REVIEW 4 major objections 6 minor 17 references
QoS-Aware Token Scheduling and Private Data Valuation for Multi-Modal Agentic Networks
T0 review · 4 major / 6 minor · reviewed 2026-07-13 · grok-4.5
Pith's one-line read DP-protected prototypes let edge agentic systems value private multi-modal data and allocate scarce AI quotas by contribution.
desk verdict Solid systems integration of DP prototypes and token-bucket AI quotas; fairness claims are real relative to the authors' proxy, not to external utility. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The data-anchoring-token (DAT) valuation f(d_n)=ϕ(d_n) ln(1+|d_n|), where ϕ is a novelty score from DP-protected prototypes, which sets each agent's refill rate in the token-bucket rule of Equation (15).
What would settle it
Replace the novelty rule with a more accurate utility estimate (or a true Shapley-style valuation on the same multi-modal set) and check whether Jain and Gini fairness and reconstruction resistance still beat the random, round-robin, and max-min baselines under the same scarcity and heterogeneity conditions.
Extended reading notes
Core claim
When multi-modal contributions are represented as differentially private prototypes and valued by a novelty-quantity score, the resulting values can parameterize a token-bucket allocator so that scarce AI quotas are issued proportionally to effective contribution, remaining fair under heterogeneous data and limited resources while reducing semantic leakage under reconstruction attacks.
Load-bearing premise
The simple novelty rule that gives full credit only to the first batch of samples in a new cluster is a good enough proxy for true data utility that the resulting values can safely drive quota allocation.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes a two-stage framework for decentralized multi-modal agentic systems: (i) differentially private (DP) prototypes of multi-modal embeddings, released as iDAO catalog entries for private data valuation, and (ii) a token-bucket scheduler that converts those valuations into AI-quota refill rates. Data value is defined as f(d_n)=φ(d_n) ln(1+|d_n|), where φ is a novelty score that awards 1 to the first Γ samples of a newly discovered cluster, 0.5 to later in-cluster samples, and 0 to unassigned points. Given valuations v, the closed-form allocation in Eq. (15) fills per-agent buckets proportionally to v while respecting residual capacity. Simulations on COCO embeddings under Dirichlet heterogeneity report higher Jain index and lower Gini coefficient than Random/Round-Robin/Max-Min baselines, and lower reconstruction cosine similarity than an image-only DP baseline under a direct-ascent attack.
Significance. If the valuation proxy is accepted as a reasonable stand-in for contribution quality, the work offers a concrete systems primitive that couples formal (ε,δ)-DP prototype release with contribution-proportional AI quota scheduling—an attractive design point for resource-scarce edge agentic networks that want data sovereignty without full raw-data sharing. Strengths include a clear two-stage reformulation of a non-convex joint objective, an explicit Gaussian-mechanism derivation for prototype DP (supplementary materials), extensive Monte Carlo fairness sweeps (1000 rounds) over α and N, and an empirical reconstruction-attack comparison that goes beyond pure theory. The main significance is therefore architectural and empirical within the authors’ proxy regime, not a new optimality theorem for data markets.
major comments (4)
- [Section IV-B, Eq. (20)] Section IV-B and Eq. (20): The central claim that the scheduler “rewards effective contributions” rests on φ(d_n) being a faithful proxy for true data utility. The rule (first Γ samples of a new cluster get φ=1, later samples 0.5, unassigned 0) is stipulated without any experiment linking φ or the resulting v_n to an external utility measure (downstream task accuracy, held-out validation loss, model improvement, or even embedding-space coverage relative to a ground-truth ontology). Without such a link, “effective contribution” is defined by the same function that parameterizes the objective, so the incentive interpretation is not yet substantiated.
- [Section II-D, Figs. 4–5] Section II-D–E and Figs. 4–5: Fairness is evaluated on reward rates r_n = u_n / v_n (Eq. 5), where v_n is produced by the authors’ own f. Jain/Gini gains therefore show that the token-bucket rule is consistent with its input valuations under heterogeneous v, not that it fairly rewards true data quality. If φ systematically mis-ranks multi-modal or noisy contributions, the reported “contribution-based fairness” can improve while true utility allocation worsens. At minimum the paper should either (a) validate φ against an external utility oracle, or (b) reframe the claim as “valuation-consistent fairness under a stated proxy.”
- [Abstract; Section V] Abstract and Section V: The abstract and conclusions claim improved “QoS” versus benchmarks, but the reported experiments measure Jain index, Gini coefficient, prototype discovery counts, and reconstruction cosine similarity. Social welfare F(v,u)=∑ v_n u_n (Eq. 10) is the stated objective, yet no table or figure reports F (or latency, task success rate, or other service-level QoS) under the proposed vs. baseline allocators. Either add a QoS metric that is independent of the valuation proxy, or narrow the claim to fairness and privacy under the simulated regime.
- [Section II-E; Eq. (15)] Section II-E and III: Problem (11) is acknowledged non-convex; the paper solves it by a two-stage heuristic (optimize u given v via Eq. 15, then set v_n = f(d_n) when u_n>0). There is no analysis of sub-optimality relative to the joint problem, nor of fixed-point consistency when v and u co-evolve over multiple contribution rounds. For a systems paper this is acceptable if limitations are explicit, but the current text presents Eq. (15) as “the optimized token allocation” without quantifying how far the two-stage solution can be from a joint stationary point under the fairness constraints (7)–(9).
minor comments (6)
- [Section II-A, II-C] Notation: both the number of edge servers and the DP mechanism are denoted M (Section II-A vs. II-C); rename one to avoid collision.
- [Figs. 4–5] Fig. 4 caption and body use “Gini Coefficent” (missing ‘i’); same typo appears in Fig. 5.
- [Section IV-B; Section V-A] Eq. (21) and the experimental text mix ε and ϵ for the privacy budget; standardize on one symbol.
- [Section V-A] The Image-DP baseline is cited as [11] (federated domain adaptation prototypes) while the reconstruction attack is [12]; a short sentence clarifying that Image-DP applies Gaussian noise only to image embeddings (leaving text unchanged) would help readers who do not open the supplement.
- [Introduction; Section II-B] iDAO and DAT are introduced via footnotes/URLs; a one-sentence formal definition in the main text (what on-chain fields a DAT carries; what governance an iDAO performs) would improve self-containment.
- [Section V; Supplementary 4.B] Sensitivity of results to the CLIP encoder (ViT-B/32 only) is discussed only in the supplement limitations; a brief main-text note that richer encoders may worsen reconstruction risk under fixed (ε,δ) would be useful.
Circularity Check
No significant circularity: token-bucket allocation and novelty valuation are author-defined mechanisms whose fairness/attack outcomes are external simulation results against independent benchmarks, not algebraic identities of the inputs.
full rationale
The paper proposes a two-stage design (token-bucket refill rates parameterized by v_n, then a novelty-based f(d_n)=ϕ ln(1+|d_n|) with an ad-hoc Γ threshold for ϕ) and evaluates it via Monte-Carlo simulations of Jain/Gini on reward rates r_n=u_n/v_n plus reconstruction-attack cosine similarities. These numerical outcomes are not forced by construction: unconstrained proportional allocation does make r constant (hence J=1, G=0), yet the reported gains appear under binding bucket constraints, heterogeneous Dirichlet contributions, and scarcity (Figs. 4–5), and are measured against external schedulers (random, round-robin, max-min) that ignore v entirely. The DP prototypes follow the standard Gaussian mechanism; attack resistance is an independent empirical comparison to Image-DP. Self-citations (e.g., [2]) supply background on prior blockchain valuation and are not load-bearing uniqueness theorems or fitted parameters renamed as predictions. The unvalidated status of ϕ as a proxy for true utility is a correctness/assumption issue, not a circular derivation. Hence the derivation chain is self-contained against external benchmarks and circularity is negligible.
Assumptions & free parameters
free parameters (6)
- Γ (novelty threshold) =
50
- ε (privacy budget) =
1
- δ =
1e-5
- α (Dirichlet concentration) =
0.5
- B (bucket size) =
4
- q,s (noise scale parameters) =
q=0.2,s=0.05 (baseline)
assumptions (5)
- standard math Gaussian mechanism yields (ε,δ)-DP when noise variance satisfies the ℓ2-sensitivity bound of the mean embedding (Eq. 21 / S.3).
- domain assumption Token-bucket scheduling with refill rate proportional to data value v_n produces a feasible, incentive-compatible allocation of scarce AI quotas.
- ad hoc to paper Novelty ϕ equals 1 for the first Γ samples of a newly discovered cluster, 0.5 thereafter, and 0 for unassigned points; this approximates true data utility.
- domain assumption Rational agents will contribute data to maximize f(d_n) and will actively spend tokens rather than hoard them.
- domain assumption Embeddings are ℓ2-clipped by constant C so that mean sensitivity is at most 2C/I.
invented entities (3)
-
iDAO (individual-centric decentralized autonomous organization)
-
DAT (data anchoring token)
independent evidence
-
DP-protected prototype catalog as semantic market primitive
Cite this review
Pith. "Pith review of QoS-Aware Token Scheduling and Private Data Valuation for Multi-Modal Agentic Networks." pith.science (2026). https://pith.science/paper/D7Z3ONIL
@misc{pith2026260615573,
author = {Pith},
title = {Pith review of: QoS-Aware Token Scheduling and Private Data Valuation for Multi-Modal Agentic Networks},
year = {2026},
howpublished = {\url{https://pith.science/paper/D7Z3ONIL}},
note = {Machine review of arXiv:2606.15573}
}
read the original abstract
In agentic systems, human-generated data records anchor the value of AI services. Yet cloud compute pipelines centralize processing on remote servers. Data centralization reduces personal data sovereignty and may potentially degrade the quality of service (QoS). Meanwhile, user contributions are diverse in quantity and quality: decentralized records can be biased, noisy, and heterogeneously distributed. To address the data challenge, we study fair token allocation and private data valuation for decentralized and resource-constrained agentic systems. Our approach embeds multi-modal representations in a shared semantic space and releases differentially private (DP) prototypes to preserve utility while reducing semantic leakage. With the DP guarantee, we design a fair token allocation scheme that rewards effective contributions and remains robust to data heterogeneity and AI resource scarcity. Extensive simulations demonstrate improved contribution-based fairness and QoS compared to standard benchmarks. The improved resistance to image reconstruction attacks indicates enhanced privacy for multi-modal personal data.
Reference graph
Works this paper leans on
-
[1]
Enabling mobile AI agent in 6G era: Architecture and key technologies,
Z. Chen, Q. Sun, N. Liet al., “Enabling mobile AI agent in 6G era: Architecture and key technologies,”IEEE Network, vol. 38, no. 5, pp. 66–75, 2024
2024
-
[2]
Towards collaborative edge intelligence: Blockchain-based data valuation and scheduling for improved quality of service,
Y . Du, Z. Wang, C. Leung, and V . C. Leung, “Towards collaborative edge intelligence: Blockchain-based data valuation and scheduling for improved quality of service,”Future Internet, vol. 16, no. 8, p. 267, 2024
2024
-
[3]
The shapley value in machine learning,
B. Rozemberczki, L. Watson, P. Bayeret al., “The shapley value in machine learning,” inInternational Joint Conference on Artificial Intelligence and European Conference on Artificial Intelligence, 2022, pp. 5572–5579
2022
-
[4]
Olive: Oblivious federated learning on trusted execution environment against the risk of sparsification,
F. Kato, Y . Cao, and M. Yoshikawa, “Olive: Oblivious federated learning on trusted execution environment against the risk of sparsification,” Proceedings of the VLDB Endowment, vol. 16, no. 10, pp. 2404–2417, 2023
2023
-
[5]
ZKML: An optimizing system for ML inference in zero-knowledge proofs,
B.-J. Chen, S. Waiwitlikhit, I. Stoicaet al., “ZKML: An optimizing system for ML inference in zero-knowledge proofs,” inEuropean Conference on Computer Systems, 2024, pp. 560–574
2024
-
[6]
The algorithmic foundations of differential privacy,
C. Dwork, A. Rothet al., “The algorithmic foundations of differential privacy,”Foundations and Trends in Theoretical Computer Science, vol. 9, no. 3–4, pp. 211–407, 2014
2014
-
[7]
Towards the fairness of traffic policer,
D. Shan, P. Zhang, W. Jianget al., “Towards the fairness of traffic policer,” inIEEE Conference on Computer Communications, 2021, pp. 1–10
2021
-
[8]
Microsoft COCO: Common objects in context,
T.-Y . Lin, M. Maire, S. Belongieet al., “Microsoft COCO: Common objects in context,” inEuropean Conference on Computer Vision. Springer, 2014, pp. 740–755
2014
Show all 17 references
-
[9]
Multimodal remote inference,
K. Zhang, Y . Sun, and B. Ji, “Multimodal remote inference,”arXiv preprint arXiv:2508.07555, 2025
2025 arXiv
-
[10]
Fairness in serving large language models,
Y . Sheng, S. Cao, D. Liet al., “Fairness in serving large language models,” inUSENIX Symposium on Operating Systems Design and Implementation, 2024, pp. 965–988
2024
-
[11]
Enhancing federated domain adaptation with multi-domain prototype-based federated fine-tuning,
J. Zhang, Y . Duan, S. Niuet al., “Enhancing federated domain adaptation with multi-domain prototype-based federated fine-tuning,” inInterna- tional Conference on Learning Representations, 2025, pp. 1–23
2025
-
[12]
Direct ascent synthesis: Revealing hid- den generative capabilities in discriminative models,
S. Fort and J. Whitaker, “Direct ascent synthesis: Revealing hid- den generative capabilities in discriminative models,”arXiv preprint arXiv:2502.07753, 2025
2025 arXiv
-
[13]
Learning transferable visual models from natural language supervision,
A. Radford, J. W. Kim, C. Hallacyet al., “Learning transferable visual models from natural language supervision,” inInternational Conference on Machine Learning, 2021, pp. 8748–8763. SUPPLEMENTARYMATERIALS
2021
-
[14]
, I}, letθ i denote thei-th embedding vector, clipped byC
DERIVATIONS OFDIFFERENTIALPRIVACYGUARANTEE ForIdata instances in the same cluster, indexed byi∈ I= {1,2, . . . , I}, letθ i denote thei-th embedding vector, clipped byC. That is, θi 2 ≤C,∀i∈ I. LetDandD ′ be neighboring datasets differing in exactly one example, with|D|=|D ′|=...
-
[15]
Across 200 rounds of reconstruction attacks [12], the benchmark Fig
ADDITIONALRECONSTRUCTIONATTACKRESULTS We further provide additional experimental results in Fig.S.1 to support the DP guarantee of our shared prototypes. Across 200 rounds of reconstruction attacks [12], the benchmark Fig. S.1. Reconstruction attack comparison: (top) Image-DP ...
-
[16]
S.2 to illustrate the severe data heterogeneity in decentralized settings
DATASETVISUALIZATION We visualize the COCO dataset 7 in Fig. S.2 to illustrate the severe data heterogeneity in decentralized settings. The COCO categories are dominated by the “person” category, as human- related activities and objects are the most common subjects in data sam...
-
[17]
Bucket sizing and dynamic adjustment The bucket sizeB n is determined by both the purchased service tier and the agent’s data contribution level
LIMITATIONS ANDCLARIFICATIONS A. Bucket sizing and dynamic adjustment The bucket sizeB n is determined by both the purchased service tier and the agent’s data contribution level. Agents with higher tiers or greater contributions receive larger buckets and hence larger AI quota...
Reviewed July 13, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.