语言大模型 / LLM - arXivDaily 专题

2402.00367 2024-07-02 cs.CL 87%

Don't Hallucinate, Abstain: Identifying LLM Knowledge Gaps via Multi-LLM Collaboration

Shangbin Feng, Weijia Shi, Yike Wang, Wenxuan Ding, Vidhisha Balachandran, Yulia Tsvetkov

专题命中指令微调：LLM(title,abstract)；large language model(abstract)；language model(abstract)；prompting(abstract)

Comments ACL 2024

2406.17923 2024-06-27 cs.CL 87%

PAFT: A Parallel Training Paradigm for Effective LLM Fine-Tuning

Shiva Kumar Pentyala, Zhichao Wang, Bin Bi, Kiran Ramnath, Xiang-Bo Mao, Regunathan Radhakrishnan, Sitaram Asur, Na, Cheng

专题命中指令微调：LLM(title,abstract)；large language model(abstract)；language model(abstract)；SFT(abstract)

2405.04403 2024-05-08 cs.CV cs.CL 87%

Learning To See But Forgetting To Follow: Visual Instruction Tuning Makes LLMs More Prone To Jailbreak Attacks

Georgios Pantazopoulos, Amit Parekh, Malvina Nikandrou, Alessandro Suglia

专题命中指令微调：instruction tuning(title,abstract)；LLM(abstract)；large language model(abstract)；language model(abstract)

2404.11978 2024-04-19 cs.CL 87%

EVIT: Event-Oriented Instruction Tuning for Event Reasoning

Zhengwei Tao, Xiancai Chen, Zhi Jin, Xiaoying Bai, Haiyan Zhao, Yiwei Lou

专题命中指令微调：instruction tuning(title,abstract)；LLM(abstract)；large language model(abstract)；language model(abstract)

2401.10222 2024-04-12 cs.CV cs.AI 87%

Supervised Fine-tuning in turn Improves Visual Foundation Models

Xiaohu Jiang, Yixiao Ge, Yuying Ge, Dachuan Shi, Chun Yuan, Ying Shan

专题命中指令微调：foundation model(title,abstract)；instruction tuning(abstract)；pretraining(abstract)；SFT(abstract)

Comments 23 pages, 3 figures, Project page: https://github.com/TencentARC/ViSFT/tree/main

2403.12776 2024-03-20 cs.CL 87%

Automated Data Curation for Robust Language Model Fine-Tuning

Jiuhai Chen, Jonas Mueller

专题命中指令微调：language model(title,abstract)；LLM(abstract)；large language model(abstract)；instruction tuning(abstract)

2402.11251 2024-02-20 cs.CL 87%

LLM can Achieve Self-Regulation via Hyperparameter Aware Generation

Siyin Wang, Shimin Li, Tianxiang Sun, Jinlan Fu, Qinyuan Cheng, Jiasheng Ye, Junjie Ye, Xipeng Qiu, Xuanjing Huang

专题命中指令微调：LLM(title,abstract)；large language model(abstract)；language model(abstract)；instruction tuning(abstract)

2306.17107 2024-02-06 cs.CV cs.CL 87%

LLaVAR: Enhanced Visual Instruction Tuning for Text-Rich Image Understanding

Yanzhe Zhang, Ruiyi Zhang, Jiuxiang Gu, Yufan Zhou, Nedim Lipka, Diyi Yang, Tong Sun

专题命中指令微调：instruction tuning(title,abstract)；LLM(abstract)；large language model(abstract)；language model(abstract)

Comments Preprint

2212.00616 2023-12-04 cs.CL 87%

Extensible Prompts for Language Models on Zero-shot Language Style Customization

Tao Ge, Jing Hu, Li Dong, Shaoguang Mao, Yan Xia, Xun Wang, Si-Qing Chen, Furu Wei

专题命中指令微调：language model(title,abstract)；LLM(abstract)；large language model(abstract)；prompting(abstract)

Comments Accepted by NeurIPS 2023

2311.15653 2023-11-28 cs.CL 87%

MoDS: Model-oriented Data Selection for Instruction Tuning

Qianlong Du, Chengqing Zong, Jiajun Zhang

专题命中指令微调：instruction tuning(title,abstract)；LLM(abstract)；large language model(abstract)；language model(abstract)

2311.10367 2023-11-20 cs.CL 87%

Exploring the Relationship between In-Context Learning and Instruction Tuning

Hanyu Duan, Yixuan Tang, Yi Yang, Ahmed Abbasi, Kar Yan Tam

专题命中指令微调：instruction tuning(title,abstract)；LLM(abstract)；large language model(abstract)；language model(abstract)

2305.13225 2023-10-10 cs.CL 87%

Multi-Task Instruction Tuning of LLaMa for Specific Scenarios: A Preliminary Study on Writing Assistance

Yue Zhang, Leyang Cui, Deng Cai, Xinting Huang, Tao Fang, Wei Bi

专题命中指令微调：instruction tuning(title,abstract)；LLM(abstract)；large language model(abstract)；language model(abstract)

2305.14898 2023-05-25 cs.CL 87%

PIVOINE: Instruction Tuning for Open-world Information Extraction

Keming Lu, Xiaoman Pan, Kaiqiang Song, Hongming Zhang, Dong Yu, Jianshu Chen

专题命中指令微调：instruction tuning(title,abstract)；LLM(abstract)；large language model(abstract)；language model(abstract)

2607.27566 2026-07-31 cs.CV 新提交 87%

Objective-Aligned Direct Answer SFT for Robust Multi-Frame Medical VQA

面向稳健多帧医学视觉问答的目标对齐直接答案监督微调

Site Li, Jianyi Hao, Xiaofeng Liu

机构 * Yale University（耶鲁大学）

专题命中指令微调：SFT(title,summary_cn)；foundation model(comments)

AI总结该研究针对MedFrameQA数据集，提出目标对齐的直接答案SFT是最强稳健适配系列，其在保留报告准确率上优于冻结基线且稳定性好，还可迁移至其他骨干网络，旨在引导研究聚焦稳健优化而非架构复杂度。

Comments Presented at the CVPR 2026 Workshop on Multimodal Foundation Models for Biomedicine: Challenges and Opportunities

2512.19399 2025-12-24 cs.LG cs.AI cs.CL 87%

Brain-Grounded Axes for Reading and Steering LLM States

基于大脑活动的阅读与操控LLM状态的轴

Sandro Andric

专题命中指令微调：LLM(title,abstract)；large language model(abstract)；language model(abstract)；分类 cs.CL、cs.AI、cs.LG

AI总结本文提出利用人类大脑活动作为坐标系，通过ICA提取潜在轴，操控LLM状态，实现可解释和可控的LLM行为。

Comments 10 pages, 4 figures. Code: https://github.com/sandroandric/Brain-Grounded-Axes-for-Reading-and-Steering-LLM-States

2507.09499 2025-07-15 eess.AS cs.SD 87%

The DKU System for Multi-Speaker Automatic Speech Recognition in MLC-SLM Challenge

Yuke Lin, Ming Cheng, Ze Li, Ming Li

机构 * School of Computer Science, Wuhan University, China（武汉大学计算机学院）； Suzhou Municipal Key Laboratory of Multimodal Intelligent Systems, Digital Innovation Research Center, Duke Kunshan University, China（多克大学数字创新研究中心）

专题命中指令微调：SLM(title,abstract)；LLM(abstract)；large language model(abstract)；language model(abstract)

Comments Technical Report for MLC-SLM Challenge in Interspeech2025

2311.13133 2023-12-29 cs.LG cs.AI cs.CL 87%

LIMIT: Less Is More for Instruction Tuning Across Evaluation Paradigms

Aditi Jha, Sam Havens, Jeremy Dohmann, Alex Trott, Jacob Portes

专题命中指令微调：instruction tuning(title,comments)；LLM(abstract)；large language model(abstract)；language model(abstract)

Comments 36 pages, 12 figures, NeurIPS 2023 Workshop on Instruction Tuning and Instruction Following

2309.00615 2023-09-04 cs.CV cs.AI cs.CL cs.LG cs.MM 87%

Point-Bind & Point-LLM: Aligning Point Cloud with Multi-modality for 3D Understanding, Generation, and Instruction Following

Ziyu Guo, Renrui Zhang, Xiangyang Zhu, Yiwen Tang, Xianzheng Ma, Jiaming Han, Kexin Chen, Peng Gao, Xianzhi Li, Hongsheng Li, Pheng-Ann Heng

专题命中指令微调：LLM(title,abstract)；large language model(abstract)；language model(abstract)；分类 cs.CL、cs.AI、cs.LG

Comments Work in progress. Code is available at https://github.com/ZiyuGuo99/Point-Bind_Point-LLM

2608.25428 2026-08-27 cs.CL cs.AI 新提交 87%

DCGC: Draft-Conditioned Global Correction for Complex Reasoning with Masked Diffusion Models

DCGC：基于草稿条件的全局修正，用于带掩码扩散模型的复杂推理

Minhae Oh, Nakyung Lee, Jungwoo Lee

机构 * Seoul National University（首尔大学）

专题命中指令微调：SFT(summary_cn,abstract)；large language model(abstract)；language model(abstract)；分类 cs.CL、cs.AI

AI总结 DCGC是结合SFT与动态双CFG机制的MDM框架，用于修正LLMs推理错误，在多推理基准中优于现有方法，可作为无验证器的全局修正模块。

Comments 21 pages, 3 figures, 12 tables

2608.25243 2026-08-27 cs.CL cs.LG 新提交 87%

From Memorization to Absorption: Mixed-Policy RL for Continual Knowledge Injection

从记忆到吸收：用于持续知识注入的混合策略强化学习

Zhibo Hou, Fan Zhao, Zhiyu An, Wan Du

机构 * University of California, Merced（加州大学默塞德分校）

专题命中指令微调：SFT(summary_cn,abstract)；large language model(abstract)；language model(abstract)；分类 cs.CL、cs.LG

AI总结针对大语言模型持续知识注入中监督微调泛化不足的问题，提出三阶段自学习框架GRIN，其核心为混合策略RL算法Golden-GRPO，引入两个基准验证，结果表明GRIN性能优于SFT及混合策略RL基线。

2605.12264 2026-08-27 cs.CR cs.CL cs.LG 版本更新 87%

Reconstruction of Personally Identifiable Information from Proprietary Data in Supervised Fine-Tuned Models

从监督微调模型中重建个人身份信息

Sae Furukawa, Alina Oprea

机构 * Northeastern University（东北大学）

专题命中指令微调：LLM(abstract,abstract_cn)；SFT(abstract,abstract_cn)；large language model(abstract)；language model(abstract)

AI总结本文研究了从监督微调模型中重建个人身份信息的问题，提出COVA算法在前缀攻击下优于现有提取方法，揭示了不同PII类型泄露程度差异。

2608.22963 2026-08-25 cs.AI cs.CL 新提交 87%

Buried in Textual Debt: Context Pruning with Visual Evidence Preservation for MLLM Agents

被文本债务掩埋：面向多模态大语言模型智能体的保留视觉证据的上下文剪枝

Yuchen Huang, Sijia Li, Jun Zhang, Yi R. Fung

机构 * Hong Kong University of Science and Technology（香港科技大学）

专题命中指令微调：SFT(summary_cn,abstract)；large language model(abstract)；language model(abstract)；分类 cs.CL、cs.AI

AI总结针对多模态大语言模型智能体长轨迹中文本主导上下文、抑制视觉证据的问题，提出基于KL引导的SPARE框架，结合OPSD与SFT实现高效剪枝，在多步视觉工具使用基准中达到最优准确率与剪枝比例的权衡。

Comments 14 pages, 2 figures, 4 tables

2608.18767 2026-08-20 cs.CL cs.LG 新提交 87%

Gradient Mirage: Trainable yet Label-Unidentifiable Gradients in Large Language Model Split Learning

梯度幻影：大语言模型拆分学习中可训练且标签不可识别的梯度

Shiyu Miao, Yunlong Mao, Zirui Huang, Liang Yao, Tianshuo Zheng, Yanhui Gu, Fan Liu, Sheng Zhong

专题命中指令微调：large language model(title)；language model(title)；LLM(abstract)；分类 cs.CL、cs.LG

AI总结针对大语言模型拆分学习中梯度匹配攻击的标签泄露问题，提出Gradient Mirage防御方法，通过三个维度的不一致性打破梯度-目标一致性，在保留优化效用的同时提升隐私保护，实现更优的隐私-效用权衡。

2608.14563 2026-08-18 cs.LG cs.AI 新提交 87%

Forward Pass Domain Adaptation (Without Cross-Layer Backpropagation)

仅前向传播的领域适配（无需跨层反向传播）

Rivaan Patil, Simon Dennis, Hao Guo, Kevin Shabahang

机构 * University of California, Santa Cruz（加州大学圣克鲁兹分校）； University of Melbourne（墨尔本大学）

专题命中指令微调：SFT(summary_cn,abstract)；large language model(abstract)；language model(abstract)；分类 cs.AI、cs.LG

AI总结该研究提出仅前向传播的MLP训练（FPO）方法，无需跨层反向传播即可适配大语言模型，提升吞吐量、降低内存开销，在保持域外基准性能的同时优化域内困惑度，且耗时低于限定SFT的方案。

Comments 15 pages

2608.10414 2026-08-12 cs.CL cs.LG 新提交 87%

How Robust Are LLMs to Vietnamese Dialects?

大型语言模型对越南方言的鲁棒性如何？

Minh Tran, Trinh Chau, Thanh-Nhan Le, Nam Tran, Luan Thanh Nguyen, Cuong Dang, Duc Hoang

专题命中指令微调：LLM(summary_cn,abstract)；large language model(abstract)；language model(abstract)；分类 cs.CL、cs.LG

AI总结本研究通过构建VialectBench基准，评估指令微调LLM对越南方言的鲁棒性，发现方言输入会导致模型性能下降，且不同方言组的影响存在显著差异，标准越南语上的性能不代表方言场景下的可靠性。

Comments 8 pages, 3 figures, 4 tables

2607.19364 2026-08-12 cs.AI cs.CL 版本更新 87%

SAE-StatSteer: Statistical Consensus Feature Selection for Optimization-Free Activation Steering of Large Language Models

用于大语言模型激活空间控制的基于统计的稀疏特征干预

Oshayer Siddique, J. M Areeb Uzair Alam, Md Jobayer Rahman Rafy, Syed Rifat Raiyan, Hasan Mahmud, Md Kamrul Hasan

机构 * Islamic University of Technology（伊斯兰科技大学）； Systems and Software Lab (SSL)（系统与软件实验室）

专题命中指令微调：large language model(title)；language model(title)；LLM(abstract)；分类 cs.CL、cs.AI

AI总结研究大语言模型激活空间控制问题时，提出基于统计的稀疏特征干预方法，先经可靠性过滤，再用三个统计量排序特征，构建引导方向。在多模型、领域和配置上实验，发现该方法有特定领域转移效果，且引导受多种因素影响，强调评估应兼顾质量与行为转移。

Comments Under review, 26 pages, 5 figures, 16 tables

2607.06160 2026-08-04 cs.CL cs.AI 版本更新 87%

LongCrafter: Towards Diverse Long-Context Understanding via Evidence-Graph-Guided Instruction Synthesis

LongCrafter：通过证据图引导的指令合成实现多样化的长上下文理解

Chenhao Yuan, Yinhao Xu, Shuwen Xu, Xizhi Yang, Jiaxiang Liu, Chenxi Zhou, Shaoping Huang, Haolin Ren, Pengfei Cao, Jun Zhao, Kang Liu

机构 * University of Chinese Academy of Sciences（中国科学院大学）； The Key Laboratory of Cognition and Decision Intelligence for Complex Systems, Institute of Automation, Chinese Academy of Sciences（中国科学院自动化研究所认知与决策智能复杂系统重点实验室）

专题命中指令微调：SFT(summary_cn,abstract)；large language model(abstract)；language model(abstract)；分类 cs.CL、cs.AI

AI总结针对现有长上下文理解方法的局限，提出LongCrafter框架，结合分层任务分类法与证据管道，生成多样化长上下文SFT数据，微调后的模型在多个数据集上表现优异，能有效缓解“中间迷失”问题。

2604.13627 2026-08-04 cs.LG cs.CL 版本更新 87%

(How) Learning Rates Regulate Catastrophic Overtraining

(如何) 学习率调节灾难性过训练

Mark Rofin, Aditya Varre, Nicolas Flammarion

机构 * EPFL（瑞士联邦理工学院）

专题命中指令微调：LLM(abstract,abstract_cn)；SFT(abstract,abstract_cn)；pretraining(abstract)；post-training(abstract)

AI总结研究通过隐式正则化分析学习率对微调中灾难性遗忘的影响，发现学习率大小影响模型收敛方向，并揭示学习率衰减加剧预训练模型尖锐性从而导致灾难性过训练。

Comments COLM 2026

2607.29238 2026-08-03 cs.CL cs.AI cs.CY cs.ET cs.HC 新提交 87%

Small Is Enough: Per-User Style Rewriting of AI-Edited Text via LoRA Adapters

小模型已足够：通过LoRA适配器对AI编辑文本进行单用户风格重写

Antorweep Chakravorty

机构 * University of Stavanger（斯塔万格大学）

专题命中指令微调：LLM(summary_cn,abstract)；language model(abstract)；small language model(abstract)；分类 cs.CL、cs.AI

AI总结 InMyStyle是注重隐私的单用户系统，用多本地辅助LLM构建配对训练示例，微调0.5B-7B参数模型的LoRA适配器，可将AI编辑文本重写为用户风格，小模型即可完成重写任务，其输出感知AI性低于辅助AI生成内容

2510.18874 2026-06-29 cs.LG cs.CL 版本更新 87%

Retaining by Doing: The Role of On-Policy Data in Mitigating Forgetting

在做中保留：在线策略数据在缓解遗忘中的作用

Howard Chen, Noam Razin, Karthik Narasimhan, Danqi Chen

专题命中指令微调：SFT(summary_cn,abstract)；language model(abstract)；post-training(abstract)；分类 cs.CL、cs.LG

AI总结本文系统比较了监督微调（SFT）和强化学习（RL）在语言模型后训练中的遗忘模式，发现RL因使用在线策略数据而更少遗忘，并验证了在线策略数据是缓解遗忘的关键因素。

Journal ref Proceedings of the 43rd International Conference on Machine Learning (ICML), 2026

1. 指令微调 11692 篇

Don't Hallucinate, Abstain: Identifying LLM Knowledge Gaps via Multi-LLM Collaboration

PAFT: A Parallel Training Paradigm for Effective LLM Fine-Tuning

Learning To See But Forgetting To Follow: Visual Instruction Tuning Makes LLMs More Prone To Jailbreak Attacks

EVIT: Event-Oriented Instruction Tuning for Event Reasoning

Supervised Fine-tuning in turn Improves Visual Foundation Models

Automated Data Curation for Robust Language Model Fine-Tuning

LLM can Achieve Self-Regulation via Hyperparameter Aware Generation

LLaVAR: Enhanced Visual Instruction Tuning for Text-Rich Image Understanding

Extensible Prompts for Language Models on Zero-shot Language Style Customization

MoDS: Model-oriented Data Selection for Instruction Tuning

Exploring the Relationship between In-Context Learning and Instruction Tuning

Multi-Task Instruction Tuning of LLaMa for Specific Scenarios: A Preliminary Study on Writing Assistance

PIVOINE: Instruction Tuning for Open-world Information Extraction

Objective-Aligned Direct Answer SFT for Robust Multi-Frame Medical VQA

Brain-Grounded Axes for Reading and Steering LLM States

The DKU System for Multi-Speaker Automatic Speech Recognition in MLC-SLM Challenge

LIMIT: Less Is More for Instruction Tuning Across Evaluation Paradigms

Point-Bind & Point-LLM: Aligning Point Cloud with Multi-modality for 3D Understanding, Generation, and Instruction Following

DCGC: Draft-Conditioned Global Correction for Complex Reasoning with Masked Diffusion Models

From Memorization to Absorption: Mixed-Policy RL for Continual Knowledge Injection

Reconstruction of Personally Identifiable Information from Proprietary Data in Supervised Fine-Tuned Models

Buried in Textual Debt: Context Pruning with Visual Evidence Preservation for MLLM Agents

Gradient Mirage: Trainable yet Label-Unidentifiable Gradients in Large Language Model Split Learning

Forward Pass Domain Adaptation (Without Cross-Layer Backpropagation)

How Robust Are LLMs to Vietnamese Dialects?

SAE-StatSteer: Statistical Consensus Feature Selection for Optimization-Free Activation Steering of Large Language Models

LongCrafter: Towards Diverse Long-Context Understanding via Evidence-Graph-Guided Instruction Synthesis

(How) Learning Rates Regulate Catastrophic Overtraining

Small Is Enough: Per-User Style Rewriting of AI-Edited Text via LoRA Adapters

Retaining by Doing: The Role of On-Policy Data in Mitigating Forgetting