AI-Model Network: Concept, Current State and Future

Huang Jijun; Lai Junjie; Liu Zhongren; Li Zhetao; Long Saiqin; Wang Jianhui; Wu Junru; Xiao Yong; Zeng Xiyu

arxiv: 2606.27382 · v1 · pith:OIWIUXGTnew · submitted 2026-05-25 · 💻 cs.AI

AI-Model Network: Concept, Current State and Future

Li Zhetao , Zeng Xiyu , Wang Jianhui , Xiao Yong , Liu Zhongren , Wu Junru , Lai Junjie , Huang Jijun

show 1 more author

Long Saiqin

This is my paper

Pith reviewed 2026-06-29 22:04 UTC · model grok-4.3

classification 💻 cs.AI

keywords AI-Model NetworkAI-ModelNetmodel interconnectioncollaborative reasoningheterogeneous AI modelslarge language modelsAI paradigm

0 comments

The pith

Pathways between heterogeneous AI models enable interconnection, capability sharing, and collaborative reasoning in AI-ModelNet.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper argues that just as computers gained value through the Internet's sharing and collaboration, AI models need a similar network to overcome high training costs and isolation of large models. It proposes AI-ModelNet to connect these models via pathways for effective interaction among heterogeneous systems. This paradigm shift from single or multi-model approaches to a networked system is validated through a prototype and application cases. The approach aims to facilitate capability sharing among lightweight and domain-specific models by drawing directly from Internet development patterns.

Core claim

By establishing pathways between models, AI-ModelNet achieves interconnection, capability sharing, and collaborative reasoning among heterogeneous AI models. Drawing from the Internet's history, where computation leads to sharing that empowers further computation, the framework addresses the bottleneck of model interaction in the era of large models by proposing a hierarchical architecture for world wide AI-model networking.

What carries the argument

Pathways between models that enable interconnection and collaborative reasoning, analogous to Internet connections between computers.

If this is right

Models can collaborate without needing to be retrained into a single large system.
Lightweight private models can leverage capabilities from others in the network.
Collaborative reasoning emerges from combining domain-specific expertise across models.
The shift reduces reliance on centralized large model training and deployment.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

New protocols for model communication would need to be standardized to realize the network at scale.
Questions of model ownership, data privacy, and contribution incentives remain to be resolved in practice.
Testing the network with real heterogeneous models could reveal compatibility issues not addressed in the prototype.

Load-bearing premise

Heterogeneous models can interact and collaborate effectively through established pathways, with the Internet providing an adequate model for technical and structural compatibility.

What would settle it

An experiment showing that attempts to connect models via pathways fail to produce effective capability sharing or collaborative outputs due to fundamental incompatibilities in model architectures or representations.

read the original abstract

While the primary function of computers lies in computation and processing, the core value of the Internet is rooted in sharing and collaboration. Computers create the Internet, and the Internet empowers the value of computers. The rapid development of the Internet, cloud computing, and big data is pushing artificial intelligence into the era of large models (LMs). However, the practical application of LMs is currently hindered by high training costs and deployment complexities, driving a shift toward lightweight, private, and domain-specific models. With the rapid proliferation and wide distribution of heterogeneous models, enabling effective interaction and collaboration among them has emerged as a critical bottleneck that urgently needs to be addressed in LM development. Drawing inspiration from the development of the Internet, this paper proposes the concept, vision, and system architecture of world wide AI-model network (AI-ModelNet). It is a novel paradigm that achieves interconnection, capability sharing, and collaborative reasoning by establishing pathways between models. We first briefly review the current state of single-model and multi-model research. Subsequently, the systemic vision and hierarchical architecture of AI-ModelNet are articulated, followed by validation of the framework's feasibility through a prototype system and diverse application cases. Finally, key directions for future research are discussed preliminarily.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

This is a high-level vision paper proposing an Internet-style network for interconnecting AI models, but the architecture stays too abstract to assess whether the claimed collaboration would actually work.

read the letter

The paper's core idea is to treat the growing number of domain-specific models like computers on the Internet, linking them through pathways so they can share capabilities and do collaborative reasoning. That framing is the main new element.

It does a reasonable job reviewing the move from large models to lighter, private ones and sketching a hierarchical architecture with layers for connectivity and higher-level functions. The authors also flag some future research directions.

The problems start with the lack of substance behind the vision. The architecture is described at a conceptual level with no details on model interfaces, data formats, invocation protocols, or how outputs from one model get reconciled with another. The prototype is offered as validation, yet nothing is said about what it actually implements, what it measures, or how it performs against existing multi-agent or distributed inference approaches.

This makes the central claim—that establishing pathways produces interconnection and collaborative reasoning—impossible to evaluate. The Internet analogy does not supply the missing technical pieces, since models do not come with standardized packet formats or routing tables. The argument ends up somewhat circular: success is defined by the interconnection the paper assumes can be created.

The work is aimed at readers who follow high-level infrastructure ideas in AI and want to think about deployment at scale. Anyone looking for concrete mechanisms, reproducible results, or falsifiable claims will not find them.

I would not bring this to a reading group and would not cite it. It does not yet deserve serious referee time because the claims cannot be checked without substantially more specification and evidence.

Referee Report

2 major / 1 minor

Summary. The paper proposes the concept of a worldwide AI-ModelNet as a novel paradigm for interconnecting heterogeneous AI models to enable capability sharing and collaborative reasoning, drawing an analogy from the Internet. It briefly reviews single-model and multi-model research, articulates a systemic vision and hierarchical architecture, validates feasibility via an undescribed prototype system and application cases, and discusses future research directions.

Significance. If the architecture could be specified with concrete protocols and shown to produce genuine collaboration, the proposal could address a real bottleneck in distributed large-model ecosystems. The Internet-inspired vision provides an intuitive high-level framing, but the manuscript offers no quantitative validation, error analysis, or derivation of the architecture from requirements, limiting its current contribution to a conceptual outline.

major comments (2)

[prototype system section] Prototype system section: the claim that the prototype validates the framework's feasibility is unsupported because the manuscript provides no description of model interfaces, data exchange formats, invocation protocols, or mechanisms for handling heterogeneous outputs and collaborative reasoning. This leaves the central claim of achieving interconnection and capability sharing unverified.
[systemic vision and hierarchical architecture section] Systemic vision and hierarchical architecture section: the architecture is defined by direct transfer of the Internet development pattern (sharing and collaboration) without addressing AI-specific differences such as the lack of standardized packet formats or routing tables, rendering the pathways at the level of analogy rather than implementable design.

minor comments (1)

[current state review] The review of current single-model and multi-model research is described as brief; expanding it with explicit citations to key limitations that AI-ModelNet is intended to solve would improve grounding.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments. We address each major comment below.

read point-by-point responses

Referee: [prototype system section] Prototype system section: the claim that the prototype validates the framework's feasibility is unsupported because the manuscript provides no description of model interfaces, data exchange formats, invocation protocols, or mechanisms for handling heterogeneous outputs and collaborative reasoning. This leaves the central claim of achieving interconnection and capability sharing unverified.

Authors: We agree the prototype description is insufficient to support the validation claim. The revised manuscript will expand the prototype system section with details on model interfaces, data exchange formats, invocation protocols, and mechanisms for heterogeneous outputs and collaborative reasoning. revision: yes
Referee: [systemic vision and hierarchical architecture section] Systemic vision and hierarchical architecture section: the architecture is defined by direct transfer of the Internet development pattern (sharing and collaboration) without addressing AI-specific differences such as the lack of standardized packet formats or routing tables, rendering the pathways at the level of analogy rather than implementable design.

Authors: The architecture is presented as an Internet-inspired vision for intuitive framing, as stated in the manuscript. We will revise the section to explicitly address AI-specific differences such as heterogeneous outputs and lack of standardized formats, and outline initial directions toward implementable mechanisms while retaining the high-level conceptual contribution. revision: yes

Circularity Check

0 steps flagged

No significant circularity; conceptual proposal with independent vision and prototype validation

full rationale

The paper is a high-level conceptual proposal defining AI-ModelNet as a paradigm for model interconnection inspired by the Internet, followed by architecture description, a prototype for feasibility, and future directions. No mathematical derivations, fitted parameters presented as predictions, or load-bearing self-citations appear in the text. The central claim is presented as a definitional vision rather than a result derived from prior equations or inputs that reduce by construction. The prototype serves as external-to-the-claim validation rather than a self-referential loop. This matches the default expectation for non-circular conceptual papers.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 1 invented entities

The central claim rests on the unproven transfer of Internet-style interconnection to AI models plus the new entity AI-ModelNet itself; no free parameters are fitted because the work is non-quantitative.

axioms (2)

domain assumption Heterogeneous AI models can achieve effective interaction and collaboration once pathways are established.
Invoked in the description of the systemic vision and architecture.
ad hoc to paper The historical development pattern of the Internet (sharing and collaboration) directly transfers to AI model ecosystems.
Stated as the inspirational source for the proposed paradigm.

invented entities (1)

AI-ModelNet no independent evidence
purpose: Global network enabling model interconnection, capability sharing, and collaborative reasoning.
New conceptual entity introduced to address the stated bottleneck; no independent falsifiable evidence supplied beyond the prototype mention.

pith-pipeline@v0.9.1-grok · 5770 in / 1335 out tokens · 37007 ms · 2026-06-29T22:04:16.456832+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

48 extracted references · 9 canonical work pages · 5 internal anchors

[1]

The world-wide web[J]

Berners-Lee T, Cailliau R, Luotonen A, et al. The world-wide web[J]. Communications of the ACM, 1994, 37(8): 76-82

1994
[2]

Attention is all you need [C] // Proc of the 31st Conf on Neural Information Processing Systems

Vaswani A, Shazeer N, Parmar N, et al. Attention is all you need [C] // Proc of the 31st Conf on Neural Information Processing Systems. New York: ACM, 2017: 6000–6010

2017
[3]

Investigating bias in LLM-based bias detection: Disparities between LLMs and human perception [C] //Proc of the 31st Int Conf on Computational Linguistics

Lin Luyuan, Wang Lingzhi, Guo Jinsong, et al. Investigating bias in LLM-based bias detection: Disparities between LLMs and human perception [C] //Proc of the 31st Int Conf on Computational Linguistics. Stroudsburg,PA: ACL, 2025: 10634-10649

2025
[4]

Thinking on networking and computing promoted by domain-specific method [J]

Su Jinshu. Thinking on networking and computing promoted by domain-specific method [J]. Journal of Computer Research and Development, 2025, 62(12): 2889-2894(in Chinese) （苏金树. 领域定制方法促进网络与计算发展的思考[J]. 计算机研究与发展, 2025, 62(12): 2889-2894）

2025
[5]

Don’t hallucinate, abstain: Identifying LLM knowledge gaps via multi-LLM collaboration [C] //Proc of the 62nd Annual Meeting of the Association for Computational Linguistics

Feng Shagbin, Shi Weijia, Wang Yike, et al. Don’t hallucinate, abstain: Identifying LLM knowledge gaps via multi-LLM collaboration [C] //Proc of the 62nd Annual Meeting of the Association for Computational Linguistics. Stroudsburg,PA: ACL, 2024: 14664-14690 15 Journal of Computer Research and Development 2025 年

2024
[6]

Rethinking the reversal curse of LLMs: A prescription from human knowledge reversal [C] //Proc of the 2024 Conf on Empirical Methods in Natural Language Processing

Lu Zhicong, Li Jin, Li Peiguang, et al. Rethinking the reversal curse of LLMs: A prescription from human knowledge reversal [C] //Proc of the 2024 Conf on Empirical Methods in Natural Language Processing. Stroudsburg,PA: ACL, 2024: 7518-7530

2024
[7]

Small LLMs are weak tool learners: A multi-LLM agent [C] //Proc of the 2024 Conf on Empirical Methods in Natural Language Processing

Shen Weizhou, Li Chenliang, Chen Hongzhan, et al. Small LLMs are weak tool learners: A multi-LLM agent [C] //Proc of the 2024 Conf on Empirical Methods in Natural Language Processing. Stroudsburg,PA: ACL, 2024: 16658-16680 (没有届)

2024
[9]

Uncertainty-aware answer selection for improved reasoning in multi-LLM systems [C] //Proc of the Findings of the Association for Computational Linguistics (EMNLP 2025)

Agrawal A, Aralikatti R, Satheesh A, et al. Uncertainty-aware answer selection for improved reasoning in multi-LLM systems [C] //Proc of the Findings of the Association for Computational Linguistics (EMNLP 2025). Stroudsburg,PA: ACL, 2025: 25090-25098

2025
[10]

MiniCheck: Efficient fact-checking of LLMs on grounding documents[C]// Proc of the 2024 Conf on Empirical Methods in Natural Language Processing

Tang Liyan, Laban P, Durrett G. MiniCheck: Efficient fact-checking of LLMs on grounding documents[C]// Proc of the 2024 Conf on Empirical Methods in Natural Language Processing. Stroudsburg,PA: ACL, 2024: 8818–8847

2024
[11]

China home to over one-third of world's AI large language models [N/OL]

Dong jing. China home to over one-third of world's AI large language models [N/OL]. 中国日报 . (2024-07-03)[2026-01-01]. https://language.chinadaily.com.cn/a/202407/03/WS66850bcda31095c51c50c2 9e.html

2024
[12]

Chat GPT-4 significantly surpasses GPT-3.5 in drug information queries[J]

He Na, Yan Yinging, Wu Ziyang, et al. Chat GPT-4 significantly surpasses GPT-3.5 in drug information queries[J]. Journal of Telemedicine and Telecare, 2025, 31(2): 306-308

2025
[13]

Llama-nemotron: Efficient reasoning models [J]

Akhiad B, Itay L, Izik G, et al. Llama-nemotron: Efficient reasoning models [J]. arXiv preprint arXiv:2505.00949, 2025

work page arXiv 2025
[14]

Exploring DeepSeek: A survey on advances, applications, challenges and future directions[J]

Deng Zehang, Ma Wanlun, Han Qing-Long, et al. Exploring DeepSeek: A survey on advances, applications, challenges and future directions[J]. IEEE/CAA Journal of Automatica Sinica, 2025, 12(5): 872-893

2025
[15]

Bai Shuai, Chen Keqin, Liu Xuejing, et al. Qwen2. 5-vl technical report[J]. arXiv preprint arXiv:2502.13923, 2025

work page internal anchor Pith review Pith/arXiv arXiv 2025
[16]

Kimi-VL Technical Report

Team Kimi, Du Angang, Yin Boohong, et al. Kimi-vl technical report[J]. arXiv preprint arXiv:2504.07491, 2025

work page internal anchor Pith review Pith/arXiv arXiv 2025
[17]

Aiersilan A, Liu Mingzhe. LLM-enhanced traffic editor for accelerated testing of autonomous vehicles under various pedestrian behaviors[C]//Proc of the 5th Int Conf on Smart Transportation and City Engineering. Bellingham,WA:SPIE, 2025, 13575: 1054-1060

2025
[18]

Intelligent agents with llm-based process automation[C]// Proc of the 30th ACM SIGKDD Conf on Knowledge Discovery and Data Mining

Guan Yanchu, Wang Dong, Chu Zhixuan, et al. Intelligent agents with llm-based process automation[C]// Proc of the 30th ACM SIGKDD Conf on Knowledge Discovery and Data Mining. New York: ACM, 2024: 5018-5027

2024
[19]

Automated literature research and review-generation method based on large language models[J]

Wu Shican, Ma Xiao, Luo Dehui, et al. Automated literature research and review-generation method based on large language models[J]. National Science Review, 2025, 12(6): nwaf169

2025
[20]

Xu Derong, Zhang Ziheng, Zhu Zhihong, et al. Mitigating hallucinations of large language models in medical information extraction via contrastive decoding[C]// Proc of the Findings of the Association for Computational Linguistics（EMNLP 2024）. Stroudsburg,PA: ACL, 2024. 2024: 7744-7757

2024
[21]

Multi-tier multi-node scheduling of llm for collaborative AI computing[C]// Proc of the 2025 IEEE Conf on Computer Communications (IEEE INFOCOM 2025)

Ma Mulei, Gong Chenyu, Zeng Liekang, et al. Multi-tier multi-node scheduling of llm for collaborative AI computing[C]// Proc of the 2025 IEEE Conf on Computer Communications (IEEE INFOCOM 2025). Piscataway,NJ: IEEE, 2025: 1-10

2025
[22]

Merge, ensemble, and cooperate! a survey on collaborative strategies in the era of large language models[J]

Lu Jinliang, Pang Ziliang, Xiao Min, et al. Merge, ensemble, and cooperate! a survey on collaborative strategies in the era of large language models[J]. arXiv preprint arXiv:2407.06089, 2024

work page arXiv 2024
[23]

Ties-merging: Resolving interference when merging models[J]

Yadav P, Tam D, Choshen L, et al. Ties-merging: Resolving interference when merging models[J]. Advances in Neural Information Processing Systems, 2023, 36: 7093-7115

2023
[24]

Twin-merging: dynamic integration of modular expertise in model merging[J]

Lu Zheyi, Fan Chenghao, Wei Wei, et al. Twin-merging: dynamic integration of modular expertise in model merging[J]. Advances in Neural Information Processing Systems, 2024, 37: 78905-78935

2024
[25]

Ensemble learning for heterogeneous large language models with deep parallel collaboration[J]

Huang Yichong, Feng Xiaocheng, Li Baohang, et al. Ensemble learning for heterogeneous large language models with deep parallel collaboration[J]. Advances in Neural Information Processing Systems, 2024, 37: 119838-119860

2024
[26]

Token-level collaborative reasoning for parallel multi-models[J]

Wang jianhui, Lizhetao, Wu Tao, et al. Token-level collaborative reasoning for parallel multi-models[J]. Chinese Journal of Computers, 2025, 48(11): 2579-2593 (in Chinese) （王建辉, 李哲涛, 伍涛, 等. Token 级多模型并联协作推理[J].计算机学报, 2025, 48(11): 2579-2593）

2025
[27]

Dynamic model routing based on collaborative relationship [J/OL]

Wu Junru, Li Zhetao, Wang Jianhui, et al. Dynamic model routing based on collaborative relationship [J/OL]. Journal of Software, 2025[2026-01-02]. http://www.jos.org.cn/1000-9825/7498.html (in Chinese) （吴俊儒,李哲涛,王建辉,等. 基于协作关系的模型动态路由. [J/OL]. 软件学报, 2025[2026-01-02]. http://www.jos.org.cn/1000-9825/7498.html）

2025
[28]

Unifying large language models and knowledge graphs: a roadmap[J]

Pan Shirui, Luo Linhao, Wang Yufei, et al. Unifying large language models and knowledge graphs: a roadmap[J]. IEEE Transactions on Knowledge and Data Engineering, 2024, 36(7): 3580-3599

2024
[29]

From role-play to drama-interaction: An LLM solution [C]// Proc of the Findings of the Association for Computational Linguistics (ACL 2024 )

Wu Weiqi, Wu Hongqiu, Jiang Lai, et al. From role-play to drama-interaction: An LLM solution [C]// Proc of the Findings of the Association for Computational Linguistics (ACL 2024 ). Stroudsburg,PA: ACL, 2024: 3271-3290

2024
[30]

RouteLLM: Learning to Route LLMs with Preference Data

Ong I, Almahairi A, Wu V, et al. Routellm: Learning to route llms with preference data[J]. arXiv preprint arXiv:2406.18665, 2024

work page internal anchor Pith review Pith/arXiv arXiv 2024
[31]

BEST-Route: Adaptive LLM routing with test-time optimal compute[C/OL]// Proc of the 42nd Int Conf on Machine Learning

Ding Dujian, Mallick A, Zhang Shaokun, et al. BEST-Route: Adaptive LLM routing with test-time optimal compute[C/OL]// Proc of the 42nd Int Conf on Machine Learning. New York: ACM, 2025[2026-01-03]. https://openreview.net/forum?id=tFBIbCVXkG

2025
[32]

GraphRouter: A graph-based router for LLM selections[C]// Proc of the 13th Int Conf on Learning Representations

Feng Tao, Shen Yanzhen, You Jiaxuan. GraphRouter: A graph-based router for LLM selections[C]// Proc of the 13th Int Conf on Learning Representations. New York: ACM, 2025[2026-01-03]. https://openreview.net/forum?id=eU39PDsZtT 16 Journal of Computer Research and Development 2025 年

2025
[33]

RCR-Router: Efficient role-aware context routing for multi-agent LLM systems with structured memory[J]

Liu Jun, Kong Zhenglun, Yang Changdi, et al. RCR-Router: Efficient role-aware context routing for multi-agent LLM systems with structured memory[J]. arXiv preprint, arXiv:2508.04903, 2025

work page arXiv 2025
[34]

TensorOpera router: A multi-model router for efficient LLM inference[C]// Proc of the 2024 Conf on Empirical Methods in Natural Language Processing (EMNLP 2024)

Stripelis D, Xu Zhaozhou, Hu Zijian, et al. TensorOpera router: A multi-model router for efficient LLM inference[C]// Proc of the 2024 Conf on Empirical Methods in Natural Language Processing (EMNLP 2024). Stroudsburg,PA: ACL, 2024: 452-462

2024
[35]

Improving factuality and reasoning in language models through multiagent debate[C]//Proc of the 41st Int Conf on Machine Learning

Du Yilun, Li Shuang, Torralba A, et al. Improving factuality and reasoning in language models through multiagent debate[C]//Proc of the 41st Int Conf on Machine Learning. New York: ACM, 2024:11733-11763

2024
[36]

C-eval: A multi-level multi-discipline chinese evaluation suite for foundation models [C]// Proc of the 37th Conf on Neural Information Processing Systems (NeurIPS 2023)

Huang Yuzhen, Bai Yuzhou, Zhu Zhihao, et al. C-eval: A multi-level multi-discipline chinese evaluation suite for foundation models [C]// Proc of the 37th Conf on Neural Information Processing Systems (NeurIPS 2023). New York: ACM, 2023: 62991-63010

2023
[37]

Clark C, Lee K, Chang Mingwei, et al. BoolQ: Exploring the Surprising Difficulty of Natural Yes/No Questions[C]// Proc of the 2019 Conf of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. Stroudsburg,PA: ACL, 2019: 2924-2936

2019
[38]

Measuring Massive Multitask Language Understanding

Hendrycks D, Burns C, Basart S, et al. Measuring massive multitask language understanding[J]. arXiv preprint arXiv:2009.03300, 2020

work page internal anchor Pith review Pith/arXiv arXiv 2009
[39]

IRT-Router: Effective and interpretable multi-LLM routing via item response theory[C]// Proc of the 63rd Annual Meeting of the Association for Computational Linguistics

Song Wei, Huang Zhenya, Cheng Cheng, et al. IRT-Router: Effective and interpretable multi-LLM routing via item response theory[C]// Proc of the 63rd Annual Meeting of the Association for Computational Linguistics. Stroudsburg,PA: ACL, 2025: 15629-15644

2025
[40]

TETRIS: optimal draft token selection for batch speculative decoding

Wu Zhaoxuan, Zhou Zijian, Verma1 A, et al. TETRIS: optimal draft token selection for batch speculative decoding. [C]// Proc of the 63rd Annual Meeting of the Association for Computational Linguistics. Stroudsburg,PA: ACL, 2025: 33329-33345

2025
[41]

Announcing the agent2agent protocol (A2A)[EB/OL].2025[2025-10-19]

Rao S, Philip S. Announcing the agent2agent protocol (A2A)[EB/OL].2025[2025-10-19]. https://developers.googleblog.com/en/a2a-a-new-era-of-agent-interoperability/

2025
[42]

The stepwise deception: Simulating the evolution from true news to fake news with llm agents[C]// Proc of the 2025 Conf on Empirical Methods in Natural Language Processing

Liu Yuhan, Song Zirui, Zhang Juntian, et al. The stepwise deception: Simulating the evolution from true news to fake news with llm agents[C]// Proc of the 2025 Conf on Empirical Methods in Natural Language Processing. Stroudsburg,PA: ACL, 2025: 26187-26203

2025
[43]

A survey on trustworthy LLM agents: Threats and countermeasures[C]// Proc of the 31st ACM SIGKDD Conf on Knowledge Discovery and Data Mining V

Yu Miao, Meng Fanci, Zhou Xinyun, et al. A survey on trustworthy LLM agents: Threats and countermeasures[C]// Proc of the 31st ACM SIGKDD Conf on Knowledge Discovery and Data Mining V. 2. New York: ACM, 2025: 6216-6226

2025
[44]

No free lunch theorem for privacy-preserving LLM inference[J]

Zhang Xiaojin, Pang Yahao, Kang Yan, et al. No free lunch theorem for privacy-preserving LLM inference[J]. Artificial Intelligence, 2025, 341: 104293

2025
[45]

三个臭皮匠，顶个诸葛亮

IBM BeeAI. Introduction to agent communication protocol (acp). [EB/OL] (2024-12-03)[2026-01-03]. https://docs.beeai.dev/acp/alpha/introduction 模型互联网：概念、现状和未来李哲涛曾曦玉王建辉肖勇刘忠仁吴俊儒赖俊杰黄纪俊龙赛琴（暨南大学信息科学技术学院广州 510632）（liztchina@hotmail.com） AI-Model Network: Concept, Current State and Future Li Zhetao, Zeng Xiyu, Wang Jianhui, Xiao Yong, Liu Zhongren, W...

work page doi:10.7544/issn1000-1239.202550223 2024
[46]

**冰**：纯冰在0 ℃时熔化，熔化潜热约为334 J/g。这意味着它需要吸收大量的热量才能从固态转变为液态。
[47]

**冰+食盐**：食盐 (NaCl) 溶于冰水中可以降低水的凝固点，但不会显著增加制冷效果，因为食盐溶解后主要影响的是水的凝固点，而不是直接增加冰的熔化潜热。
[48]

**冰+$CaCl_2·6H_2O$**：$CaCl_2·6H_2O$ (六水合氯化钙) 是一种吸湿性很强的化合物，在吸湿过程中会释放大量热量，但其主要作用是吸湿而非直接作为制冷剂。因此，它对制冷效果的影响有限。
[49]

RouterBench: A Benchmark for Multi-LLM Routing System

**冰+$CaCl_2$**：$CaCl_2$ (氯化钙) 在水中溶解时会释放大量的热量，这会导致溶液温度显著下降。这种现象称为 “溶解热”，可以显著提高制冷效果。因此，加入氯化钙可以显著增强冰的制冷效果。综合以上分析，**冰+$CaCl_2$**组合的制冷效果最好，因为它利用了氯化钙溶解时释放的大量热量来进一步降低温度。 D 为了确定制冷效果最好的选项，我们需要考虑每种组合的熔点和吸热能力。冰的熔点是0 ℃，而加入食盐或$CaCl_2$会降低冰的熔点，使其在更低的温度下融化，从而吸收更多的热量。 -A. 冰：在 0 ℃时融化，吸热能力一般。 -B. 冰+食盐：食盐溶于水会形成盐水溶液，其凝固点低于0 ℃，但具体数值取决于食盐的浓度。通常，10%的食盐水可以将冰点降低到 −18 ℃左右，制冷效果较好...

work page internal anchor Pith review Pith/arXiv arXiv 2026

[1] [1]

The world-wide web[J]

Berners-Lee T, Cailliau R, Luotonen A, et al. The world-wide web[J]. Communications of the ACM, 1994, 37(8): 76-82

1994

[2] [2]

Attention is all you need [C] // Proc of the 31st Conf on Neural Information Processing Systems

Vaswani A, Shazeer N, Parmar N, et al. Attention is all you need [C] // Proc of the 31st Conf on Neural Information Processing Systems. New York: ACM, 2017: 6000–6010

2017

[3] [3]

Investigating bias in LLM-based bias detection: Disparities between LLMs and human perception [C] //Proc of the 31st Int Conf on Computational Linguistics

Lin Luyuan, Wang Lingzhi, Guo Jinsong, et al. Investigating bias in LLM-based bias detection: Disparities between LLMs and human perception [C] //Proc of the 31st Int Conf on Computational Linguistics. Stroudsburg,PA: ACL, 2025: 10634-10649

2025

[4] [4]

Thinking on networking and computing promoted by domain-specific method [J]

Su Jinshu. Thinking on networking and computing promoted by domain-specific method [J]. Journal of Computer Research and Development, 2025, 62(12): 2889-2894(in Chinese) （苏金树. 领域定制方法促进网络与计算发展的思考[J]. 计算机研究与发展, 2025, 62(12): 2889-2894）

2025

[5] [5]

Don’t hallucinate, abstain: Identifying LLM knowledge gaps via multi-LLM collaboration [C] //Proc of the 62nd Annual Meeting of the Association for Computational Linguistics

Feng Shagbin, Shi Weijia, Wang Yike, et al. Don’t hallucinate, abstain: Identifying LLM knowledge gaps via multi-LLM collaboration [C] //Proc of the 62nd Annual Meeting of the Association for Computational Linguistics. Stroudsburg,PA: ACL, 2024: 14664-14690 15 Journal of Computer Research and Development 2025 年

2024

[6] [6]

Rethinking the reversal curse of LLMs: A prescription from human knowledge reversal [C] //Proc of the 2024 Conf on Empirical Methods in Natural Language Processing

Lu Zhicong, Li Jin, Li Peiguang, et al. Rethinking the reversal curse of LLMs: A prescription from human knowledge reversal [C] //Proc of the 2024 Conf on Empirical Methods in Natural Language Processing. Stroudsburg,PA: ACL, 2024: 7518-7530

2024

[7] [7]

Small LLMs are weak tool learners: A multi-LLM agent [C] //Proc of the 2024 Conf on Empirical Methods in Natural Language Processing

Shen Weizhou, Li Chenliang, Chen Hongzhan, et al. Small LLMs are weak tool learners: A multi-LLM agent [C] //Proc of the 2024 Conf on Empirical Methods in Natural Language Processing. Stroudsburg,PA: ACL, 2024: 16658-16680 (没有届)

2024

[8] [9]

Uncertainty-aware answer selection for improved reasoning in multi-LLM systems [C] //Proc of the Findings of the Association for Computational Linguistics (EMNLP 2025)

Agrawal A, Aralikatti R, Satheesh A, et al. Uncertainty-aware answer selection for improved reasoning in multi-LLM systems [C] //Proc of the Findings of the Association for Computational Linguistics (EMNLP 2025). Stroudsburg,PA: ACL, 2025: 25090-25098

2025

[9] [10]

MiniCheck: Efficient fact-checking of LLMs on grounding documents[C]// Proc of the 2024 Conf on Empirical Methods in Natural Language Processing

Tang Liyan, Laban P, Durrett G. MiniCheck: Efficient fact-checking of LLMs on grounding documents[C]// Proc of the 2024 Conf on Empirical Methods in Natural Language Processing. Stroudsburg,PA: ACL, 2024: 8818–8847

2024

[10] [11]

China home to over one-third of world's AI large language models [N/OL]

Dong jing. China home to over one-third of world's AI large language models [N/OL]. 中国日报 . (2024-07-03)[2026-01-01]. https://language.chinadaily.com.cn/a/202407/03/WS66850bcda31095c51c50c2 9e.html

2024

[11] [12]

Chat GPT-4 significantly surpasses GPT-3.5 in drug information queries[J]

He Na, Yan Yinging, Wu Ziyang, et al. Chat GPT-4 significantly surpasses GPT-3.5 in drug information queries[J]. Journal of Telemedicine and Telecare, 2025, 31(2): 306-308

2025

[12] [13]

Llama-nemotron: Efficient reasoning models [J]

Akhiad B, Itay L, Izik G, et al. Llama-nemotron: Efficient reasoning models [J]. arXiv preprint arXiv:2505.00949, 2025

work page arXiv 2025

[13] [14]

Exploring DeepSeek: A survey on advances, applications, challenges and future directions[J]

Deng Zehang, Ma Wanlun, Han Qing-Long, et al. Exploring DeepSeek: A survey on advances, applications, challenges and future directions[J]. IEEE/CAA Journal of Automatica Sinica, 2025, 12(5): 872-893

2025

[14] [15]

Bai Shuai, Chen Keqin, Liu Xuejing, et al. Qwen2. 5-vl technical report[J]. arXiv preprint arXiv:2502.13923, 2025

work page internal anchor Pith review Pith/arXiv arXiv 2025

[15] [16]

Kimi-VL Technical Report

Team Kimi, Du Angang, Yin Boohong, et al. Kimi-vl technical report[J]. arXiv preprint arXiv:2504.07491, 2025

work page internal anchor Pith review Pith/arXiv arXiv 2025

[16] [17]

Aiersilan A, Liu Mingzhe. LLM-enhanced traffic editor for accelerated testing of autonomous vehicles under various pedestrian behaviors[C]//Proc of the 5th Int Conf on Smart Transportation and City Engineering. Bellingham,WA:SPIE, 2025, 13575: 1054-1060

2025

[17] [18]

Intelligent agents with llm-based process automation[C]// Proc of the 30th ACM SIGKDD Conf on Knowledge Discovery and Data Mining

Guan Yanchu, Wang Dong, Chu Zhixuan, et al. Intelligent agents with llm-based process automation[C]// Proc of the 30th ACM SIGKDD Conf on Knowledge Discovery and Data Mining. New York: ACM, 2024: 5018-5027

2024

[18] [19]

Automated literature research and review-generation method based on large language models[J]

Wu Shican, Ma Xiao, Luo Dehui, et al. Automated literature research and review-generation method based on large language models[J]. National Science Review, 2025, 12(6): nwaf169

2025

[19] [20]

Xu Derong, Zhang Ziheng, Zhu Zhihong, et al. Mitigating hallucinations of large language models in medical information extraction via contrastive decoding[C]// Proc of the Findings of the Association for Computational Linguistics（EMNLP 2024）. Stroudsburg,PA: ACL, 2024. 2024: 7744-7757

2024

[20] [21]

Multi-tier multi-node scheduling of llm for collaborative AI computing[C]// Proc of the 2025 IEEE Conf on Computer Communications (IEEE INFOCOM 2025)

Ma Mulei, Gong Chenyu, Zeng Liekang, et al. Multi-tier multi-node scheduling of llm for collaborative AI computing[C]// Proc of the 2025 IEEE Conf on Computer Communications (IEEE INFOCOM 2025). Piscataway,NJ: IEEE, 2025: 1-10

2025

[21] [22]

Merge, ensemble, and cooperate! a survey on collaborative strategies in the era of large language models[J]

Lu Jinliang, Pang Ziliang, Xiao Min, et al. Merge, ensemble, and cooperate! a survey on collaborative strategies in the era of large language models[J]. arXiv preprint arXiv:2407.06089, 2024

work page arXiv 2024

[22] [23]

Ties-merging: Resolving interference when merging models[J]

Yadav P, Tam D, Choshen L, et al. Ties-merging: Resolving interference when merging models[J]. Advances in Neural Information Processing Systems, 2023, 36: 7093-7115

2023

[23] [24]

Twin-merging: dynamic integration of modular expertise in model merging[J]

Lu Zheyi, Fan Chenghao, Wei Wei, et al. Twin-merging: dynamic integration of modular expertise in model merging[J]. Advances in Neural Information Processing Systems, 2024, 37: 78905-78935

2024

[24] [25]

Ensemble learning for heterogeneous large language models with deep parallel collaboration[J]

Huang Yichong, Feng Xiaocheng, Li Baohang, et al. Ensemble learning for heterogeneous large language models with deep parallel collaboration[J]. Advances in Neural Information Processing Systems, 2024, 37: 119838-119860

2024

[25] [26]

Token-level collaborative reasoning for parallel multi-models[J]

Wang jianhui, Lizhetao, Wu Tao, et al. Token-level collaborative reasoning for parallel multi-models[J]. Chinese Journal of Computers, 2025, 48(11): 2579-2593 (in Chinese) （王建辉, 李哲涛, 伍涛, 等. Token 级多模型并联协作推理[J].计算机学报, 2025, 48(11): 2579-2593）

2025

[26] [27]

Dynamic model routing based on collaborative relationship [J/OL]

Wu Junru, Li Zhetao, Wang Jianhui, et al. Dynamic model routing based on collaborative relationship [J/OL]. Journal of Software, 2025[2026-01-02]. http://www.jos.org.cn/1000-9825/7498.html (in Chinese) （吴俊儒,李哲涛,王建辉,等. 基于协作关系的模型动态路由. [J/OL]. 软件学报, 2025[2026-01-02]. http://www.jos.org.cn/1000-9825/7498.html）

2025

[27] [28]

Unifying large language models and knowledge graphs: a roadmap[J]

Pan Shirui, Luo Linhao, Wang Yufei, et al. Unifying large language models and knowledge graphs: a roadmap[J]. IEEE Transactions on Knowledge and Data Engineering, 2024, 36(7): 3580-3599

2024

[28] [29]

From role-play to drama-interaction: An LLM solution [C]// Proc of the Findings of the Association for Computational Linguistics (ACL 2024 )

Wu Weiqi, Wu Hongqiu, Jiang Lai, et al. From role-play to drama-interaction: An LLM solution [C]// Proc of the Findings of the Association for Computational Linguistics (ACL 2024 ). Stroudsburg,PA: ACL, 2024: 3271-3290

2024

[29] [30]

RouteLLM: Learning to Route LLMs with Preference Data

Ong I, Almahairi A, Wu V, et al. Routellm: Learning to route llms with preference data[J]. arXiv preprint arXiv:2406.18665, 2024

work page internal anchor Pith review Pith/arXiv arXiv 2024

[30] [31]

BEST-Route: Adaptive LLM routing with test-time optimal compute[C/OL]// Proc of the 42nd Int Conf on Machine Learning

Ding Dujian, Mallick A, Zhang Shaokun, et al. BEST-Route: Adaptive LLM routing with test-time optimal compute[C/OL]// Proc of the 42nd Int Conf on Machine Learning. New York: ACM, 2025[2026-01-03]. https://openreview.net/forum?id=tFBIbCVXkG

2025

[31] [32]

GraphRouter: A graph-based router for LLM selections[C]// Proc of the 13th Int Conf on Learning Representations

Feng Tao, Shen Yanzhen, You Jiaxuan. GraphRouter: A graph-based router for LLM selections[C]// Proc of the 13th Int Conf on Learning Representations. New York: ACM, 2025[2026-01-03]. https://openreview.net/forum?id=eU39PDsZtT 16 Journal of Computer Research and Development 2025 年

2025

[32] [33]

RCR-Router: Efficient role-aware context routing for multi-agent LLM systems with structured memory[J]

Liu Jun, Kong Zhenglun, Yang Changdi, et al. RCR-Router: Efficient role-aware context routing for multi-agent LLM systems with structured memory[J]. arXiv preprint, arXiv:2508.04903, 2025

work page arXiv 2025

[33] [34]

TensorOpera router: A multi-model router for efficient LLM inference[C]// Proc of the 2024 Conf on Empirical Methods in Natural Language Processing (EMNLP 2024)

Stripelis D, Xu Zhaozhou, Hu Zijian, et al. TensorOpera router: A multi-model router for efficient LLM inference[C]// Proc of the 2024 Conf on Empirical Methods in Natural Language Processing (EMNLP 2024). Stroudsburg,PA: ACL, 2024: 452-462

2024

[34] [35]

Improving factuality and reasoning in language models through multiagent debate[C]//Proc of the 41st Int Conf on Machine Learning

Du Yilun, Li Shuang, Torralba A, et al. Improving factuality and reasoning in language models through multiagent debate[C]//Proc of the 41st Int Conf on Machine Learning. New York: ACM, 2024:11733-11763

2024

[35] [36]

C-eval: A multi-level multi-discipline chinese evaluation suite for foundation models [C]// Proc of the 37th Conf on Neural Information Processing Systems (NeurIPS 2023)

Huang Yuzhen, Bai Yuzhou, Zhu Zhihao, et al. C-eval: A multi-level multi-discipline chinese evaluation suite for foundation models [C]// Proc of the 37th Conf on Neural Information Processing Systems (NeurIPS 2023). New York: ACM, 2023: 62991-63010

2023

[36] [37]

Clark C, Lee K, Chang Mingwei, et al. BoolQ: Exploring the Surprising Difficulty of Natural Yes/No Questions[C]// Proc of the 2019 Conf of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. Stroudsburg,PA: ACL, 2019: 2924-2936

2019

[37] [38]

Measuring Massive Multitask Language Understanding

Hendrycks D, Burns C, Basart S, et al. Measuring massive multitask language understanding[J]. arXiv preprint arXiv:2009.03300, 2020

work page internal anchor Pith review Pith/arXiv arXiv 2009

[38] [39]

IRT-Router: Effective and interpretable multi-LLM routing via item response theory[C]// Proc of the 63rd Annual Meeting of the Association for Computational Linguistics

Song Wei, Huang Zhenya, Cheng Cheng, et al. IRT-Router: Effective and interpretable multi-LLM routing via item response theory[C]// Proc of the 63rd Annual Meeting of the Association for Computational Linguistics. Stroudsburg,PA: ACL, 2025: 15629-15644

2025

[39] [40]

TETRIS: optimal draft token selection for batch speculative decoding

Wu Zhaoxuan, Zhou Zijian, Verma1 A, et al. TETRIS: optimal draft token selection for batch speculative decoding. [C]// Proc of the 63rd Annual Meeting of the Association for Computational Linguistics. Stroudsburg,PA: ACL, 2025: 33329-33345

2025

[40] [41]

Announcing the agent2agent protocol (A2A)[EB/OL].2025[2025-10-19]

Rao S, Philip S. Announcing the agent2agent protocol (A2A)[EB/OL].2025[2025-10-19]. https://developers.googleblog.com/en/a2a-a-new-era-of-agent-interoperability/

2025

[41] [42]

The stepwise deception: Simulating the evolution from true news to fake news with llm agents[C]// Proc of the 2025 Conf on Empirical Methods in Natural Language Processing

Liu Yuhan, Song Zirui, Zhang Juntian, et al. The stepwise deception: Simulating the evolution from true news to fake news with llm agents[C]// Proc of the 2025 Conf on Empirical Methods in Natural Language Processing. Stroudsburg,PA: ACL, 2025: 26187-26203

2025

[42] [43]

A survey on trustworthy LLM agents: Threats and countermeasures[C]// Proc of the 31st ACM SIGKDD Conf on Knowledge Discovery and Data Mining V

Yu Miao, Meng Fanci, Zhou Xinyun, et al. A survey on trustworthy LLM agents: Threats and countermeasures[C]// Proc of the 31st ACM SIGKDD Conf on Knowledge Discovery and Data Mining V. 2. New York: ACM, 2025: 6216-6226

2025

[43] [44]

No free lunch theorem for privacy-preserving LLM inference[J]

Zhang Xiaojin, Pang Yahao, Kang Yan, et al. No free lunch theorem for privacy-preserving LLM inference[J]. Artificial Intelligence, 2025, 341: 104293

2025

[44] [45]

三个臭皮匠，顶个诸葛亮

IBM BeeAI. Introduction to agent communication protocol (acp). [EB/OL] (2024-12-03)[2026-01-03]. https://docs.beeai.dev/acp/alpha/introduction 模型互联网：概念、现状和未来李哲涛曾曦玉王建辉肖勇刘忠仁吴俊儒赖俊杰黄纪俊龙赛琴（暨南大学信息科学技术学院广州 510632）（liztchina@hotmail.com） AI-Model Network: Concept, Current State and Future Li Zhetao, Zeng Xiyu, Wang Jianhui, Xiao Yong, Liu Zhongren, W...

work page doi:10.7544/issn1000-1239.202550223 2024

[45] [46]

**冰**：纯冰在0 ℃时熔化，熔化潜热约为334 J/g。这意味着它需要吸收大量的热量才能从固态转变为液态。

[46] [47]

**冰+食盐**：食盐 (NaCl) 溶于冰水中可以降低水的凝固点，但不会显著增加制冷效果，因为食盐溶解后主要影响的是水的凝固点，而不是直接增加冰的熔化潜热。

[47] [48]

**冰+$CaCl_2·6H_2O$**：$CaCl_2·6H_2O$ (六水合氯化钙) 是一种吸湿性很强的化合物，在吸湿过程中会释放大量热量，但其主要作用是吸湿而非直接作为制冷剂。因此，它对制冷效果的影响有限。

[48] [49]

RouterBench: A Benchmark for Multi-LLM Routing System

**冰+$CaCl_2$**：$CaCl_2$ (氯化钙) 在水中溶解时会释放大量的热量，这会导致溶液温度显著下降。这种现象称为 “溶解热”，可以显著提高制冷效果。因此，加入氯化钙可以显著增强冰的制冷效果。综合以上分析，**冰+$CaCl_2$**组合的制冷效果最好，因为它利用了氯化钙溶解时释放的大量热量来进一步降低温度。 D 为了确定制冷效果最好的选项，我们需要考虑每种组合的熔点和吸热能力。冰的熔点是0 ℃，而加入食盐或$CaCl_2$会降低冰的熔点，使其在更低的温度下融化，从而吸收更多的热量。 -A. 冰：在 0 ℃时融化，吸热能力一般。 -B. 冰+食盐：食盐溶于水会形成盐水溶液，其凝固点低于0 ℃，但具体数值取决于食盐的浓度。通常，10%的食盐水可以将冰点降低到 −18 ℃左右，制冷效果较好...

work page internal anchor Pith review Pith/arXiv arXiv 2026