Introduces ContextualJudgeBench benchmark, evaluating LLM judges' performance in contextual settings, with top accuracy only reaching 55%.

Austin Xu, Srijan Bansal, Yifei Ming et al.

AlignX leverages 1.3 million preference samples to enhance personalized LLM alignment, achieving 17.06% accuracy gain.

Jia-Nan Li, Jian Guan, Songhao Wu et al.

HDLCoRe uses prompt engineering and retrieval augmentation to improve HDL code generation without fine-tuning, reducing hallucinations by 28%.

Heng Ping, Shixuan Li, Peiyu Zhang et al.

Introduced a multimodal sarcasm analysis framework, evaluating 12 LVLMs over 2409 samples.

Junjie Chen, Xuyang Liu, Subin Huang et al.

Proposes MoC framework with Boundary Clarity and Chunk Stickiness metrics, significantly improving text chunking and retrieval-augmented generation performance.

Jihao Zhao, Zhiyuan Ji, Zhaoxin Fan et al.

Proposes Plan-and-Act framework with synthetic data augmentation, achieving 57.58% success in long-horizon web tasks.

Lutfi Eren Erdogan, Nicholas Lee, Sehoon Kim et al.

Search-R1 employs reinforcement learning enabling LLMs to generate multi-turn search queries, boosting QA performance by 41%.

Bowen Jin, Hansi Zeng, Zhenrui Yue et al.

OpenRAG end-to-end tuning improves retrieval relevance by 4.0%, surpassing SOTA retrievers by 2.1%, using contrastive learning and approximate labels.

Jiawei Zhou, Lei Chen

Proposes a knowledge-centric RAG framework emphasizing knowledge lifecycle management for improved NLP tasks.

Mingyue Cheng, Yucong Luo, Jie Ouyang et al.

LMM-R1 employs a two-stage rule-based RL, boosting 3B multimodal models' reasoning by 4.83% on average, enabling effective transfer from text to vision tasks.

Yingzhe Peng, Gongrui Zhang, Miaosen Zhang et al.

START enhances reasoning by integrating tools, achieving 63.6% accuracy on GPQA.

Chengpeng Li, Mingfeng Xue, Zhenru Zhang et al.

Proposes dual-objective optimization (DOOR/W-DOOR) to improve LLM safety alignment, significantly reducing jailbreak success rates while maintaining utility.

Xuandong Zhao, Will Cai, Tianneng Shi et al.

Using Sparse Autoencoders (SAE) to extract interpretable features from Gemma-2-2b residual streams enhances AI text detection and interpretability.

Kristian Kuznetsov, Laida Kushnareva, Polina Druzhinina et al.

The GA-Rollback framework employs generator-assistant collaboration with stepwise rollback to significantly improve LLM agent reasoning accuracy, achieving up to 10.5× success rate improvements on benchmark tasks.

Xingzuo Li, Kehai Chen, Yunfei Long et al.

EpicPRM framework constructs high-quality math reasoning dataset Epic50k using adaptive binary search, enhancing PRM performance.

Wei Sun, Qianlong Du, Fuwei Cui et al.

This paper introduces ToolRet, a benchmark for large language model tool retrieval, revealing poor IR model performance and proposing 200k training samples for improvement.

Zhengliang Shi, Yuhan Wang, Lingyong Yan et al.

SePer measures retrieval utility via semantic perplexity reduction, showing high correlation with human judgments and outperforming traditional metrics.

Lu Dai, Yijie Xu, Jinhui Ye et al.

GAS(P) evaluates gender bias in unprompted generated text, revealing persistent biases across models.

Jennifer Mickel, Maria De-Arteaga, Leqi Liu et al.

This study uses translation chains to reveal iterative generation causes information distortion, with key metrics showing increased deviation over rounds.

Amr Mohamed, Mingmeng Geng, Michalis Vazirgiannis et al.

Bi'an introduces a bilingual benchmark and lightweight models, achieving 88.2% accuracy in RAG hallucination detection, close to GPT-4o performance.

Zhouyu Jiang, Mengshu Sun, Zhiqiang Zhang et al.