DeCoRe : Décodage par Contraste des Têtes de Récupération pour Atténuer les Hallucinations

DeCoRe: Decoding by Contrasting Retrieval Heads to Mitigate Hallucinations

October 24, 2024
Auteurs: Aryo Pradipta Gema, Chen Jin, Ahmed Abdulaal, Tom Diethe, Philip Teare, Beatrice Alex, Pasquale Minervini, Amrutha Saseendran
cs.AI

Résumé

Les grands modèles de langage (LLM) hallucinent souvent, produisant des sorties non fidèles ou incorrectes sur le plan factuel en déformant le contexte fourni ou en rappelant incorrectement des connaissances internes. Des études récentes ont identifié des têtes d'attention spécifiques au sein de l'architecture Transformer, appelées têtes de récupération, responsables de l'extraction d'informations contextuelles pertinentes. Nous émettons l'hypothèse que le masquage de ces têtes de récupération peut induire des hallucinations et que la comparaison des sorties du LLM de base et du LLM masqué peut réduire les hallucinations. À cette fin, nous proposons Décodage par Contraste des Têtes de Récupération (DeCoRe), une nouvelle stratégie de décodage sans entraînement qui amplifie les informations trouvées dans le contexte et les paramètres du modèle. DeCoRe atténue les réponses potentiellement hallucinées en contrastant dynamiquement les sorties du LLM de base et du LLM masqué, en utilisant l'entropie conditionnelle comme guide. Nos expériences approfondies confirment que DeCoRe améliore significativement les performances sur des tâches nécessitant une fidélité contextuelle élevée, telles que la résumé (XSum de 18,6 %), le suivi des instructions (MemoTrap de 10,9 %) et la réponse à des questions ouvertes (NQ-Open de 2,4 % et NQ-Swap de 5,5 %).
English
Large Language Models (LLMs) often hallucinate, producing unfaithful or factually incorrect outputs by misrepresenting the provided context or incorrectly recalling internal knowledge. Recent studies have identified specific attention heads within the Transformer architecture, known as retrieval heads, responsible for extracting relevant contextual information. We hypothesise that masking these retrieval heads can induce hallucinations and that contrasting the outputs of the base LLM and the masked LLM can reduce hallucinations. To this end, we propose Decoding by Contrasting Retrieval Heads (DeCoRe), a novel training-free decoding strategy that amplifies information found in the context and model parameters. DeCoRe mitigates potentially hallucinated responses by dynamically contrasting the outputs of the base LLM and the masked LLM, using conditional entropy as a guide. Our extensive experiments confirm that DeCoRe significantly improves performance on tasks requiring high contextual faithfulness, such as summarisation (XSum by 18.6%), instruction following (MemoTrap by 10.9%), and open-book question answering (NQ-Open by 2.4% and NQ-Swap by 5.5%).

Summary

AI-Generated Summary

Paper Overview

This literature introduces DeCoRe, a decoding strategy for language models that contrasts base model outputs with masked retrieval heads to enhance contextual fidelity and reduce hallucinations. DeCoRe significantly improves model accuracy in tasks requiring contextual faithfulness and factual recall. It outperforms baselines in various tasks like summarization, instruction-following, and open-book question answering.

Core Contribution

DeCoRe introduces a novel decoding strategy that leverages masked retrieval heads and dynamic entropy-controlled contrastive decoding to enhance contextual fidelity and reduce hallucinations in language models.

Research Context

The study positions itself within the realm of large language models (LLMs) and decoding strategies, focusing on improving contextual fidelity, factuality, and reasoning capabilities while mitigating hallucinations.

Keywords

DeCoRe, Language Models, Contextual Fidelity, Factuality, Hallucinations, Contrastive Decoding, Masked Retrieval Heads

Background

This research addresses the challenges faced by large language models in maintaining contextual fidelity and reducing hallucinations by proposing the DeCoRe decoding strategy. The rationale lies in the need to enhance model performance in tasks requiring accurate contextual understanding and factual recall.

Research Gap

Existing literature lacks efficient strategies to mitigate hallucinations and improve contextual fidelity in language models, especially in tasks like summarization and open-book question answering.

Technical Challenges

The primary technical obstacles include reducing hallucinations, enhancing contextual fidelity, and improving factuality in language model outputs while maintaining high performance across various tasks.

Prior Approaches

Previous solutions have focused on internal mechanisms of LLMs and constrained decoding methods but have not adequately addressed the challenges of hallucinations and contextual fidelity.

Methodology

The research methodology involves studying the impact of masked retrieval heads on language model performance, implementing the DeCoRe strategy without training, and evaluating its effectiveness in various tasks.

Theoretical Foundation

DeCoRe is based on the principle of contrasting base model outputs with masked retrieval heads to improve contextual fidelity and reduce hallucinations in language model generations.

Technical Architecture

DeCoRe utilizes masked retrieval heads and dynamic entropy-controlled contrastive decoding to enhance model outputs' contextual fidelity and reduce hallucinations.

Implementation Details

The study implements DeCoRe without training, focusing on the contrast between base model predictions and masked model outputs to improve contextual fidelity and factuality.

Innovation Points

DeCoRe introduces a novel approach to decoding in language models, significantly improving contextual fidelity, factuality, and reasoning capabilities across various tasks.

Experimental Validation

The experimental validation involves evaluating DeCoRe's performance in tasks like summarization, instruction-following, and question answering to demonstrate its effectiveness in enhancing contextual fidelity and reducing hallucinations.

Setup

Exact configurations, parameters, and datasets like NQ-Swap, NQ-Open, and MuSiQue are used to evaluate model performance in various tasks requiring factual recall and contextual fidelity.

Metrics

Evaluation criteria include metrics like EM scores, conditional entropy, and factuality assessments to quantify the improvements in contextual fidelity and factuality achieved by DeCoRe.

Results

Quantitative and qualitative findings show significant improvements in contextual fidelity, factuality, and reasoning capabilities of language models using DeCoRe compared to baselines.

Comparative Analysis

DeCoRe outperforms traditional strategies like DoLA and static contrastive decoding in tasks like summarization, instruction-following, and open-book question answering, showcasing its superiority in enhancing model performance.

Impact and Implications

The study's findings have significant implications for improving language model performance in tasks requiring contextual fidelity, factuality, and reasoning capabilities while reducing hallucinations.

Key Findings

DeCoRe significantly enhances contextual fidelity, factuality, and reasoning capabilities of language models, leading to improved performance in various tasks like summarization and open-book question answering.

Limitations

The study acknowledges limitations in certain tasks and the need for further research to optimize DeCoRe's performance across a broader range of language model applications.

Future Directions

Concrete research opportunities include exploring the application of DeCoRe in sensitive domains, refining the contrastive decoding process, and enhancing model performance in tasks requiring high contextual fidelity.

Practical Significance

DeCoRe offers practical applications in improving language model outputs' accuracy, especially in tasks requiring accurate contextual understanding, factuality, and reasoning capabilities.

Articles en Vedette

DeepSeek-R1 : Encourager la capacité de raisonnement dans les LLMs via l'apprentissage par renforcement
DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, Aixin Liu, Bing Xue, Bingxuan Wang, Bochao Wu, Bei Feng, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, Damai Dai, Deli Chen, Dongjie Ji, Erhang Li, Fangyun Lin, Fucong Dai, Fuli Luo, Guangbo Hao, Guanting Chen, Guowei Li, H. Zhang, Han Bao, Hanwei Xu, Haocheng Wang, Honghui Ding, Huajian Xin, Huazuo Gao, Hui Qu, Hui Li, Jianzhong Guo, Jiashi Li, Jiawei Wang, Jingchang Chen, Jingyang Yuan, Junjie Qiu, Junlong Li, J. L. Cai, Jiaqi Ni, Jian Liang, Jin Chen, Kai Dong, Kai Hu, Kaige Gao, Kang Guan, Kexin Huang, Kuai Yu, Lean Wang, Lecong Zhang, Liang Zhao, Litong Wang, Liyue Zhang, Lei Xu, Leyi Xia, Mingchuan Zhang, Minghua Zhang, Minghui Tang, Meng Li, Miaojun Wang, Mingming Li, Ning Tian, Panpan Huang, Peng Zhang, Qiancheng Wang, Qinyu Chen, Qiushi Du, Ruiqi Ge, Ruisong Zhang, Ruizhe Pan, Runji Wang, R. J. Chen, R. L. Jin, Ruyi Chen, Shanghao Lu, Shangyan Zhou, Shanhuang Chen, Shengfeng Ye, Shiyu Wang, Shuiping Yu, Shunfeng Zhou, Shuting Pan, S. S. Li, Shuang Zhou, Shaoqing Wu, Shengfeng Ye, Tao Yun, Tian Pei, Tianyu Sun, T. Wang, Wangding Zeng, Wanjia Zhao, Wen Liu, Wenfeng Liang, Wenjun Gao, Wenqin Yu, Wentao Zhang, W. L. Xiao, Wei An, Xiaodong Liu, Xiaohan Wang, Xiaokang Chen, Xiaotao Nie, Xin Cheng, Xin Liu, Xin Xie, Xingchao Liu, Xinyu Yang, Xinyuan Li, Xuecheng Su, Xuheng Lin, X. Q. Li, Xiangyue Jin, Xiaojin Shen, Xiaosha Chen, Xiaowen Sun, Xiaoxiang Wang, Xinnan Song, Xinyi Zhou, Xianzu Wang, Xinxia Shan, Y. K. Li, Y. Q. Wang, Y. X. Wei, Yang Zhang, Yanhong Xu, Yao Li, Yao Zhao, Yaofeng Sun, Yaohui Wang, Yi Yu, Yichao Zhang, Yifan Shi, Yiliang Xiong, Ying He, Yishi Piao, Yisong Wang, Yixuan Tan, Yiyang Ma, Yiyuan Liu, Yongqiang Guo, Yuan Ou, Yuduan Wang, Yue Gong, Yuheng Zou, Yujia He, Yunfan Xiong, Yuxiang Luo, Yuxiang You, Yuxuan Liu, Yuyang Zhou, Y. X. Zhu, Yanhong Xu, Yanping Huang, Yaohui Li, Yi Zheng, Yuchen Zhu, Yunxian Ma, Ying Tang, Yukun Zha, Yuting Yan, Z. Z. Ren, Zehui Ren, Zhangli Sha, Zhe Fu, Zhean Xu, Zhenda Xie, Zhengyan Zhang, Zhewen Hao, Zhicheng Ma, Zhigang Yan, Zhiyu Wu, Zihui Gu, Zijia Zhu, Zijun Liu, Zilin Li, Ziwei Xie, Ziyang Song, Zizheng Pan, Zhen Huang, Zhipeng Xu, Zhongyu Zhang, Zhen ZhangJan 22, 20253735

Rapport technique de Qwen2.5
Qwen2.5 Technical Report

Qwen, An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, Le Yu, Mei Li, Mingfeng Xue, Pei Zhang, Qin Zhu, Rui Men, Runji Lin, Tianhao Li, Tingyu Xia, Xingzhang Ren, Xuancheng Ren, Yang Fan, Yang Su, Yichang Zhang, Yu Wan, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, Zihan QiuDec 19, 202436311

MiniMax-01 : Mise à l'échelle des modèles de base avec Attention Éclair.
MiniMax-01: Scaling Foundation Models with Lightning Attention

MiniMax, Aonian Li, Bangwei Gong, Bo Yang, Boji Shan, Chang Liu, Cheng Zhu, Chunhao Zhang, Congchao Guo, Da Chen, Dong Li, Enwei Jiao, Gengxin Li, Guojun Zhang, Haohai Sun, Houze Dong, Jiadai Zhu, Jiaqi Zhuang, Jiayuan Song, Jin Zhu, Jingtao Han, Jingyang Li, Junbin Xie, Junhao Xu, Junjie Yan, Kaishun Zhang, Kecheng Xiao, Kexi Kang, Le Han, Leyang Wang, Lianfei Yu, Liheng Feng, Lin Zheng, Linbo Chai, Long Xing, Meizhi Ju, Mingyuan Chi, Mozhi Zhang, Peikai Huang, Pengcheng Niu, Pengfei Li, Pengyu Zhao, Qi Yang, Qidi Xu, Qiexiang Wang, Qin Wang, Qiuhui Li, Ruitao Leng, Shengmin Shi, Shuqi Yu, Sichen Li, Songquan Zhu, Tao Huang, Tianrun Liang, Weigao Sun, Weixuan Sun, Weiyu Cheng, Wenkai Li, Xiangjun Song, Xiao Su, Xiaodong Han, Xinjie Zhang, Xinzhu Hou, Xu Min, Xun Zou, Xuyang Shen, Yan Gong, Yingjie Zhu, Yipeng Zhou, Yiran Zhong, Yongyi Hu, Yuanxiang Fan, Yue Yu, Yufeng Yang, Yuhao Li, Yunan Huang, Yunji Li, Yunpeng Huang, Yunzhi Xu, Yuxin Mao, Zehan Li, Zekang Li, Zewei Tao, Zewen Ying, Zhaoyang Cong, Zhen Qin, Zhenhua Fan, Zhihang Yu, Zhuo Jiang, Zijia WuJan 14, 20252836

PDF113November 16, 2024