I grandi modelli linguistici possono auto-migliorarsi nel ragionamento a lungo contesto.

Large Language Models Can Self-Improve in Long-context Reasoning

November 12, 2024

Autori: Siheng Li, Cheng Yang, Zesen Cheng, Lemao Liu, Mo Yu, Yujiu Yang, Wai Lam

cs.AI

Abstract

I grandi modelli linguistici (LLM) hanno ottenuto progressi sostanziali nel trattamento di contesti lunghi, ma faticano ancora con il ragionamento a lungo contesto. Gli approcci esistenti coinvolgono tipicamente il raffinamento dei LLM con dati sintetici, che dipendono da annotazioni di esperti umani o modelli avanzati come il GPT-4, limitando così ulteriori progressi. Per affrontare questo problema, esaminiamo il potenziale dei LLM di auto-migliorarsi nel ragionamento a lungo contesto e proponiamo \ours, un approccio appositamente progettato per questo scopo. Questo approccio è diretto: campioniamo più output per ogni domanda, li valutiamo con il Rischio Minimo di Bayes, e quindi applichiamo un raffinamento supervisionato o un'ottimizzazione delle preferenze basata su questi output. Estesi esperimenti su diversi principali LLM dimostrano l'efficacia di \ours, con un miglioramento assoluto di 4,2 punti per Llama-3.1-8B-Instruct. Inoltre, \ours raggiunge prestazioni superiori rispetto agli approcci precedenti che dipendono da dati prodotti da esperti umani o modelli avanzati. Prevediamo che questo lavoro aprirà nuove vie per le tecniche di auto-miglioramento in scenari a lungo contesto, essenziali per il continuo avanzamento dei LLM.

English

Large language models (LLMs) have achieved substantial progress in processing long contexts but still struggle with long-context reasoning. Existing approaches typically involve fine-tuning LLMs with synthetic data, which depends on annotations from human experts or advanced models like GPT-4, thus restricting further advancements. To address this issue, we investigate the potential for LLMs to self-improve in long-context reasoning and propose \ours, an approach specifically designed for this purpose. This approach is straightforward: we sample multiple outputs for each question, score them with Minimum Bayes Risk, and then apply supervised fine-tuning or preference optimization based on these outputs. Extensive experiments on several leading LLMs demonstrate the effectiveness of \ours, with an absolute improvement of 4.2 points for Llama-3.1-8B-Instruct. Furthermore, \ours achieves superior performance compared to prior approaches that depend on data produced by human experts or advanced models. We anticipate that this work will open new avenues for self-improvement techniques in long-context scenarios, which are essential for the continual advancement of LLMs.

Summary

AI-Generated Summary

Paper Overview

The paper introduces SEALONG, a self-improvement approach for Large Language Models (LLMs) in long-context reasoning tasks. SEALONG significantly enhances LLM performance without human annotations and outperforms prior approaches, showcasing the potential for LLM self-improvement in challenging question types.

Core Contribution

SEALONG introduces a novel self-improvement approach for LLMs in long-context reasoning tasks.
The method involves multiple output sampling, Minimum Bayes Risk scoring, and supervised fine-tuning or preference optimization.
Demonstrates substantial absolute improvements in LLM performance, particularly with the Llama-3.1-8B-Instruct model.

Research Context

Addresses the limitations of existing approaches in covering challenging question types requiring full-context reasoning.
Proposes a self-improvement strategy for LLMs in long-context reasoning, advancing the capabilities of LLMs in processing extended contexts.

Keywords

Large Language Models (LLMs)
Self-improvement
Long-context reasoning
Minimum Bayes Risk (MBR)
Supervised fine-tuning

Background

The paper focuses on enhancing Large Language Models' (LLMs) performance in long-context reasoning tasks. Existing approaches relying on synthetic data for fine-tuning LLMs have limitations, prompting the development of SEALONG to enable self-improvement in LLMs without human annotations.

Research Gap

Existing approaches struggle with long-context reasoning despite advancements in processing extended contexts.
Synthetic data-based fine-tuning methods hinder further progress in improving LLM performance.

Technical Challenges

LLMs face difficulties in handling challenging question types requiring full-context reasoning.
Experimental setups for implementing SEALONG are limited to LLMs with up to 14B parameters.

Prior Approaches

Previous methods involved fine-tuning LLMs with synthetic data, restricting advancements in enhancing LLM performance.
SEALONG surpasses prior approaches by enabling self-improvement in LLMs for long-context reasoning tasks.

Methodology

The methodology of SEALONG involves multiple technical components to facilitate self-improvement in LLMs for long-context reasoning tasks.

Theoretical Foundation

Utilizes multiple output sampling, Minimum Bayes Risk scoring, and supervised fine-tuning or preference optimization for LLM self-improvement.
Implements self-supervision strategies to guide LLMs in generating accurate responses and evaluating their performance.

Technical Architecture

Incorporates Sequence parallelization and QLoRA for efficient fine-tuning in long-context scenarios.
Specifies key parameters like LoRA rank, alpha, dropout, batch size, learning rate, and maximum sequence length for model fine-tuning.

Implementation Details

Fine-tunes models for one epoch on a computing setup with 8×H100GPUs.
Utilizes jina-embeddings-v3 and ORPO for fine-tuning Llama-3.1 and Qwen-2.5 models in SEALONG.

Innovation Points

SEALONG introduces innovative self-improvement strategies for LLMs in long-context reasoning tasks.
Implements advanced techniques like MBR decoding and supervised fine-tuning to enhance LLM performance.

Experimental Validation

The experimental validation of SEALONG showcases its effectiveness in improving LLM performance in long-context reasoning tasks.

Setup

Utilizes MuSiQue's training set with self-supervision for generalization ability.
Synthesizes training data by combining related and unrelated documents for context in the experiments.

Metrics

Evaluates performance using SubEM metric, emphasizing top-performing results.
Compares SEALONG's performance across various long-context tasks like Qasper, MultiFieldQA-En, HotpotQA, MuSiQue, and 2WikiMultihopQA.

Results

Demonstrates significant enhancements in LLM performance across different models with SEALONG.
Outperforms previous datasets and shows strong data efficiency with minimal synthetic examples.

Comparative Analysis

Compares SEALONG with prior approaches, highlighting its superior performance in long-context reasoning tasks.
Shows improvements in long-context performance without compromising short-context task performance.

Impact and Implications

SEALONG's contributions have significant implications for the advancement of LLMs in long-context reasoning tasks.

Key Findings

SEALONG achieves notable absolute improvements in LLM performance, particularly with the Llama-3.1-8B-Instruct model.
Demonstrates the potential for LLM self-improvement without human annotations, paving the way for enhanced long-context reasoning capabilities.

Limitations

The experimental setup for SEALONG is currently limited to LLMs with up to 14B parameters.
Further investigation is needed to assess SEALONG's effectiveness at larger scales and explore longer context lengths.

Future Directions

Future research should focus on creating high-quality prompt sets for long-context LLMs to enhance performance.
Exploration of longer context lengths and scalability of SEALONG at larger parameter scales is essential for comprehensive evaluation.

Practical Significance

SEALONG's self-improvement strategies offer practical applications in enhancing LLM performance for challenging question types requiring full-context reasoning.
The methodology and findings of SEALONG can be leveraged to advance the capabilities of LLMs in processing extended contexts effectively.

Articoli in Evidenza

DeepSeek-R1: Incentivizzare la capacità di ragionamento nei LLM tramite Apprendimento per Rinforzo
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 Zhang•Jan 22, 2025•3735

Rapporto Tecnico 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 Qiu•Dec 19, 2024•36311

MiniMax-01: Scalare i modelli di base con attenzione lampeggiante
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 Wu•Jan 14, 2025•2836