Robust Multi-Agent Bandits with Heavy-Tailed Rewards and Information Asymmetry

Daphne Feng; Lily Jiang; Ricardo Parada; Sophia Yi; William Chang

REVIEW 2 major objections 4 minor 17 references

Robust Multi-Agent Bandits with Heavy-Tailed Rewards and Information Asymmetry

T0 review · 2 major / 4 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read Three decentralized multi-agent bandit algorithms match centralized heavy-tailed regret rates under information asymmetry.

desk verdict Solid multi-agent heavy-tailed bandit paper with a real but fixable anytime-consistency gap in Problem C. read the letter →

arxiv 2608.10529 v1 pith:UHYSBJ4Y submitted 2026-08-11 cs.LG cs.AI

Daphne Feng , Ricardo Parada , Lily Jiang , Sophia Yi , William Chang This is my paper · ORCID

classification cs.LGcs.AI

keywords multi-armedbanditsheavy-tailedrewardsmulti-agentlearninginformationasymmetrydecentralizedalgorithmsregretboundstruncated-meanestimatorcoordinationwithoutcommunication

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

The paper studies a multi-armed bandit problem in which several agents act simultaneously, agree on a strategy beforehand, and cannot communicate once learning begins, while reward distributions have heavy tails. Its central claim is that each of three information-asymmetry regimes admits a decentralized algorithm whose regret matches the best centralized heavy-tailed rate up to constants and, in one case, a single log factor. The three mechanisms are shared rewards with hidden actions, observable actions with independent rewards, and full asymmetry; the corresponding algorithms are named mRUCB-A, mRUCB-Intervals, and mHT-DSEE. Under the moment condition $E|X-\mu|^{1+\varepsilon}\le v$, the regret bounds are $O(\log T\sum_{a\neq a^*}\Delta_a^{-1/\varepsilon})$ for the first two regimes and $O(K^M\log^2 T)$ for the third. The result matters because real deployments such as spectrum sharing, federated experiments, and sensor teams without backhaul often combine heavy-tailed observations with limited attribution and no online communication; the paper argues that this combination need not degrade the asymptotic learning rate.

What carries the argument

The load-bearing object is the robust upper confidence bound built on the truncated-mean estimator $\hat\mu_a(t)$ with confidence radius $\alpha_a(t)=v^{1/(1+\varepsilon)}(c\log(T^\gamma)/n_a(t))^{\varepsilon/(1+\varepsilon)}$, for which the paper imports the concentration guarantee $\Pr(|\hat\mu_a(t)-\mu_a|>\alpha_a(t))\le t^{-\gamma}$. Everything else is designed so that this radius is common knowledge among the players: in Problem A the shared reward synchronizes all estimates; in Problem B the shared pull count $n_a(t)$ makes $\alpha_a(t)$ identical across players, so a single observed deviation can signal arm elimination through a 1-bit implicit channel (intervals that separate once $4\alpha_a(t)<\Delta_a$); in Problem C the deterministic anytime schedule $D(t)=\lceil w(t)\log t\rceil$ lets every player reproduce the exploration sequence from the round index alone. The same concentration bound is the only probabilistic input to all three theorems, so any estimator satisfying it can replace the truncated mean.

What would settle it

Run the truncated-mean estimator (3) on a Pareto distribution with $\varepsilon=0.5$, $v=1$, and the exact truncation threshold $(v s/\log(T^\gamma))^{1/(1+\varepsilon)}$, and estimate $\Pr(|\hat\mu_a(t)-\mu_a|>\alpha_a(t))$ over many independent runs; if the empirical tail probability exceeds $t^{-\gamma}$ at any $t$ by a non-negligible margin, the concentration step used in Theorems 1–3 fails for this estimator and the bounds would need re-derivation.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that decentralization and information asymmetry are asymptotically almost free in heavy-tailed multi-agent bandits. Under the moment condition $E|X-\mu|^{1+\varepsilon}\le v$, the algorithm mRUCB-A (common reward, hidden actions) attains $R_T=O(\log T\sum_{a\neq a^*}\Delta_a^{-1/\varepsilon})$, the same rate a single-agent heavy-tailed learner achieves over the joint action space; mRUCB-Intervals (visible actions, independent rewards) attains the same leading rate with a larger constant; and mHT-DSEE (full asymmetry) attains $O(K^M\log^2 T)$. The only unavoidable cost is the $K^M$ dependence on the joint action space, which a centralized learner also pays, and the extra log factor in the fully asymmetric regime is the price of not knowing the gaps and having no shared observations. The paper argues that observable actions compensate for the loss of shared rewards at leading order, while the full-asymmetry regime shows the value of even minimal observability.

Load-bearing premise

All three regret bounds rest on the imported concentration inequality for the truncated-mean estimator: the estimate must stay within its confidence radius with probability at least $1-t^{-\gamma}$ at every round, and the paper uses a truncation threshold for which that inequality is borrowed rather than proved.

Editorial extensions

If this is right

When all agents observe the same reward, hiding their individual actions costs nothing in the regret rate: the team matches a centralized learner over the joint action space.
When actions are observable but rewards are independent, a deliberate deviation works as a one-bit implicit signal, and the total cost of all such signaling rounds is at most $(K^M-1)\Delta_{\max}$, independent of the horizon and of the number of players.
In the fully asymmetric regime a pre-committed anytime schedule with $w(t)=\lceil\log t\rceil$ gives $O(K^M\log^2 T)$ regret; knowing $\Delta_{\min}$ would allow a constant schedule and recover $O(K^M\log T)$.
Because only the concentration bound is used, the algorithms can run with any robust estimator, such as median-of-means or Catoni-style confidence sequences, without changing the rates; only the constants change.
The rate hierarchy is asymptotic; at moderate horizons the constants dominate, and in the experiments Problem B's elimination threshold $4\alpha_a<\Delta_a$ costs $2^{(1+\varepsilon)/\varepsilon}$ more samples per arm than the index-based methods.

Reading between the lines

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

The paper does not pursue this, but the formal role played by the common confidence radius suggests a general design rule: any information structure that keeps active-set statistics common knowledge across agents should support elimination at centralized rates, while structures that privatize them will likely pay at least a log factor; testing this on other asymmetry patterns would be a natural ex
Beyond the paper, the anytime schedule in mHT-DSEE can be viewed as a tunable knob rather than a fixed choice: choosing $w(t)=t^{\eta}$ trades exploration cost against the time until agents first agree on the optimal arm, yielding intermediate rates of $O(K^M\log^{1+\eta}T)$ for any $\eta>0$; the paper's Remark 1 gestures at this but does not develop it.
A testable extension is to replace the truncated mean in Problem B with a Catoni-style estimator, since the protocol already supports one-bit communication; the paper predicts only constant improvements, but sharper constants would shrink Problem B's moderate-horizon disadvantage relative to the other two regimes.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

The paper is worth a look. It gives the first systematic treatment of three information-asymmetry regimes for heavy-tailed multi-agent bandits, and the comparison between them is genuinely useful. Problems A and B are clean: shared rewards make coordination free, and observable actions support the signaling-by-deviation mechanism in mRUCB-Intervals. The proofs there are standard union-bound arguments, and they check out. The authors also credit the single-agent machinery in [10] and [11] properly, and no fitted parameters or circular reasoning appear. That part earns its place.

The soft spot is Problem C. The stress-test concern lands. The confidence radius in Section II-C, Eq. (5), uses log(T^γ), and the estimator in Eq. (3) truncates with log(T^γ). But Theorem 3's Step 2 uses α(n) with log(t^γ), and the crossing condition in Eq. (10) cancels log t to make t0 independent of T. Algorithm 3 points back to the RUCB defined in Section II-C, which uses log T. So the paper is running two different confidence radii. With the stated log T version, t0 becomes exp(O(sqrt(log T))) at best, and the claimed O(log^2 T) bound does not follow. This is a load-bearing flaw for the central claim of Problem C, not a cosmetic typo.

That said, it looks fixable. The anytime concentration result from [10] holds with log t, so the authors can restate the estimator and RUCB for Algorithm 3 with log(t^γ), or define the general α(t) with log t and treat log T as a special case for the fixed-horizon algorithms. They should also double-check that the exact truncation threshold in Eq. (3) is covered by [10, Prop. 1] as written, since the per-sample s with log(T^γ) is a nonstandard variant. Those are minor-to-moderate gaps in presentation, but Problem C's theorem is the one that needs the fix.

Experiments are illustrative only, no code shipped, so they add little beyond confirming sublinear behavior. The paper's value is the framework and the A/B results.

For the field: anyone working on multi-agent bandits with limited communication will find the problem formulations and the signaling mechanism useful. I would send this to peer review, but I would require the authors to fix the anytime inconsistency before publication. As it stands, the claims outrun the proof for the hardest regime.

Referee Report

2 major / 4 minor

Summary. The paper studies multi-agent stochastic bandits with heavy-tailed rewards satisfying a (1+ε)-moment condition, under three information-asymmetry regimes: common rewards with unobserved actions (Problem A), independent rewards with observed actions (Problem B), and independent rewards with unobserved actions (Problem C). For each regime it proposes a decentralized algorithm (mRUCB-A, mRUCB-Intervals, mHT-DSEE) and proves regret bounds: O(log T Σ_a Δ_a^{-1/ε}) for Problems A and B and O(K^M log^2 T) for Problem C, based on the truncated-mean robust UCB estimator of [10] and a deterministic exploration schedule in the style of [11]. The paper also reports experiments on an infinite-variance Pareto reward instance with M=2 players and K=2 arms per player.

Significance. If the bounds are established, the paper gives a useful separation among information structures for heavy-tailed multi-agent bandits: shared rewards give costless synchronization, observed actions provide an implicit signaling channel whose cost is independent of the horizon, and full asymmetry costs an extra log factor. The strengths include a clean exact reduction in Problem A, a natural deviation-based elimination protocol in Problem B, explicit constants in the regret bounds, and experiments run with the same estimator as the analysis on a genuinely infinite-variance reward family. The principal weakness is that the Problem C analysis as written relies on a confidence radius that uses log t, while the preliminaries and Algorithm 3 define the radius with log T; this inconsistency is load-bearing for the O(log^2 T) claim. With that issue resolved and the concentration statements restated in terms of pull counts, the paper would be a solid contribution.

major comments (2)

[Section V, Theorem 3, Eq. (10), Algorithm 3] The central O(K^M log^2 T) bound for Problem C is not supported by the algorithm as written. Section II-C defines the robust estimator (3) and the confidence radius (5) with log(T^γ) in the truncation threshold and in α_a(t), and Algorithm 3 (line 6) computes RUCB_i^a(t) without any modification. The proof of Theorem 3, however, uses α(n) = v^{1/(1+ε)}(c log(t^γ)/n)^{ε/(1+ε)} and derives the crossing condition (10) with log t on both sides, so that t0 is independent of T. With the log T radius of (5), the crossing condition is D(t) > cγ log T (2v^{1/(1+ε)}/Δ_min)^{(1+ε)/ε}; since D(t) = ⌈w(t) log t⌉ and w(t)=⌈log t⌉, this requires log t = Ω(sqrt(log T)), giving t0 = exp(Ω(sqrt(log T))). The term Δ_max t0 in (9) then dominates O(log^2 T). The paper must either specify that mHT-DSEE uses a separately defined anytime version of (3)/(5) with log t and prove or cite the corresponding concentration bound, or revise the claimed Problem C rate.
[Section II-C, Eq. (5); proofs of Theorems 1–3] The stated concentration inequality Pr(|bµ_a(t) − μ_a| > α_a(t)) ≤ t^{-γ} uses the global round index t, while α_a(t) in Eq. (5) depends on the pull count n_a(t). If t is the global round, the assertion is false when n_a(t) is small: for n_a(t)=1 the available moment inequality cannot give a t^{-γ}-type tail for large t. The bound from [10, Prop. 1] is in terms of the number of samples s. Consequently the lines 'Pr(G_t^c) ≤ K^M t^{-γ}' and 'the tail sum converges for γ>1' appearing in the proofs of Theorems 1, 2, and 3 should be restated with sums over pull counts s of s^{-γ} (or D(t)^{-γ} in Theorem 3). The final rates are unchanged under this standard correction, but the proofs as written rely on an invalid probability bound.

minor comments (4)

[Algorithm 2, Problem B] The signaling step 'That player pulls a different individual arm' is undefined if the detecting player has an action set of size 1; the paper should either assume K_i ≥ 2 for all players or describe how a singleton player can participate in the signaling protocol.
[Section V, first paragraph] The sentence 'the confidence radius uses log t rather than log T' contradicts the formal definitions in Eqs. (3) and (5) and Algorithm 3; this ambiguity is part of Major Comment 1 and should be resolved by giving a single consistent definition.
[Section VI-A] The text refers to Figure 1, but no figure appears in the manuscript; please verify that the figure is included in the submitted version.
[Section II-C] The remark that any estimator obeying a bound of the form (5) may be substituted is helpful, but the constants in the regret bounds should be made explicitly dependent on the substituted estimator's concentration constants rather than left implicit through c.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the regret bounds are derived from imported external concentration and DSEE results, with no parameter fitted to the experimental outcome and no load-bearing self-citation.

full rationale

All three regret bounds are derived from external concentration results ([10, Prop. 1] and [11]'s DSEE schedule) applied to the truncated-mean estimator, not from the paper's own conclusions. Eqs. (3)-(5) are stated as imported tools, and the paper explicitly says 'Only this concentration property is used below, so any estimator obeying a bound of the form (5)—median-of-means, or the Catoni-style confidence sequences of [15]—may be substituted,' which makes the analysis independent of any particular estimator choice. No parameter is fitted to experimental outcomes; the experiments use the same (c,γ)=(1,2) estimator the analysis assumes, but the analysis does not use experimental data. The self-citations [7], [8] appear only as background on cooperative MMAB and are not load-bearing for the heavy-tailed regret theorems. The reviewer's flagged T/t mismatch in Problem C—Eq. (5) uses log(T^γ) while Theorem 3's Step 2 uses log(t^γ)—is an internal consistency or correctness concern, not a circularity: it does not make any theorem reduce to its assumptions by construction. Therefore no circular step is exhibited and the derivation chain is, on its face, self-contained against external benchmarks.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new free parameters or entities. It depends on a cited concentration inequality and on the heavy-tail moment assumption with known tail parameters, both standard in this literature. The only unusual element is the per-sample truncation in Eq. (3), which is not flagged as a variant of the cited estimator.

assumptions (4)

domain assumption Concentration of the robust estimator: Pr(|bμ_a(t) - μ_a| > α_a(t)) ≤ t^{-γ} for every arm and round.
Stated in Section II-C and attributed to [10, Prop. 1] (or substituted estimators like Catoni sequences). All three theorems use only this property, so its correctness is load-bearing.
domain assumption Rewards satisfy E[|X_a - μ_a|^{1+ε}] ≤ v with known ε and v.
This is the heavy-tail model (2). The algorithms require ε and v as inputs; the paper flags unknown tail parameters as a limitation.
domain assumption Players can pre-agree on a deterministic protocol and know the joint action space and tie-breaking order.
Section II-B states this; all algorithms rely on a shared protocol and shared initialization (e.g., lexicographic ordering, cyclic order).
standard math Standard probability bounds, including union bounds and Borel-Cantelli-type tail summation.
Used throughout the proofs to convert per-round failure probabilities into O(1) regret contributions; these are standard.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Robust Multi-Agent Bandits with Heavy-Tailed Rewards and Information Asymmetry." pith.science (2026). https://pith.science/paper/UHYSBJ4Y

@misc{pith2026260810529,
  author       = {Pith},
  title        = {Pith review of: Robust Multi-Agent Bandits with Heavy-Tailed Rewards and Information Asymmetry},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UHYSBJ4Y}},
  note         = {Machine review of arXiv:2608.10529}
}

read the original abstract

The multi-armed bandit problem is a central framework in sequential decision-making, extensively studied under sub-Gaussian reward assumptions. However, real-world applications often involve heavy-tailed reward distributions and decentralized, information-asymmetric interactions. We study multi-agent multi-armed bandits with heavy-tailed rewards under three information-asymmetry regimes: unobserved actions with common rewards, observed actions with independent rewards, and unobserved actions with independent rewards. We develop robust decentralized algorithms for each setting and derive regret guarantees that nearly match centralized heavy-tailed rates. Experiments on a Pareto-distributed reward environment validate our theoretical findings and illustrate the trade-offs between synchronization, coordination, and exploration across the three regimes.

Figures

Figures reproduced from arXiv: 2608.10529 by the authors.

**Figure 1.** Mean cumulative regret over 10 runs under Pareto rewards with infinite variance (M=2, K=2, T =106 ), log–log axes. place. mRUCB-A ends at 214±24 and mHT-DSEE at 292±21, both still growing slowly, while mRUCB-Intervals ends at 4115 ± 285 but is exactly flat beyond ≈7 × 104 rounds. The ordering at this horizon is governed by constants rather than by rates, and is instructive. Elimination in Problem B requires two inte… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

17 extracted references · 16 canonical work pages

[10]

Bandits with heavy tail,

S. Bubeck, N. Cesa-Bianchi, and G. Lugosi, “Bandits with heavy tail,” 2012

work page 2012
[11]

Deterministic sequencing of exploration and exploitation for multi-armed bandit problems,

S. Vakili, K. Liu, and Q. Zhao, “Deterministic sequencing of exploration and exploitation for multi-armed bandit problems,” 2013

work page 2013
[1]

Some aspects of the sequential design of experiments,

H. E. Robbins, “Some aspects of the sequential design of experiments,” Bulletin of the American Mathematical Society, vol. 58, pp. 527–535, 1952

work page 1952
[2]

On the likelihood that one unknown probability exceeds another in view of the evidence of two samples,

W. R. Thompson, “On the likelihood that one unknown probability exceeds another in view of the evidence of two samples,”Biometrika, vol. 25, no. 3/4, pp. 285–294, 1933

1933
[3]

Competitive collaborative learning,

B. Awerbuch and R. Kleinberg, “Competitive collaborative learning,” Journal of Computer and System Sciences, vol. 74, no. 8, pp. 1271– 1288, 2008. Learning Theory 2005

work page 2008
[4]

Gossip-based distributed stochastic bandit algorithms,

B. Szorenyi, R. Busa-Fekete, I. Hegedus, R. Ormandi, M. Jelasity, and B. Kegl, “Gossip-based distributed stochastic bandit algorithms,” in Proceedings of the 30th International Conference on Machine Learning (S. Dasgupta and D. McAllester, eds.), vol. 28 ofProceedings of Ma- chine Learning Research, (Atlanta, Georgia, USA), pp. 19–27, PMLR, 17–19 Jun 2013

work page 2013
[5]

Decentralized learning for multi- player multiarmed bandits,

D. Kalathil, N. Nayyar, and R. Jain, “Decentralized learning for multi- player multiarmed bandits,”IEEE Transactions on Information Theory, vol. 60, no. 4, pp. 2331–2345, 2014

work page 2014
[6]

Multi-player multi-armed bandits with collision- dependent reward distributions,

C. Shi and C. Shen, “Multi-player multi-armed bandits with collision- dependent reward distributions,”IEEE Transactions on Signal Process- ing, vol. 69, p. 4385–4402, 2021

work page 2021

Show all 17 references

[7]

Online learning for coop- erative multi-player multi-armed bandits,

W. Chang, M. Jafarnia-Jahromi, and R. Jain, “Online learning for coop- erative multi-player multi-armed bandits,”CoRR, vol. abs/2109.03818, 2021

2021 arXiv
[8]

Optimal cooperative multiplayer learning bandits with noisy rewards and no communication,

W. Chang and Y . Lu, “Optimal cooperative multiplayer learning bandits with noisy rewards and no communication,”arXiv preprint arXiv:2311.06210, 2023

2023 arXiv
[9]

A survey on multi-player bandits,

E. Boursier and V . Perchet, “A survey on multi-player bandits,” 2024

2024
[12]

Pure exploration of multi- armed bandits with heavy-tailed payoffs.,

X. Yu, H. Shao, M. R. Lyu, and I. King, “Pure exploration of multi- armed bandits with heavy-tailed payoffs.,” inUAI, pp. 937–946, 2018

2018
[13]

Almost optimal algorithms for linear stochastic bandits with heavy-tailed payoffs,

H. Shao, X. Yu, I. King, and M. R. Lyu, “Almost optimal algorithms for linear stochastic bandits with heavy-tailed payoffs,” inAdvances in Neural Information Processing Systems(S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, eds.), vol. 31, C...

2018
[14]

Minimax optimal bandits for heavy tail rewards,

K. Lee and S. Lim, “Minimax optimal bandits for heavy tail rewards,” IEEE Transactions on Neural Networks and Learning Systems, vol. 35, no. 4, pp. 5280–5294, 2024

2024
[15]

Catoni-style confidence sequences for heavy- tailed mean estimation,

H. Wang and A. Ramdas, “Catoni-style confidence sequences for heavy- tailed mean estimation,”Stochastic Processes and their Applications, vol. 163, p. 168–202, Sept. 2023

2023
[16]

Cooperative multi-agent bandits with heavy tails,

A. Dubeyet al., “Cooperative multi-agent bandits with heavy tails,” in International conference on machine learning, pp. 2730–2739, PMLR, 2020

2020
[17]

Multi-agent multi-armed bandit with fully heavy- tailed dynamics,

X. Wang and M. Xu, “Multi-agent multi-armed bandit with fully heavy- tailed dynamics,” 2025

2025

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.

[1] [10]

Bandits with heavy tail,

S. Bubeck, N. Cesa-Bianchi, and G. Lugosi, “Bandits with heavy tail,” 2012

work page 2012

[2] [11]

Deterministic sequencing of exploration and exploitation for multi-armed bandit problems,

S. Vakili, K. Liu, and Q. Zhao, “Deterministic sequencing of exploration and exploitation for multi-armed bandit problems,” 2013

work page 2013

[3] [1]

Some aspects of the sequential design of experiments,

H. E. Robbins, “Some aspects of the sequential design of experiments,” Bulletin of the American Mathematical Society, vol. 58, pp. 527–535, 1952

work page 1952

[4] [2]

On the likelihood that one unknown probability exceeds another in view of the evidence of two samples,

W. R. Thompson, “On the likelihood that one unknown probability exceeds another in view of the evidence of two samples,”Biometrika, vol. 25, no. 3/4, pp. 285–294, 1933

1933

[5] [3]

Competitive collaborative learning,

B. Awerbuch and R. Kleinberg, “Competitive collaborative learning,” Journal of Computer and System Sciences, vol. 74, no. 8, pp. 1271– 1288, 2008. Learning Theory 2005

work page 2008

[6] [4]

Gossip-based distributed stochastic bandit algorithms,

B. Szorenyi, R. Busa-Fekete, I. Hegedus, R. Ormandi, M. Jelasity, and B. Kegl, “Gossip-based distributed stochastic bandit algorithms,” in Proceedings of the 30th International Conference on Machine Learning (S. Dasgupta and D. McAllester, eds.), vol. 28 ofProceedings of Ma- chine Learning Research, (Atlanta, Georgia, USA), pp. 19–27, PMLR, 17–19 Jun 2013

work page 2013

[7] [5]

Decentralized learning for multi- player multiarmed bandits,

D. Kalathil, N. Nayyar, and R. Jain, “Decentralized learning for multi- player multiarmed bandits,”IEEE Transactions on Information Theory, vol. 60, no. 4, pp. 2331–2345, 2014

work page 2014

[8] [6]

Multi-player multi-armed bandits with collision- dependent reward distributions,

C. Shi and C. Shen, “Multi-player multi-armed bandits with collision- dependent reward distributions,”IEEE Transactions on Signal Process- ing, vol. 69, p. 4385–4402, 2021

work page 2021