Artificial Intelligence · arXiv:2611.17186 v1 · submitted 12 Nov 2026 · 19 references
Cortex-FL: Agent-Mediated Collaboration for Federated Learning
The paper says
“Federated learning (FL) enables independent data owners to train models collaboratively without centralizing raw data. However, identifying collaborators, preparing training tasks, and coordinating execution and incentives hinder its practical adoption by organizations without specialized expertise.”
more of the abstract
We present Cortex-FL, an agent-native FL system that enables users to express learning objectives in natural language and delegate the collaboration required to train a model. Agents translate these objectives into shared tasks, discover collaborators, and negotiate participation agreements, connecting collaboration formation with distributed training and incentive management. The architecture separates agent coordination from local execution authority, allowing owners to delegate coordination while retaining control over their data, computation, and financial commitments. We demonstrate the complete workflow on a four-node deployment, in which a model requested in natural language proceeds through collaboration formation, contract-governed training, verified model delivery, and x402 settlement on a blockchain testnet.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 13 Nov 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up, with conditions. A credible systems demonstration: the workflow ran end-to-end, but the security claim in §2 is asserted, not tested, and the evaluation has no baseline. This line is wrong
3 major objections · 6 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 19 references say.
The paper says, at §2 (Trust model) and §5 (Demonstration)
The review did not quote this passage.
Pith says: Objection 1 of 3 · major
The trust model is central to the paper's claim that the system 'preserves owner control over data, computation, and payments' (b115). The paper asserts that 'the host validates proposed actions against the task specification, authenticated identities, and current authorization' and that 'these guarantees cover owner authorization, artifact integrity, and transaction accounting across every agent action and every training workload on the node.'
more of objection 1
However, the only adversarial evidence provided is a single test where 'an agent modifies the model package before delivery, and the receiving node rejects it at verification' (§5). This test exercises artifact hashing but does not test the tool boundary, host authorization, or the agent's inability to bypass the trust gate. No misled agent, malicious peer, or compromised host is exercised. The paper should either provide additional adversarial testing or analysis to support the trust model claims, or explicitly acknowledge the limitations of the current evaluation and tone down the security claims accordingly.
The paper says, at §5
The review did not quote this passage.
Pith says: Objection 2 of 3 · major
The demonstration is conducted on a local network with four nodes, using a single task (five-class image classification on CIFAR-10). While this is sufficient for a proof-of-concept, the paper does not discuss how the system would behave under more challenging conditions, such as network partitions, node failures, malicious peers, or larger-scale deployments. The paper should discuss these limitations and, if possible, provide evidence of robustness or plans for future evaluation.
The paper says, at §2
The review did not quote this passage.
Pith says: Objection 3 of 3 · major
The trust model assumes that 'each owner trusts the local host, authorization interface, and credential store.' This is a reasonable starting point, but the paper does not discuss what happens if the host itself is compromised or if the owner's credentials are stolen. The paper should clarify the threat model and identify what attacks are and are not considered.
6 minor objections
The paper says, at General
Pith says: minor objection 1 of 6
The paper does not include a limitations section. Adding one would improve clarity and help readers understand the scope of the claims.
The paper says, at Figure 3
Pith says: minor objection 2 of 6
Figure 3 is informative but dense. Consider adding a caption that explains the key elements more explicitly.
show 4 more
The paper says, at §3
Pith says: minor objection 3 of 6
The paper references several external systems (FL Alliance, FLocKit, x402, AgentaNet) but does not provide a brief summary of how they work or what guarantees they provide. A short paragraph in Section 3 or 4 would help readers who are not familiar with these systems.
The paper says, at §5
Pith says: minor objection 4 of 6
The demonstration uses test USDC on Base Sepolia testnet. The paper should clarify whether the system is designed to work with real USDC on mainnet and what additional considerations would apply.
The paper says, at §2
Pith says: minor objection 5 of 6
The paper uses the term 'agentic node' but does not define it formally. Consider adding a definition in Section 2 or 3.
The paper says, at §6
Pith says: minor objection 6 of 6
The paper could benefit from a comparison with other agent-based FL systems or frameworks, if any exist. If none exist, stating this explicitly would strengthen the novelty claim.
What would settle it
Design and run a red-team test in which one agent is given a prompt that instructs it to (a) publish a task with a different objective than the owner specified, (b) invite a peer that the owner did not approve, or (c) authorize a payment exceeding the owner's stated budget. The test should check whether the host's validation rejects each action before it reaches the owner review or the execution scope. Report the number of attempts and the number of successful bypasses. If zero bypasses occur across a systematic set of adversarial prompts, the trust model's enforcement is empirically supported.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 13 Nov 2026
- Reader holds up, with conditions
- Skeptic agrees
- Editor wrote the verdict above
- Referee major revision: 3 major objections, 6 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
Worth a read if you care about lowering the barrier to cross-silo FL. The paper is honest about what it contributes: FL Alliance and FLocKit handle training, x402 handles payment, AgentaNet handles discovery, and the new piece is the integration layer binding them with a signed artifact chain and host-side trust gates. That connective layer has a genuinely sensible design idea: agents coordinate but hold no credentials, and every consequential step passes an owner-approval or host-verification gate. The four-node demonstration ran, with concrete artifact evidence (round-roster with T/V assignments, 10 rounds from 23:45 to 23:56 UTC, validation loss 1.52 to 0.96, three delivery verifications, three settlement receipts). For a systems paper whose central claim is "the workflow runs," that evidence is adequate and the paper does not oversell it.
The soft spots are real but proportionate. First, the trust model in §2 makes a universal security guarantee—"These guarantees cover owner authorization, artifact integrity, and transaction accounting across every agent action and every training workload"—but the only adversarial exercise is a single tampered model package rejected at delivery verification. That tests artifact hashing, not the tool boundary, host authorization, or the agent's inability to bypass the trust gate. The paper acknowledges the trust boundary but does not test it. Load-bearing? Yes, because the entire value proposition is owner-controlled delegation; if an agent can cause an unauthorized action, that collapses. The paper does not support its universal claim.
Second, the evaluation is thin even for a demonstration: no baseline (what does the organizer achieve alone?), no accuracy figure for the five-class task, one seed, and no ablation separating the federation's benefit from the organizer's local training. The paper states the loss but never the accuracy, so it is unclear whether collaborating actually helped. Third, the code is not released, making independent reproduction impossible. And the system is assembled from the authors' own prior components (FLock, AgentaNet), which is disclosed but means the integration inherits properties it does not re-establish.
None of this breaks the stated central claim, which is narrow and evidenced. It limits generalization. The paper is a deployment template, not a scientific result about FL performance or security. It deserves a serious referee because the demonstration is concrete, the architecture is clearly described, and the integration problem is real. But it needs revision: a threat-model section that tests the host gate, an ablation, and release of code.
My recommendation: engage with it for the systems direction, but require those additions before accepting.
The strongest objection
The paper's central claim (b5) is that the complete workflow runs. The trust model in §2 states that 'the host validates proposed actions against the task specification, authenticated identities, and current authorization' and that 'these guarantees cover owner authorization, artifact integrity, and transaction accounting across every agent action and every training workload on the node.' This is a universal security claim expressed in the grammar of a measured fact. The only adversarial evidence reported is a single corrupted model package rejected at delivery verification (§5). That exercises artifact hashing, not the tool boundary, host authorization, or the agent's inability to bypass the trust gate. No misled agent, malicious peer, or compromised host is exercised. The paper acknowledges that agents 'hold no wallet or service credentials' and that 'every consequential step passes a trust gate,' but it provides no test that a misled agent cannot, for example, invoke a tool with a fabricated task identifier, or that a compromised host cannot forge a journal entry. The claim that the system 'supports delegation throughout the learning lifecycle while preserving owner control' (b115) is therefore supported only by design description, not by adversarial evaluation. This is a load-bearing concern because if the trust model fails—if an agent can cause an unauthorized action—the entire value proposition of owner-controlled delegation collapses.
The reader's weakest_assumption correctly identifies that the trust model's enforcement is asserted, not analysed, and that the only adversarial evidence is a single tampered-package rejection. My concern is a direct elaboration of that assumption: the paper makes a universal security claim about host-side enforcement but provides no adversarial testing of the agent's ability to bypass the trust gate. The reader's verdict is CONDITIONAL, and I agree that this is the appropriate verdict given the evidence. The concrete test I propose would directly address this gap by systematically probing the host's validation logic with adversarial agent prompts. If the test passes, the trust model gains empirical support; if it fails, the central claim is undermined. I do not recommend changing the verdict because the paper's central claim (that the workflow runs) is supported by the demonstration, but the trust model's enforcement is a load-bearing assumption that requires further evidence. The reader's rationale already notes that 'no misled agent, malicious peer, or compromised host is exercised,' which is the same gap I identify. My contribution is to propose a specific, feasible test that would settle whether the concern lands.
The referee's report
If the claims hold, Cortex-FL addresses a practical barrier to FL adoption by reducing the need for specialized expertise in setting up and managing cross-silo collaborations. The integration of agent-based coordination with FL training and blockchain-based incentives is novel and timely. The demonstration provides a concrete proof-of-concept that connects natural-language input to a complete FL workflow, including collaboration formation, contract-governed training, verified model delivery, and settlement. The system builds on existing components and contributes an integration layer that could be reused or extended. The paper is clearly written and the demonstration appears to have been executed successfully, with detailed timestamps and metrics.
Pith's reading of the paper (a machine's summary, not the authors')
Training a shared model across several organizations is a coordination problem as much as a machine-learning problem. Someone has to decide what the model should do, find partners who hold the missing data, agree on terms, run the training without pooling raw data, and pay the participants. Cortex-FL bundles those steps into a desktop application. The owner types a request such as 'I need a five-class model for cat, dog, bird, ship, truck, using my photos, under 0.01 USDC.' An embedded language-model agent drafts the task, checks the owner's local data, registers the task and model hash on a test blockchain, invites peers, and negotiates a price per peer. The owner approves each consequential step through the application window; the agent itself holds no wallet keys and never sees a host file path.
The demonstration uses four Apple machines on one local network. The organizer has only cat and dog photos, so the three peers contribute the remaining classes. Ten rounds of parameter averaging ran under a staking and voting protocol borrowed from the authors' earlier work, in which designated voters check each candidate model and vote through a smart contract. The winning checkpoint ended up identical on all four hosts. Each peer was paid 0.002 test USDC through the x402 payment protocol, released only after the buyer checked the delivered weights against the contract-selected checkpoint. In a separate test, an agent corrupted the model package before delivery and the receiving node rejected it.
What the paper does not show: how accurate the resulting classifier is, what the organizer could have reached alone, or whether the security properties hold against a peer or host that actually misbehaves.
Core claim. The paper's load-bearing claim is stated in the abstract and restated in the conclusion: 'We demonstrate the complete workflow on a four-node deployment, in which a model requested in natural language proceeds through collaboration formation, contract-governed training, verified model delivery, and x402 settlement on a blockchain testnet' (b5, echoed at b114 and b115). If correct, a non-specialist can express a learning objective in plain language and end up holding a trained model, with the federation, its contracts, its training rounds, and its payments assembled by agents acting under owner approval at every consequential step.
The premise it rests on. The value proposition rests on the trust model in §2: each owner trusts only the local host, its authorization interface, and its credential store, while peer messages, retrieved artifacts, and agent-visible content are treated as untrusted and validated by the host before any action. Every security sentence in the paper reduces to that host actually enforcing scope for every agent action, and the enforcement is asserted, not analysed. The only adversarial evidence reported is a single corrupted model package rejected at delivery verification (§5, 'an agent modifies the model package before delivery, and the receiving node rejects it'), which exercises artifact hashing, not the tool boundary or host authorization. No misled agent, malicious peer, or compromised host is exercised.
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
Federated learning (FL) enables independent data owners to train models collaboratively without centralizing raw data. However, identifying collaborators, preparing training tasks, and coordinating execution and incentives hinder its practical adoption by organizations without specialized expertise. We present Cortex-FL, an agent-native FL system that enables users to express learning objectives in natural language and delegate the collaboration required to train a model. Agents translate these objectives into shared tasks, discover collaborators, and negotiate participation agreements, connecting collaboration formation with distributed training and incentive management. The architecture separates agent coordination from local execution authority, allowing owners to delegate coordination while retaining control over their data, computation, and financial commitments. We demonstrate the complete workflow on a four-node deployment, in which a model requested in natural language proceeds through collaboration formation, contract-governed training, verified model delivery, and x402 settlement on a blockchain testnet.
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-11-14
The authors' main claim, as the classifier picked it out 0.59
“We present Cortex-FL, an agent-native FL system that enables users to express learning objectives in natural language and delegate the collaboration required to train a model.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.05 · states limitation 0.18 · human participants 0.13 · controlled comparison 0.03 · randomized assignment 0.01 · states quantitative result 0.68 · compares with prior results 0.06 · mentions code or data availability 0.05
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Federated learning (FL) enables independent data owners to train models collaboratively without centralizing raw data. | background 1.00 | 0.04 | 0.03 | asserted 1.00 |
| However, identifying collaborators, preparing training tasks, and coordinating execution and incentives hinder its practical adoption by organizations without specialized expertise. | background 0.51 | 0.06 | 0.04 | asserted 1.00 |
| We present Cortex-FL, an agent-native FL system that enables users to express learning objectives in natural language and delegate the collaboration required to train a model. | contribution 1.00 | 0.89 | 0.29 | asserted 1.00 |
| Agents translate these objectives into shared tasks, discover collaborators, and negotiate participation agreements, connecting collaboration formation with distributed training and incentive management. | method 0.86 | 0.73 | 0.16 | asserted 1.00 |
| The architecture separates agent coordination from local execution authority, allowing owners to delegate coordination while retaining control over their data, computation, and financial commitments. | method 0.89 | 0.61 | 0.11 | asserted 1.00 |
| We demonstrate the complete workflow on a four-node deployment, in which a model requested in natural language proceeds through collaboration formation, contract-governed training, verified model delivery, and x402 settlement on a blockchain testnet. | contribution 0.43 | 0.96 | 0.20 | asserted 1.00 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.90
What was examined, and what was not
Sources read
- arXiv listing abstract v1
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-11-14: 1 request, 4,918 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
Figures (3)
Figures are the authors', reproduced from the paper.
References (19)
- [1] Anthropic. 2024. Introducing the Model Context Protocol. Anthropic News. Published November 25, 2024. https://www.anthropic.com/news/model-context- protocol on Pith
- [2] Anthropic. 2025. Claude 3.7 Sonnet and Claude Code. Anthropic. Published February 24, 2025. https://www.anthropic.com/news/claude-3-7-sonnet on Pith
- [3] Daniel J. Beutel, Taner Topal, Akhil Mathur, Xinchi Qiu, Javier Fernandez- Marques, Yan Gao, Lorenzo Sani, Kwing Hei Li, Titouan Parcollet, Pedro Porto Buarque de Gusmão, and Nicholas D. Lane. 2020. Flower: A Friendly Federated Learning Research Framework. arXiv:2007.14390 [cs.LG] on PithPith's reviewarXiv
- [4] Weize Chen, Ziming You, Ran Li, Yitong Guan, Chen Qian, Chenyang Zhao, Cheng Yang, Ruobing Xie, Zhiyuan Liu, and Maosong Sun. 2025. Internet of Agents: Weaving a Web of Heterogeneous Agents for Collaborative Intelligence. InThe Thirteenth International Conference on Learning R... on Pith
- [5] Marco De Rossi, Davide Crapis, Jordan Ellis, and Erik Reppel. 2025. ERC-8004: Trustless Agents. Ethereum Improvement Proposals. Draft. Accessed October 27, 2026. https://eips.ethereum.org/EIPS/eip-8004 on Pith
- [6] Nanqing Dong, Zhipeng Wang, Jiahao Sun, Michael C. Kampffmeyer, William J. Knottenbelt, and Eric P. Xing. 2024. Defending Against Poisoning Attacks in Federated Learning With Blockchain.IEEE Transactions on Artificial Intelligence 5, 7 (2024), 3743–3756. doi:10.1109/TAI.2024.3... on Pithdoi:10.1109/tai.2024.3376651
- [7] Chaoyang He, Songze Li, Jinhyun So, Xiao Zeng, Mi Zhang, Hongyi Wang, Xi- aoyang Wang, Praneeth Vepakomma, Abhishek Singh, Hang Qiu, Xinghua Zhu, Jianzong Wang, Li Shen, Peilin Zhao, Yan Kang, Yang Liu, Ramesh Raskar, Qiang Yang, Murali Annavaram, and Salman Avestimehr. 2020. ... on Pith
- [8] Peter Kairouz, H. Brendan McMahan, Brendan Avent, Aurélien Bellet, Mehdi Bennis, Arjun Nitin Bhagoji, Keith Bonawitz, Zachary Charles, Graham Cormode, Rachel Cummings, et al . 2021. Advances and Open Problems in Federated Learning.Foundations and Trends in Machine Learning14, ... on Pithdoi:10.1561/2200000083
- [9] Peter Jihoon Kim, Kevin Britz, and David Knott. 2020. ERC-3009: Transfer With Authorization. Ethereum Improvement Proposals. Accessed: 2026-11-05. https://eips.ethereum.org/EIPS/eip-3009 on Pith
- [10] Alex Krizhevsky. 2009.Learning Multiple Layers of Features from Tiny Images. Technical Report. University of Toronto on Pith
- [11] Yang Liu, Tao Fan, Tianjian Chen, Qian Xu, and Qiang Yang. 2021. FATE: An Industrial Grade Platform for Collaborative Learning With Data Protection. Journal of Machine Learning Research22, 226 (2021), 1–6 on Pith
- [12] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Agüera y Arcas. 2017. Communication-Efficient Learning of Deep Networks from Decentralized Data. InProceedings of the 20th International Conference on Artificial Intelligence and Statistics (Proceedings of M... on Pith
- [13] OpenAI. 2025. Introducing Codex. OpenAI. Published May 16, 2025. https: //openai.com/index/introducing-codex/ on Pith
- [14] Erik Reppel, Nemil Dalal, and Dan Kim. 2025. Introducing x402: A New Standard for Internet-Native Payments. Coinbase Developer Platform. Published May 6, on Pith
- [15] Rui Sun, Zhipeng Wang, Jiahao Sun, and Rajiv Ranjan. 2025. Vision: How to Fully Unleash the Productivity of Agentic AI? Decentralized Agent Swarm Network. InICML 2025 Workshop on Collaborative and Federated Agentic Workflows on Pith
- [16] Rao Surapaneni, Miku Jha, Michael Vakoc, and Todd Segal. 2025. Announcing the Agent2Agent Protocol (A2A). Google Developers Blog. Published April 9, 2025. https://developers.googleblog.com/a2a-a-new-era-of-agent-interoperability/ on Pith
- [17] Stefanie Warnat-Herresthal, Hartmut Schultze, Krishnaprasad Lingadahalli Shas- try, Sathyanarayanan Manamohan, Saikat Mukherjee, Vishesh Garg, Ranjan Sarveswara, Kathrin Händler, Peter Pickkers, N. Al Aziz, Sofia Ktena, Florian Tran, Sebastian Briesemeister, et al . 2021. Swar... on Pithdoi:10.1038/s41586-021-03583-3
- [18] Yingxuan Yang, Mulei Ma, Yuxuan Huang, Huacan Chai, Chenyu Gong, Haoran Geng, Yuanjian Zhou, Ying Wen, Meng Fang, Muhao Chen, Shangding Gu, Ming Jin, Costas Spanos, Yang Yang, Pieter Abbeel, Dawn Song, Weinan Zhang, and Jun Wang. 2025. Agentic Web: Weaving the Next Web with AI... on PithPith's reviewarXiv
- [2025] https://www.coinbase.com/developer-platform/discover/launches/x402 on Pith
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment