Mastering Long-Context Multi-Task Reasoning with Transformers and Recurrent Memory

IF 1 Q4 OPTICS
A. Bulatov, Y. Kuratov, M. Burtsev
{"title":"Mastering Long-Context Multi-Task Reasoning with Transformers and Recurrent Memory","authors":"A. Bulatov,&nbsp;Y. Kuratov,&nbsp;M. Burtsev","doi":"10.3103/S1060992X24700735","DOIUrl":null,"url":null,"abstract":"<p>Recent advancements have significantly improved the skills and performance of language models, but have also increased computational demands due to the increasing number of parameters and the quadratic complexity of the attention mechanism. As context sizes expand into millions of tokens, making long-context processing more accessible and efficient becomes a critical challenge. Furthermore, modern benchmarks such as BABILong [1] underscore the inefficiency of even the most powerful LLMs in long context reasoning. In this paper, we employ finetuning and multi-task learning to train a model capable of mastering multiple BABILong long-context reasoning skills. We demonstrate that even models with fewer than 140 million parameters can outperform much larger counterparts by learning multiple essential tasks simultaneously. By conditioning Recurrent Memory Transformer [2] on task description, we achieve state-of-the-art results on multi-task BABILong QA1–QA5 set for up to 32k tokens. The proposed model also shows generalization abilities to new lengths and tasks, along with increased robustness to input perturbations.</p>","PeriodicalId":721,"journal":{"name":"Optical Memory and Neural Networks","volume":"33 3 supplement","pages":"S466 - S474"},"PeriodicalIF":1.0000,"publicationDate":"2025-01-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Optical Memory and Neural Networks","FirstCategoryId":"1085","ListUrlMain":"https://link.springer.com/article/10.3103/S1060992X24700735","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q4","JCRName":"OPTICS","Score":null,"Total":0}
引用次数: 0

Abstract

Recent advancements have significantly improved the skills and performance of language models, but have also increased computational demands due to the increasing number of parameters and the quadratic complexity of the attention mechanism. As context sizes expand into millions of tokens, making long-context processing more accessible and efficient becomes a critical challenge. Furthermore, modern benchmarks such as BABILong [1] underscore the inefficiency of even the most powerful LLMs in long context reasoning. In this paper, we employ finetuning and multi-task learning to train a model capable of mastering multiple BABILong long-context reasoning skills. We demonstrate that even models with fewer than 140 million parameters can outperform much larger counterparts by learning multiple essential tasks simultaneously. By conditioning Recurrent Memory Transformer [2] on task description, we achieve state-of-the-art results on multi-task BABILong QA1–QA5 set for up to 32k tokens. The proposed model also shows generalization abilities to new lengths and tasks, along with increased robustness to input perturbations.

Abstract Image

掌握长上下文多任务推理与变形和循环记忆
最近的进展显著地提高了语言模型的技能和性能,但由于参数数量的增加和注意力机制的二次复杂性,也增加了计算需求。随着上下文大小扩展到数百万个令牌,使长上下文处理更易于访问和更高效成为一项关键挑战。此外,BABILong[1]等现代基准强调了即使是最强大的llm在长上下文推理中的低效率。在本文中,我们使用微调和多任务学习来训练一个能够掌握多种BABILong长上下文推理技能的模型。我们证明,即使是少于1.4亿个参数的模型,也可以通过同时学习多个基本任务来胜过更大的模型。通过对任务描述进行循环记忆变压器[2]的调节,我们在多任务BABILong QA1-QA5设置上获得了最先进的结果,最多可达32k个令牌。该模型还显示了对新长度和任务的泛化能力,以及对输入扰动的鲁棒性增强。
本文章由计算机程序翻译,如有差异,请以英文原文为准。
求助全文
约1分钟内获得全文 求助全文
来源期刊
CiteScore
1.50
自引率
11.10%
发文量
25
期刊介绍: The journal covers a wide range of issues in information optics such as optical memory, mechanisms for optical data recording and processing, photosensitive materials, optical, optoelectronic and holographic nanostructures, and many other related topics. Papers on memory systems using holographic and biological structures and concepts of brain operation are also included. The journal pays particular attention to research in the field of neural net systems that may lead to a new generation of computional technologies by endowing them with intelligence.
×
引用
GB/T 7714-2015
复制
MLA
复制
APA
复制
导出至
BibTeX EndNote RefMan NoteFirst NoteExpress
×
提示
您的信息不完整,为了账户安全,请先补充。
现在去补充
×
提示
您因"违规操作"
具体请查看互助需知
我知道了
×
提示
确定
请完成安全验证×
copy
已复制链接
快去分享给好友吧!
我知道了
右上角分享
点击右上角分享
0
联系我们:info@booksci.cn Book学术提供免费学术资源搜索服务,方便国内外学者检索中英文文献。致力于提供最便捷和优质的服务体验。 Copyright © 2023 布克学术 All rights reserved.
京ICP备2023020795号-1
ghs 京公网安备 11010802042870号
Book学术文献互助
Book学术文献互助群
群 号:481959085
Book学术官方微信