本周精选 12篇大模型领域前沿论文,覆盖大模型安全与可信性、大模型评估与基准测试、大模型行业应用与落地、大模型技术架构与系统优化等方向。全部200多篇论文皆可扫码免费领取。

➔➔➔➔点击查看原文,获取本期大模型周报合集https://mp.weixin.qq.com/s/AmghpfYjhpsDwlmfbuLdNQ

llm词云图

llm词云图

一、大模型安全与可信性方向

1、PIShield: Detecting Prompt Injection Attacks via Intrinsic LLM Features

作者:Wei Zou, Yupei Liu, Yanting Wang, Ying Chen, Neil Gong, Jinyuan Jia

亮点:提出一种高效且有效的提示注入攻击检测方法,核心在于利用LLM特定“注入关键层”提取的提示最终token内部表示,捕捉清洁与受污染提示的区分特征;基于该特征训练简单线性分类器,在5个基准数据集和8种攻击场景下,显著优于11种基线方法,且能抵抗强自适应攻击。

Overview of LLM-integrated applications

Overview of LLM-integrated applications

论文:(arxiv论文网址:https://arxiv.org/abs/2510.14005)

开源代码:https://github.com/weizou52/PIShield

Comments:The code is available at https://github.com/weizou52/PIShield

2、Confidence-Based Response Abstinence: Improving LLM Trustworthiness via Activation-Based Uncertainty Estimation

作者:Zhiqi Huang, Vivek Datla, Chenyang Zhu, Alfy Samuel, Daben Liu, Anoop Kumar, Ritesh Soni

亮点:针对高风险领域LLM可信度问题,提出基于FFN原始激活的不确定性估计方法,避免token logits和概率投影后的信息损失;将置信度预测建模为序列分类任务,结合Huber损失正则化提升鲁棒性,在金融客服场景中实现高准确性与低延迟的平衡。

Diagram of the proposed Retrieval Augmented Generation (RAG) with the confidence model.

Diagram of the proposed Retrieval Augmented Generation (RAG) with the confidence model.

论文:(arxiv论文网址:https://arxiv.org/abs/2510.13750)

开源代码:无

Comments:UncertaiNLP at EMNLP 2025

3、ConsistencyAI: A Benchmark to Assess LLMs' Factual Consistency When Responding to Different Demographic Groups

作者:Peter Banyas, Shristi Sharma, Alistair Simmons, Atharva Vispute

亮点:提出独立基准ConsistencyAI,评估LLM对不同人群的事实一致性;通过19个LLM在15个主题、100种角色下的测试,采用句子嵌入余弦相似度计算得分,发现模型提供商和主题均影响一致性,Grok-3表现最优,就业市场主题一致性最低。

The basic pipeline of our benchmark

The basic pipeline of our benchmark

论文:(arxiv论文网址:https://arxiv.org/abs/2510.13852)

开源代码:http://github.com/banyasp/consistencyAI

Comments:For associated code repository, see http://github.com/banyasp/consistencyAI For user-friendly web app, see http://v0-llm-comparison-webapp.vercel.app/

二、大模型评估与基准测试方向

1、Think Globally, Group Locally: Evaluating LLMs Using Multi-Lingual Word Grouping Games

作者:César Guerra-Solano, Zhuochun Li, Xiang Lorraine Li

亮点:提出抽象推理评估任务GlobalGroup,灵感源自《纽约时报》Connections游戏,覆盖5种语言及对应英语翻译版本;设计游戏难度度量指标,实现受控对比,发现英语模态下模型抽象推理表现更优,开源与闭源模型存在性能差距。

An example Spanish (ES) GLOBALGROUP game

An example Spanish (ES) GLOBALGROUP game

论文:(arxiv论文网址:https://arxiv.org/abs/2510.14030)

Comments:EMNLP Main 2025

2、Attribution Quality in AI-Generated Content: Benchmarking Style Embeddings and LLM Judges

作者:Misam Abbas

亮点:在Human AI Parallel Corpus数据集上,基准测试风格嵌入和GPT-4o法官两种AI生成内容归因机制,覆盖6个领域;发现风格嵌入在GPT生成内容归因上准确率更高(82% vs 68%),LLM法官在小说和学术文本上更优,二者形成互补。

Task Formulation

Task Formulation

论文:(arxiv论文网址:https://arxiv.org/abs/2510.13898)

开源代码:无(提供GitHub代码和Hugging Face衍生数据)

Comments:Accepted for publication at the 2025 IEEE ICDM Workshop on "Grounding Documents with Reasoning, Agents, Retrieval, and Attribution". This is author submitted version. Not yet published

3、TextBandit: Evaluating Probabilistic Reasoning in LLMs Through Language-Only Decision Tasks

作者:Jimin Lim, Arjun Damerla, Arthur Jiang, Nam Le

亮点:提出纯语言反馈的多臂老虎机基准任务,LLM需仅通过语言提示推断潜在奖励结构并做序列决策;评估4个开源LLM,发现Qwen3-4B最佳臂选择率达89.2%,显著优于更大模型和传统决策算法,证实语言可催生出概率推理能力。

Comparison of cumulative regret trends for four LLMs.

Comparison of cumulative regret trends for four LLMs.

论文:(arxiv论文网址:https://arxiv.org/abs/2510.13878)

Comments:COLM 2025 @ ORIGen Workshop

三、大模型行业应用与落地方向

1、FinAI Data Assistant: LLM-based Financial Database Query Processing with the OpenAI Function Calling API

作者:Juhyeong Kim, Yejin Kim, Youngbin Lee, Hyunwoo Byun

亮点:提出金融数据库自然语言查询方案,结合LLM与OpenAI函数调用API,将用户请求路由到参数化查询库而非生成完整SQL;在价格和基本面数据实验中,相比文本到SQL基线,实现更低延迟、成本和更高可靠性。

Cost, time, and accuracy for database query processing

Cost, time, and accuracy for database query processing

论文:(arxiv论文网址:https://arxiv.org/abs/2510.14162)

Comments:4 pages, 2 figures, accepted at CIKM 2025 FinAI Workshop

2、A Methodology for Assessing the Risk of Metric Failure in LLMs Within the Financial Domain

作者:William Flanagan, Mukunda Das, Rajitha Ramanayake, Swanuja Maslekar, Meghana Mangipudi, Joong Ho Choi, Shruti Nair, Shambhavi Bhusan, Sanjana Dulam, Mouni Pendharkar, Nidhi Singh, Vashisth Doshi, Sachi Shah Paresh

亮点:针对金融领域LLM性能评估痛点,指出传统指标难以泛化到生成式AI任务,现有基准缺乏工业适用性;提出风险评估框架,整合领域专家评估与机器学习指标,助力金融机构合理选择评估方案。

论文:(arxiv论文网址:https://arxiv.org/abs/2510.13524)

Comments:NeurIPS 2025 GenAI in Finance Workshop

3、Beyond Static LLM Policies: Imitation-Enhanced Reinforcement Learning for Recommendation

作者:Yi Zhang, Lili Xie, Ruihong Qiu, Jiajun Liu, Sen Wang

亮点:提出模仿增强的离线强化学习推荐框架,通过逆强化学习从LLM生成轨迹中提取奖励模型,无需微调LLM即可转移其语义洞察;在两个基准数据集上,性能优于现有RL和上下文学习基线,且降低计算开销。

The left part illustrates the collection of demonstrations from the LLM policy

The left part illustrates the collection of demonstrations from the LLM policy

论文:(arxiv论文网址:https://arxiv.org/abs/2510.13229)

开源代码:https://github.com/ArronDZhang/IL-Rec

Comments:ICDM 2025 Accepted Paper

四、大模型技术架构与系统优化方向

1、F-BFQ: Flexible Block Floating-Point Quantization Accelerator for LLMs

作者:Jude Haris, José Cano

亮点:提出灵活块浮点量化(F-BFQ)加速器,支持动态切换两种BFP量化变体并执行矩阵乘法,适配LLM层间混合量化需求;在AMD Kria板上部署,相比Arm NEON CPU执行,平均推理时间减少1.4倍,达5.2 tokens/秒。

Overview of our F-BFQ accelerator design for Q2_K and Q3_K MatMul operations

Overview of our F-BFQ accelerator design for Q2_K and Q3_K MatMul operations

论文:(arxiv论文网址:https://arxiv.org/abs/2510.13401)

开源代码:无

Comments:Accepted to Workshop on New Approaches for Addressing the Computing Requirements of LLMs and GNNs (LG-ARC) @ ISCA 2025

2、D-SMART: Enhancing LLM Dialogue Consistency via Dynamic Structured Memory And Reasoning Tree

作者:Xiang Lei, Qin Li, Min Zhang, Min Zhang

亮点:提出模型无关框架D-SMART,通过动态结构化内存(DSM)构建对话知识图谱,结合推理树(RT)实现多步推理;在MT-Bench-101基准上,使开源和闭源模型的对话一致性得分提升超48%,开源模型质量得分最高提升10.1%。

An overview of the D-SMART framework

An overview of the D-SMART framework

论文:(arxiv论文网址:https://arxiv.org/abs/2510.13363)

开源代码:无

Comments:8 pages, 6 figures (main content); 25 pages, 18 figures (total). MSC Class: 68T50; 68T30 ACM Class: I.2.7; I.2.4

3、LLM Agent Communication Protocol (LACP) Requires Urgent Standardization: A Telecom-Inspired Protocol is Necessary

作者:Xin Li, Mengbing Liu, Chau Yuen

亮点:提出LLM智能体通信协议(LACP)标准化需求,借鉴电信领域分层协议设计,构建三层架构以保障通信语义清晰、事务完整和内置安全;认为统一协议是解决多智能体生态碎片化、实现6G复杂实时应用的关键。

LLM Agent Communication: From Ad-Hoc Chaos to LACP Clarity.

LLM Agent Communication: From Ad-Hoc Chaos to LACP Clarity.

论文:(arxiv论文网址:https://arxiv.org/abs/2510.13821)

开源代码:https://github.com/LiXin97/LACP

Comments:Accepted at NeurIPS 2025 AI4NextG Workshop

➔➔➔➔点击查看原文,获取本期大模型周报合集https://mp.weixin.qq.com/s/AmghpfYjhpsDwlmfbuLdNQ

更多推荐