Look-back Decoding for Open-Ended Text Generation
Nan Xu, Chunting Zhou, Asli Celikyilmaz, Xuezhe Ma
Abstract
Given a prefix (context), open-ended generation aims to decode texts that are coherent, which do not abruptly drift from previous topics, and informative, which do not suffer from undesired repetitions. In this paper, we propose Look-back , an improved decoding algorithm that leverages the Kullback-Leibler divergence to track the distribution distance between current and historical decoding steps. Thus Lookback can automatically predict potential repetitive phrase and topic drift, and remove tokens that may cause the failure modes, restricting the next token probability distribution within a plausible distance to the history. We perform decoding experiments on document continuation and story generation, and demonstrate that Look-back is able to generate more fluent and coherent text, outperforming other strong decoding methods significantly in both automatic and human evaluations 1 . Degeneration LM (Decoding) Continuation Prefix: In addition to live broadcasts FIFA Fan Fests offer food and beverages, merchandise and various entertainment events by local and international artists. The start of 2006 World Cup was Repetitive Continuation S1: (GPT2-XL greedy) marked by the FIFA Fan Fest in Sao Paulo, Brazil.The FIFA Fan Fest is a unique event that brings together fans from all over the world to celebrate the FIFA World Cup . The FIFA Fan Fest is a unique event that brings together fans from... S2: (davinci-002 greedy) celebrated with a concert by Bon Jovi in Berlin.The FIFA Fan Fest in Munich was located at the Olympic Park.The FIFA Fan Fest in Frankfurt was located at the Römerberg.The FIFA Fan Fest in Hamburg was located at the ... Prefix: Burkan died from an attack of acute indigestion at his country home in Great Neck on June 6, 1936. Three thousand people attended his funeral at Off-Topic Continuation S3: (ada-001 greedy) the local church.The authorities are investigating how Khedira managed to enter the house and what role he played in the attack.
Ask about this paper
Your agent reads all of it.
Lune indexed this paper to the last equation, along with the top-tier papers that cite it. Ask a question and the answer quotes them.
Your agent calls
Luneget_paper_fulltext
Free to start. No credit card required.
Terminal
Install the CLIlune papers fulltext 7b75f4ac-19d5-482c-b2bb-b5a2df9c5452Cited by top-tier papers2
- Improving Open-Ended Text Generation via Adaptive DecodingWenhong Zhu, Hongkun Hao, Zhiwei He, Yiming Ai et al.ICML 2024 · 21 citations
- Visual Description Grounding Reduces Hallucinations and Boosts Reasoning in LVLMsSreyan Ghosh, Chandra Kiran Reddy Evuru, Sonal Kumar, Utkarsh Tyagi et al.ICLR 2025
Builds on10
- Language Models are Few-Shot LearnersTom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah et al.NeurIPS 2020 · 64,255 citations
- The Curious Case of Neural Text DegenerationAri Holtzman, Jan Buys, Li Du, Maxwell Forbes et al.ICLR 2020 · 4,112 citations
- SimCSE: Simple Contrastive Learning of Sentence EmbeddingsTianyu Gao, Xingcheng Yao, Danqi ChenEMNLP 2021 · 2,496 citations
- Neural Text Generation With Unlikelihood TrainingSean Welleck, Ilia Kulikov, Stephen Roller, Emily Dinan et al.ICLR 2020 · 683 citations
- MAUVE: Measuring the Gap Between Neural Text and Human Text using Divergence FrontiersKrishna Pillutla, Swabha Swayamdipta, Rowan Zellers, John Thickstun et al.NeurIPS 2021 · 606 citations
Related papers
- Back to the Future: Unsupervised Backprop-based Decoding for Counterfactual and Abductive Commonsense ReasoningLianhui Qin, Vered Shwartz, Peter West, Chandra Bhagavatula et al.EMNLP 2020 · 25 citations
- kNN-LM Does Not Improve Open-ended Text GenerationShufan Wang, Yixiao Song, Andrew Drozdov, Aparna Garimella et al.EMNLP 2023 · 3 citations
- FOCUS & RePAIR: Mitigating Text Degeneration via Token-Level Guidance For Pruned Large Language ModelsJunyoung Lee, Sehyeon Park, Shinhyoung Jang, Seonha Ryu et al.ICML 2026
- Long Text Generation by Modeling Sentence-Level and Discourse-Level CoherenceJian Guan, Xiaoxi Mao, Changjie Fan, Zitao Liu et al.ACL 2021
- Best-k Search Algorithm for Neural Text GenerationJiacheng Xu, Caiming Xiong, Silvio Savarese, Yingbo ZhouACL 2023
