Pith. sign in

REVIEW 4 major objections 4 minor 34 references

FedJigsaw: Multi-Agent Collaborative Model Reassembly for Decentralized Heterogeneous Federated Learning

T0 review · 4 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read FedJigsaw reframes heterogeneous federated learning as decentralized module assembly, letting each client build a personalized model from neighbors' building blocks.

desk verdict The design is coherent and the integration is new, but the headline empirical claims rest on baseline numbers that are not reproducible from the paper. read the letter →

arxiv 2608.01861 v1 pith:TXQV5BNL submitted 2026-08-03 cs.DC

Jifeng Chen , Haibo Zhang , Yawen Chen This is my paper
classification cs.DC
keywords federatedlearningmodelheterogeneityreassemblymulti-agentreinforcementneuralarchitecturesearchpartialtrainingdecentralizedattentionpointernetwork
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

FedJigsaw claims that model-heterogeneous federated learning is better solved by letting each client assemble its own architecture from reusable modules borrowed from neighbor clients, rather than by slicing a shared supernet into fixed subnetworks. The paper builds this as a multi-agent reinforcement learning problem: an attention-based pointer network selects modules step by step, a decentralized exchange protocol propagates both parameters and structural metadata, and an optional centralized critic stabilizes training without collecting private data. If the claim holds, federated systems can support fully diverse client architectures while cutting memory, latency, and cross-client accuracy variance relative to prior partial-training and NAS-based methods.

What carries the argument

AttenAssemble is an attention-based pointer network actor that turns model construction into a sequential decision problem: at each step it scores candidate modules by query-key attention between the current architecture state and the modules' metadata (source, layer position, type, input/output dimensions), with a STOP action for early termination. SymbioArchitect is the exchange protocol: clients broadcast lightweight metadata, retrieve physical module weights from neighbors when selected, and send updated parameters back to the source for aggregation. CoRe-Tune adds an attention-enhanced central critic that, during training only, uses a task-agnostic auxiliary dataset to estimate a global

What would settle it

Run FedJigsaw with the auxiliary dataset deliberately replaced by one whose feature statistics contradict the local data distributions (e.g., a different domain), and check whether its accuracy and variance advantages over partial-training baselines collapse; the paper's own ablation shows gains shrink when CoRe-Tune is disabled, so this directly targets the load-bearing proxy-data assumption.

Watch

Extended reading notes

Core claim

The paper's central claim is that the 'supernet trap'—requiring all clients to derive their models from one over-parameterized global backbone—is avoidable. Instead, each client maintains a small local base model as a passive repository and, each round, sequentially assembles a fresh architecture by choosing compatible modules from the repositories of its topological neighbors, guided by an attention-based policy. The authors report that this dynamic modular reassembly outperforms state-of-the-art PT and NAS baselines by up to 13.8% relative accuracy, with tighter cross-client variance and an order-of-magnitude reduction in decision latency, and that the gains persist even when the optional

Load-bearing premise

The collaboration signal depends on data-feature similarities computed from a task-agnostic auxiliary dataset that is assumed available and representative of every client's true local data; without such a proxy, the exploration bonus and the CoRe-Tune critic have no reliable ground truth to coordinate on.

Editorial extensions

If this is right

  • MHFL no longer requires a unified supernet or predefined capability tiers; clients can assemble architecturally distinct models while still benefiting from shared knowledge.
  • Reported results: average absolute accuracy gain of 3.94% (13.87% relative) over baselines; variance drops to 0.0431 on FEMNIST High Non-IID versus 0.1877 for FedShapleX.
  • Decision latency falls to 1.11–1.53 seconds per round, roughly an order of magnitude faster than Shapley-based methods and avoiding OOM failures seen in pFedHR.
  • Because CoRe-Tune is optional and the server decouples after training, the system can run in a fully decentralized mode without the proxy dataset, albeit with reduced gains.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial: the attention policy's learned 'semantic compatibility' could be probed by testing whether it transfers to a brand-new client or to unseen module types; if it does not, the framework's adaptivity would depend on retraining the actor per deployment.
  • Editorial: the manual module specification bottleneck suggests a natural extension: automating module discovery via graph partitioning, which would let FedJigsaw handle arbitrary backbone families without human-designed blocks.
  • Editorial: the reactive exchange limitation implies the next step is proactive knowledge routing; if a server-level scheduler could direct module propagation while execution stays local, the system might converge to a single consensus model in addition to personalized ones.
  • Editorial: the Byzantine risk acknowledged in the paper is testable: injecting a poisoned module into a neighborhood should measurably degrade downstream clients' accuracy, motivating module-level anomaly detection.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes FedJigsaw, a decentralized model-heterogeneous federated learning (MHFL) framework in which each client is modeled as a reinforcement learning agent that sequentially assembles a personalized neural network by selecting modules from neighboring clients. AttenAssemble uses an attention-based pointer network to handle variable action spaces; SymbioArchitect enables metadata-based module sharing with source-based aggregation; CoRe-Tune adds an optional centralized critic under a CTDE scheme to stabilize multi-agent training. Experiments on five image datasets report up to 13.8% relative accuracy improvements over PT and NAS baselines, reduced cross-client variance, and lower decision latency/memory. Ablations attempt to separate the contributions of the sharing and tuning mechanisms.

Significance. If the reported results are reproducible, FedJigsaw would be a meaningful contribution to MHFL. The paper reframes partial training as a bottom-up, decentralized module-assembly problem, which is conceptually cleaner than top-down supernet slicing and avoids the memory cost of supernet maintenance. The use of an attention-pointer actor for variable action spaces is well motivated, and the CTDE critic is a reasonable way to address non-stationarity. The paper also deserves credit for reporting per-client variance and including ablations that isolate the sharing and tuning components. However, the empirical evidence is currently not trustworthy: the reward used to train the policy appears to be computed on the test accuracy, the baseline numbers in Table 1 are far below typical published results and are not reproducible from the text, and the convergence / Nash-avoidance claims are not supported by formal analysis. The significance is therefore conditional on a major revision and independent verification.

major comments (4)
  1. [Table 1 / §4.1] The central empirical claim rests on baseline numbers that are not reproducible from the paper. HeteroFL is reported at 49.40%/54.05% on CIFAR-10 and FedShapleX at 27.09%/50.56% on CIFAR-100, which are far below published results for these methods. No code, seed list, per-client architecture assignment, training budget, or baseline hyperparameter grid is provided. The 'fair comparison' statement in §4.1 does not establish that each baseline was tuned to its intended configuration. As Table 1 is the only evidence for the headline 13.8% relative improvement, the comparison is currently indistinguishable from undertuned baselines. The authors should release code and full baseline configurations, and rerun baselines with multiple seeds.
  2. [App. A.2, Algorithm 2, lines 7–10; Eq. (3)] The reward in Eq. (3) uses Acc^(i), and Algorithm 2 obtains this quantity as 'test accuracy' immediately before computing R. If this is literally the test accuracy, then the policy is being trained on the test set, and the accuracies in Table 1 are not held-out results but fit values. This invalidates the headline comparisons. The authors must use a validation split (or a separate held-out test set) for reward computation and report only the final held-out accuracy, ensuring no overlap between the reward signal and the reported metric.
  3. [§3.2.2 and App. A.4] The claim that CoRe-Tune guides the system away from a suboptimal Nash equilibrium is not supported by any formal analysis. App. A.4, cited for convergence analysis, provides only training curves (Figures 3 and 4); there is no proof of convergence, no formal statement of the game or equilibrium, and no comparison against a Nash baseline. Since C3 is presented as a core challenge and motivates the CTDE design, this is a missing load-bearing component. At minimum, the paper should clearly state that the convergence and Nash-avoidance claims are empirical, or provide a formal result.
  4. [Eq. (4), §3.3] The Exp^(i) exploration bonus requires a task-agnostic auxiliary dataset to compute the data-feature similarity cos(E_data^(i), E_data^(j)). The paper acknowledges in §3.3 that access to such a proxy dataset is rarely guaranteed and labels CoRe-Tune optional, yet all reported headline results are for the full system with CoRe-Tune enabled. The ablation W/O CoRe-Tune is reassuring, but it does not quantify how the quality of the proxy dataset (size, domain shift, representativeness) affects the gain. Please add a sensitivity analysis on proxy dataset mismatch, or state the conditions under which the claimed gains persist.
minor comments (4)
  1. [Eq. (1)] The temperature τ appears inside the denominator as a scaling factor; clarify whether τ is a hyperparameter or a learned parameter, and how it interacts with the softmax normalization.
  2. [§4.1 vs §A.4] The same symbol α is used for the Dirichlet distribution parameter and for the reward weight; this is confusing in the experimental discussion.
  3. [Figures 3 and 4 vs §A.4] The captions state 500 epochs while the text says 1000 policy training rounds and 1000 model training rounds; align these numbers.
  4. [A.5] In the first limitation paragraph, the sentence 'This manual uses automated graph-partitioning algorithms...' appears to be missing a verb; likely 'This manual specification can be replaced by...'.

Circularity Check

1 steps flagged · score 6.0 of 10

Reported test accuracy is the same quantity optimized in the reward (Algorithm 2 line 8 feeds test accuracy into line 10 reward), so the headline accuracy comparison is a fitted objective, not a held-out prediction.

  1. fitted input called prediction [Algorithm 2 (Appendix A.2), lines 7-10; Eq. (3), Section 3.2.2; Table 1]
    "Algorithm 2 (Appendix A.2): '8 Evaluate model to obtain test accuracy Acc_{i,k}; ... 10 Compute final reward R_{i,k} ← αAcc_{i,k} + (1−α)Exp_i + βCap_i;' Eq. (3): 'R^{(i)}_{k,T} = α·Acc^{(i)} + (1−α)·Exp^{(i)} + β·Cap^{(i)}' and Table 1 reports 'Test accuracy (%)'."

    The assembly policy is trained by REINFORCE to maximize the reward in Eq. (3). Algorithm 2 obtains Acc_{i,k} by evaluating the assembled model on the test set (line 8), then feeds exactly that test accuracy into the reward (line 10). The headline results in Table 1 are the same 'Test accuracy (%)' metric. Thus the reported accuracy is the optimized objective, not an independent prediction: the policy is fitted to the test metric and the test metric is then reported as the result. The prose around Eq. (3) calls Acc 'validation accuracy', but Algorithm 2 explicitly uses 'test accuracy'; under the algorithm's own statement, the comparison reduces by construction to the quantity being optimized. The W/O CoRe-Tune ablation does not remove this reward structure, so the central empirical claim is

full rationale

The central load-bearing claim is the accuracy comparison in Table 1. The derivation chain for that claim is: AttenAssemble policy selects modules, the assembled model is trained, then evaluated, and the evaluation accuracy is used in the terminal reward (Eq. 3) that trains the policy. Algorithm 2 specifies this evaluation as 'test accuracy' (line 8) and immediately uses it in the reward (line 10). Therefore the reported test accuracy is inside the training loop; the policy is optimized against the same numbers that are later presented as the outcome. This is a concrete fitted-input-called-prediction circularity rather than a mere 'optimization is standard' situation. The paper's Eq. (3) labels Acc as validation accuracy, creating an internal inconsistency; taking the algorithm literally, the evaluation is not held out. No other circularity pattern is present: the FedShapleX self-citation (Chen et al., 2025) is used only as a baseline and as a Dirichlet-partition citation, not as a load-bearing uniqueness theorem or ansatz source. The proxy-dataset limitation of Exp(i) (Section 3.3) is an assumption about deployment, not a circular reduction. The score is 6 rather than higher because the paper does include ablations that are compared against baselines, and because not every reported number is solely a consequence of the reward; however, the headline full-system accuracy advantage reduces in part to optimizing the evaluation metric itself.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new physical or mathematical entity: modules, metadata tuples, the pointer-network actor, and the server critic are all engineering components built from prior work. The load-bearing free parameters are the reward weights (alpha, beta), temperature, and discount factor, which are either hand-set or unspecified. The most consequential assumptions are the availability of a proxy dataset for the collaboration signal and the absence of malicious neighbors.

free parameters (5)
  • reward weight alpha = 0.7 (Low Non-IID), 0.9 (High Non-IID)
    Hand-tuned balance between accuracy and exploration bonus; discussed only in the implementation details.
  • reward weight beta = not specified
    Scales the hardware capacity bonus Cap(i); no value given anywhere in the text.
  • exploration temperature tau = not specified
    Softmax temperature in Eq. 1 controls exploration/exploitation; no value or schedule given.
  • discount factor gamma = not specified
    Used in Monte-Carlo return r = gamma^(T-t) R; no value given.
  • local training epochs E and learning rate = learning rate 0.1, epochs not specified
    Adam at LR 0.1 is unusually high and its interaction with RL stability is not analyzed; local epochs are not reported.
assumptions (3)
  • domain assumption Neighboring clients are honest and share well-formed modules.
    The whole SymbioArchitect mechanism assumes modules fetched from neighbors are legitimate and not poisoned; the broader impact section admits Byzantine poisoning is unhandled.
  • domain assumption A task-agnostic proxy/auxiliary dataset is available to clients for computing Exp(i) and training the critic.
    Section 3.3, Eq. 4, and Algorithm 2 use E_data and Exp(i) computed on an auxiliary dataset; the paper concedes such a dataset is 'rarely guaranteed', so the headline results depend on an assumption that is acknowledged as often false.
  • domain assumption REINFORCE with a learned baseline converges to a useful policy in this sparse-reward, non-stationary setting.
    Actor-critic policy gradient (Eq. 5) is a standard algorithm, but no convergence guarantee is provided; the paper only points to empirical loss plots in Appendix A.4.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FedJigsaw: Multi-Agent Collaborative Model Reassembly for Decentralized Heterogeneous Federated Learning." pith.science (2026). https://pith.science/paper/TXQV5BNL

@misc{pith2026260801861,
  author       = {Pith},
  title        = {Pith review of: FedJigsaw: Multi-Agent Collaborative Model Reassembly for Decentralized Heterogeneous Federated Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TXQV5BNL}},
  note         = {Machine review of arXiv:2608.01861}
}
read the original abstract

Model Heterogeneous Federated Learning (MHFL) addresses client-level resource heterogeneity by allowing each participant to train a personalized model architecture under a shared training objective. A prevalent paradigm, Partial Training (PT), achieves this by allowing each client to train a subnetwork of the global model. However, existing PT methods typically rely on predefined architectural templates or over-parameterized supernets, limiting fine-grained personalization and imposing substantial computational and memory overhead. We propose FedJigsaw, a novel framework that reshapes model personalization as a dynamic and decentralized model assembly problem. Instead of selecting subnetworks from a predefined supernetwork, each client constructs its model by assembling reusable modules learned from neighboring clients. At the client level, we introduce AttenAssemble to enable each participant to adaptively construct a tailored model based on local observations. To support efficient knowledge sharing under communication and privacy constraints, we design SymbioArchitect, a mechanism that allows clients to exchange granular model modules with their topological neighbors. To mitigate training instability introduced by decentralized module exchange, we design CoRe-Tune, an attention-enhanced centralized training with a decentralized execution strategy, which guides local policies to foster implicit collaboration and stabilize training dynamics, without compromising data privacy. Extensive evaluations demonstrate that FedJigsaw outperforms state-of-the-art MHFL baselines by up to 13.8% in relative accuracy while significantly shrinking cross-client performance variance, but also slashes decision-making latency and peak memory footprint compared to existing policy-driven methods.

Figures

Figures reproduced from arXiv: 2608.01861 by the authors.

Figure 1
Figure 1. The architecture of the AttenAssemble policy network. The sequence will then be contextualized by a multi-head Transformer encoder. To seamlessly handle the variable action space, a Pointer Network head is employed to compute the action distribution via dot-product attention score, which directly formulates the policy π(a | s) in our reinforcement learning framework. Specifically, it utilizes the hidden representati… view at source ↗
Figure 2
Figure 2. The architecture of the FedJigsaw MHFL system. The left side illustrates the SymbioArchitect, while the right side shows the Centralized Server used in the CTDE scheme. 3.2 SYMBIOARCHITECT: FEDERATED MODULE PROPAGATION AND SHARING MECHANISM Building upon the model assembly mechanism introduced in Section 3.1, this section broadens the scope to system-wide collaboration. Specifically, we will discuss how clients dyna… view at source ↗
Figure 3
Figure 3. Learning dynamics of FedJigsaw across five datasets over 500 communication rounds. The top row presents the convergence of test accuracy, while the bottom row illustrates the cross-client performance variance under both Low Non-IID and High Non-IID settings. 14 [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Training dynamics of the Actor (top) and Critic (bottom) losses over 500 epochs. [PITH_FULL_IMAGE:figures/full_fig_p015_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

34 extracted references · 2 linked inside Pith

  1. [1]

    Federated Learning in Mobile Edge Networks: A Comprehensive Survey , year=

    Lim, Wei Yang Bryan and Luong, Nguyen Cong and Hoang, Dinh Thai and Jiao, Yutao and Liang, Ying-Chang and Yang, Qiang and Niyato, Dusit and Miao, Chunyan , journal=. Federated Learning in Mobile Edge Networks: A Comprehensive Survey , year=

  2. [2]

    Artificial intelligence and statistics , pages=

    Communication-efficient learning of deep networks from decentralized data , author=. Artificial intelligence and statistics , pages=. 2017 , organization=

    2017

  3. [3]

    The Eleventh International Conference on Learning Representations , year=

    Personalized Federated Learning with Feature Alignment and Classifier Collaboration , author=. The Eleventh International Conference on Learning Representations , year=

  4. [4]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Model-contrastive federated learning , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  5. [5]

    Advances in neural information processing systems , volume=

    Federated learning from pre-trained models: A contrastive learning approach , author=. Advances in neural information processing systems , volume=

  6. [6]

    International Conference on Machine Learning , pages=

    Federated learning with partial model personalization , author=. International Conference on Machine Learning , pages=. 2022 , organization=

    2022

  7. [7]

    Advances in neural information processing systems , volume=

    An efficient framework for clustered federated learning , author=. Advances in neural information processing systems , volume=

  8. [8]

    33rd International Joint Conference on Artificial Intelligence, IJCAI 2024 , pages=

    Redefining Contributions: Shapley-Driven Federated Learning , author=. 33rd International Joint Conference on Artificial Intelligence, IJCAI 2024 , pages=. 2024 , organization=

    2024

Show all 34 references
  1. [9]

    Proceedings of the 2023 SIAM International Conference on Data Mining (SDM) , pages=

    Knowledge-enhanced semi-supervised federated learning for aggregating heterogeneous lightweight clients in iot , author=. Proceedings of the 2023 SIAM International Conference on Data Mining (SDM) , pages=. 2023 , organization=

    2023

  2. [10]

    International conference on machine learning , pages=

    Data-free knowledge distillation for heterogeneous federated learning , author=. International conference on machine learning , pages=. 2021 , organization=

    2021

  3. [11]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Fedproto: Federated prototype learning across heterogeneous clients , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  4. [12]

    Advances in Neural Information Processing Systems , volume=

    Federated model heterogeneous matryoshka representation learning , author=. Advances in Neural Information Processing Systems , volume=

  5. [13]

    Advances in neural information processing systems , volume=

    Ensemble distillation for robust model fusion in federated learning , author=. Advances in neural information processing systems , volume=

  6. [14]

    arXiv preprint arXiv:1910.03581 , year=

    Fedmd: Heterogenous federated learning via model distillation , author=. arXiv preprint arXiv:1910.03581 , year=

    1910 arXiv

  7. [15]

    Advances in Neural Information Processing Systems , volume=

    Federated fine-tuning of large language models under heterogeneous tasks and client resources , author=. Advances in Neural Information Processing Systems , volume=

  8. [16]

    Advances in Neural Information Processing Systems , volume=

    Flora: Federated fine-tuning large language models with heterogeneous low-rank adaptations , author=. Advances in Neural Information Processing Systems , volume=

  9. [17]

    IEEE transactions on knowledge and data engineering , volume=

    Collaborative learning with heterogeneous local models: A rule-based knowledge fusion approach , author=. IEEE transactions on knowledge and data engineering , volume=. 2023 , publisher=

    2023

  10. [18]

    Pacific-Asia Conference on Knowledge Discovery and Data Mining , pages=

    Rule-based collaborative learning with heterogeneous local learning models , author=. Pacific-Asia Conference on Knowledge Discovery and Data Mining , pages=. 2022 , organization=

    2022

  11. [19]

    IEEE Internet of Things Journal , volume=

    Decentralized federated learning via mutual knowledge transfer , author=. IEEE Internet of Things Journal , volume=. 2021 , publisher=

    2021

  12. [20]

    Advances in neural information processing systems , volume=

    Fedrolex: Model-heterogeneous federated learning with rolling sub-model extraction , author=. Advances in neural information processing systems , volume=

  13. [21]

    Advances in Neural Information Processing Systems , volume=

    Fjord: Fair and accurate federated learning under heterogeneous targets with ordered dropout , author=. Advances in Neural Information Processing Systems , volume=

  14. [22]

    9th International Conference on Learning Representations, ICLR 2021 , year=

    HETEROFL: COMPUTATION AND COMMUNICATION EFFICIENT FEDERATED LEARNING FOR HETEROGENEOUS CLIENTS , author=. 9th International Conference on Learning Representations, ICLR 2021 , year=

    2021

  15. [23]

    IEEE wireless communications letters , volume=

    Federated dropout—A simple approach for enabling federated learning on resource constrained devices , author=. IEEE wireless communications letters , volume=. 2022 , publisher=

    2022

  16. [24]

    IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems , volume=

    Flexfl: Heterogeneous federated learning via apoz-guided flexible pruning in uncertain scenarios , author=. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems , volume=. 2024 , publisher=

    2024

  17. [25]

    FedShapleX: Shapley Value Driven Context-Aware Model-Heterogeneous Federated Learning , year=

    Chen, Jifeng and Zhang, Haibo and Barnard, Amanda , booktitle=. FedShapleX: Shapley Value Driven Context-Aware Model-Heterogeneous Federated Learning , year=

  18. [26]

    arXiv preprint arXiv:2004.08546 , year=

    Towards non-IID and invisible data with FedNAS: Federated deep learning via neural architecture search , author=. arXiv preprint arXiv:2004.08546 , year=

    2004 arXiv

  19. [27]

    Advances in Neural Information Processing Systems , volume=

    Towards personalized federated learning via heterogeneous model reassembly , author=. Advances in Neural Information Processing Systems , volume=

  20. [28]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Distreal: Distributed resource-aware learning in heterogeneous systems , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  21. [29]

    IEEE Transactions on Mobile Computing , volume=

    Peaches: Personalized federated learning with neural architecture search in edge computing , author=. IEEE Transactions on Mobile Computing , volume=. 2024 , publisher=

    2024

  22. [30]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    PerFedRLNAS: One-for-all personalized federated neural architecture search , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  23. [31]

    International conference on machine learning , pages=

    Efficient neural architecture search via parameters sharing , author=. International conference on machine learning , pages=. 2018 , organization=

    2018

  24. [32]

    Journal of Artificial Intelligence Research , volume=

    On centralized critics in multi-agent reinforcement learning , author=. Journal of Artificial Intelligence Research , volume=

  25. [33]

    Advances in neural information processing systems , volume=

    Deep model reassembly , author=. Advances in neural information processing systems , volume=

  26. [34]

    Advances in neural information processing systems , volume=

    Pointer networks , author=. Advances in neural information processing systems , volume=

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.