Chi et al.

Proposes LLMCompiler for parallel function calls, achieving up to 3.7× speedup and 6.7× cost reduction.

Sehoon Kim, Suhong Moon, Ryan Tabrizi et al.

CritiqueLLM employs Eval-Instruct with pseudo-references to train a critique model, achieving near GPT-4 performance in fine-grained evaluation of LLM outputs, outperforming open-source baselines.

Pei Ke, Bosi Wen, Zhuoer Feng et al.

Introduced MMMU, a benchmark with 11.5K multimodal questions across 6 disciplines, challenging models in expert-level perception and reasoning.

Xiang Yue, Yuansheng Ni, Kai Zhang et al.

First survey of legal LLMs: Transformer/ChatGPT promise efficiency, but data, ethics, and due process remain bottlenecks.

Jinqi Lai, Wensheng Gan, Jiayang Wu et al.

ATLANTIC integrates heterogeneous graph neural networks into retrieval-augmented language models, improving scientific passage relevance and faithfulness by 15%.

Sai Munikoti, Anurag Acharya, Sridevi Wagle et al.

FinanceBench evaluates LLMs on financial QA; GPT-4-Turbo with retrieval answers only 19%, oracle 85%.

Pranab Islam, Anand Kannappan, Douwe Kiela et al.

ARES: Automated framework improves RAG evaluation accuracy, reduces manual annotations.

Jon Saad-Falcon, Omar Khattab, Christopher Potts et al.

Proposes MMC-Instruction dataset and MMCA model, significantly improving chart understanding with 600k data and nine-task benchmark.

Fuxiao Liu, Xiaoyang Wang, Wenlin Yao et al.

Proposes grounding acts as metrics to evaluate dialogue, finds LLMs generate 77.5% fewer grounding acts than humans, and preference tuning reduces these acts further.

Omar Shaikh, Kristina Gligorić, Ashna Khetan et al.

Integrating neural quality metrics (COMET-QE) via data filtering, RL training, and reranking significantly improves NMT performance, with experimental gains of up to 3.2 BLEU points.

Miguel Moura Ramos, Patrick Fernandes, António Farinhas et al.

GRASP benchmark evaluates multimodal LLMs' grounding and physics understanding, revealing poor performance in intuitive physics tests.

Serwan Jassim, Mario Holubar, Annika Richter et al.

HeLM uses evidence highlighting and summarization modules with LLaMA2, achieving SOTA ROUGE/BLEU scores in table-to-text tasks.

Junyi Bian, Xiaolei Qin, Wuhe Zou et al.

This study systematically evaluates dialogue summarization models' robustness against natural variations, revealing instruction-tuned models are more sensitive, with performance drops over 30%.

Ankita Gupta, Chulaka Gunasekara, Hui Wan et al.

Proposes DPO-based unsupervised fine-tuning to mimic MBR decoding, significantly improving translation quality without extra inference cost.

Guangyu Yang, Jinghong Chen, Weizhe Lin et al.

AMBER: An LLM-free multi-dimensional benchmark for evaluating MLLM hallucinations, enhancing reliability.

Junyang Wang, Yuhang Wang, Guohai Xu et al.

Proposes context consistency training to address training-testing context mismatch in SiMT, improving BLEU by over 3 points and latency.

Meizhi Zhong, Lemao Liu, Kehai Chen et al.