acceptodds
Under review as a conference paper at ICLR 2027

Pathwise Individual Rationality in Federated Learning: A Mechanism-Architecture Co-Design

Abstract

Participation in federated learning (FL) comes at a cost. Clients trade off privacy, communication, and compute costs for potentially greater gains in model efficacy. This paper explores this tradeoff under the aegis of individual rationality (IR) versus autarky, the basic game-theoretic requirement that the federation provide utility no worse than local training. Using the above as the design target, we examine pathwise performance of FL, as a per-round bound on cumulative surplus rather than only an asymptotic equilibrium guarantee under different degrees of client data heterogeneity. Along the learning path, clients can remain below their local-training baseline for hundreds of rounds. The natural remedy is to cap each client's per-round contribution so that this shortfall stays bounded. We prove that such a cap drives all contributions to zero, with high probability, when its tolerance is small relative to the warm-up deficit, and we observe learning collapse under it even at low heterogeneity. We then propose a design that combines a short-term participation guarantee, enforced after an announced grace window, with personalized model evaluation, while preserving the incentive properties of the underlying mechanism. We provide a theoretical basis for this approach and empirically demonstrate that, after the grace window, clients meet IR on nearly all rounds without harming overall performance under low to moderate heterogeneity; under severe heterogeneity, the design shows promising outcomes for clients compared to their local baseline at some cost in accuracy.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
federated learning, individual rationality, mechanism design, game theory, incentives, best-response dynamics, personalization, data heterogeneity

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_r5JtOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies cumulative participation losses relative to local training. It shows that early contribution clipping can stop learning, then combines a grace period with personalized heads. Experiments show substantial participation gains, but the implemented contract does not consistently enforce its deficit bound.

Critical strengths
  1. Participation during training is a useful concern beyond equilibrium welfare. The collapse analysis and experiments show why an immediate deficit constraint can be counterproductive.
  2. The appendix provides informative ablations and reports failures openly, including ledger breaches, personalization costs, and clients whose local models leave little room for improvement.
Critical weaknesses
  1. The guarantee covers an ideal regime where clipping never activates and assumes every client benefits at equilibrium. It does not cover the main deployed regime. Table 4 reports 23 deficit-bound breaches among 180 client-seed pairs in the difficult CIFAR-10 setting. Personalization without clipping has the same breach count, higher participation IR, and higher accuracy. The clipping component’s benefit is unclear.
  2. Incentive preservation assumes agents continue optimizing the original global-model utility, although their realized utility comes from personalized models. Unchanged payments and prescribed updates do not establish unchanged strategic behavior under that realized utility. The paper acknowledges this different game but leaves it unanalyzed.
Questions for authors
  1. Can the deployed rule enforce the announced post-grace bound, including estimation errors? If not, how should the contract and preservation claims be narrowed?
  2. What happens when clients choose contributions to maximize their personalized utility? Would the equilibrium and participation improvements survive?
2 comments

We thank the reviewer for taking the time to review our paper.

W1 (first part): the guarantee and the deployed regime.

Theorem 2 needs a tolerance at which the clip never engages, because it bounds the ledger along the base mechanism's path to its equilibrium, in the model in which that equilibrium is proved, where accuracy depends only on the contributions; once the clip acts, the path leaves the base mechanism's (our reply to Reviewer xvkG, W2, gives the reasons). For the regime in which the clip engages, we added a per-round result that assumes nothing about the equilibrium: Proposition 2 (Appendix A.9), stated under Q1 below. On every client benefiting at the equilibrium, the Limitations already say why Theorem 2 needs it: an agent that loses there accumulates deficit along the unconstrained trajectory itself, so no contract that follows that trajectory can save it. The revision states the theorem's scope in Section 7.1 and the Limitations.

Q1: can the deployed rule enforce the bound, with estimation errors, and how should the claims be narrowed?

In each round in which the agent is sampled, yes, as far as its own contribution in that round can, with an error of at most 0.006 in our runs. Over the run, not in general: when the shortfall is not in the agent's own cost, a ledger can still fall below −Δi-\Delta_i, as 23 of the 180 agent-seed pairs do at CIFAR-10 α=0.1\alpha=0.1. The revision states both parts as Proposition 2 (Appendix A.9). It holds for the clip as implemented (Algorithm 1), on any accuracy path and any sampling, with the clip's estimation error. It assumes nothing about the equilibrium, so it also covers agents that do not gain there. In a round from TgT_g on in which agent ii is sampled:

  • (a) If some contribution passes the clip's test, the clip selects one, and after the round the agent's ledger is at least −Δi−ei(t)-\Delta_i-e_i(t), where ei(t)e_i(t) is the error from the one-round lag in the payment the test uses.
  • (b) Otherwise, the clip sets the contribution to zero. The agent's per-round surplus is then ai(t)−uˉi−β∗n−1∑j≠icjsja_i(t)-\bar u_i-\frac{\beta^*}{n-1}\sum_{j\neq i}c_js_j, where ai(t)a_i(t) is the accuracy its ledger reads (the deployed head's in the co-design): its accuracy gap to autarky minus its share of the other agents' costs. No choice of its own contribution gives a higher surplus in that round.
  • (c) So a breach larger than the largest ei(t)e_i(t) can build up only over rounds of type (b) and rounds in which the agent is not sampled.

In the 21 co-design runs of Table 4, case (a) covers 70,994 sampled agent-rounds, with the largest breach of 0.0057 against Δi=1\Delta_i=1, and case (b) covers 2,716.

How the claims read in the revision. For tolerances at which the clip never engages, the contract's claim is Theorem 2's bound, under its assumptions, unchanged, and the contribution list now says "under a sufficient tolerance". When the clip engages, the claim is Proposition 2. The preservation claim is Theorem 3's, for agents that best-respond on the base mechanism's strategic utility (W2 and Q2 below). Once the clip engages, the path leaves the base mechanism's, as Appendix A.8 states.

W1 (second part): what the clip adds beyond personalization.

Reviewer hrA1 asks the same question, and our reply there answers it in full. In short, the clip keeps the bound in each round in which the agent is sampled, as far as the agent's own contribution in that round can, up to the clip's estimation error (Q1 above). But a lower contribution saves only the unrefunded part of the agent's own cost, under 5% of it, so almost the same agents cross with and without the clip, and most of the protection comes from the heads. The accuracy cost is set by the tolerance: at CIFAR-10 α=0.1\alpha=0.1 the co-design at Δi=20\Delta_i=20 is identical to the heads-only run (Appendix C.8).

W2 and Q2: agents that choose contributions to maximize their personalized utility.

We added this case (Appendix C.10) and revised Section 6 to reflect it: the paper's equilibrium does not survive, and the participation improvements shrink or reverse. Theorem 3 makes no claim about agents that decide on their personalized utility. It is stated for agents that best-respond on vv, the base mechanism's strategic utility, which uses the shared model's accuracy, and it shows that the heads themselves do not disturb the mechanism: they change neither the payments, nor the best-response update, nor its equilibrium, because none of these reads the heads. In the new runs, all agents choose contributions from the accuracy of the head they are deployed with, from round 1, and everything else is as in the paper (42 runs, each paired with the paper's run of the same seed, partition, and costs).

Paper's run → all agents decide on the deployed headCIFAR-10 α=10\alpha=10α=1\alpha=1α=0.1\alpha=0.1MNIST α=10\alpha=10α=1\alpha=1α=0.1\alpha=0.1
Total contribution after the window, heads only−24%−71%−79%−46%−54%−74%
Accuracy, heads only84.3 → 82.386.6 → 76.764.7 → 48.499.0 → 98.999.0 → 98.798.6 → 96.4
Participation IR, heads only98.8 → 93.3100.0 → 86.291.2 → 61.697.5 → 97.394.3 → 90.542.4 → 37.3
Accuracy, co-design84.3 → 80.986.6 → 65.059.4 → 25.999.0 → 98.999.0 → 98.798.5 → 96.2
Participation IR, co-design98.7 → 92.2100.0 → 63.987.7 → 31.497.5 → 97.394.3 → 90.542.7 → 37.8
Agents whose ledger ends the run lower, heads only90/9090/90170/18090/9090/9049/90

Accuracy is the shared model's over the final 50 rounds, participation IR is over the rounds after TgT_g, and the last row sums each agent's ledger from round 1.

Contributions fall in every configuration and every seed, because on average, an agent's deployed accuracy responds far less to its own contribution than the shared model's does. Accuracy falls with them in every configuration. At α=0.1\alpha=0.1 participation IR stays above the grace contract's without heads, and at α≥1\alpha\geq1 it ends below the grace contract's (Table 3). And 579 of the 630 agent-seed pairs end the run with a lower ledger than in the paper's runs.

So the improvements depend on contributions being decided on the shared model, as Theorem 3 assumes and as the co-design's update does (Algorithm 1). Section 6 now says so.

Reviewer_hrA1OpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies and examine the losses clients may incur while a federated model is still being trained. The authors define pathwise individual rationality in terms of cumulative utility relative to local training. They show that clipping contributions to limit early losses can cause training to collapse, and propose a grace period after which the cumulative-loss calculation starts afresh. They also add personalized heads so that clients with different data distributions can make better use of the shared model. The paper provides conditions for collapse and conditional guarantees for the post-grace period. Experiments on CIFAR-10 and MNIST vary client heterogeneity and test the components separately. A grace period is generally enough at lower heterogeneity, while personalization helps substantially in the harder CIFAR-10 setting. The results also show the limits of the approach when local training already gives clients near-perfect accuracy.

Critical strengths

The paper provides a theoretical explanation for why directly enforcing pathwise IR can cause training to collapse. Rather than only showing that contribution clipping performs poorly, Theorem 1 identifies conditions under which it drives all contributions to zero with high probability. The experiments support this result, with collapse observed in 56 of 57 runs under small tolerances. The additional analysis of client sampling rates also supports the predicted relationship between participation frequency and collapse time. I find this a meaningful contribution because it explains why a seemingly reasonable participation constraint can undermine the learning process itself.

The experimental analysis helps distinguish the effects and limitations of the proposed components. The paper evaluates grace periods and personalized heads under different levels of data heterogeneity, including settings where the proposed approach is less effective. The ablations compare the full method with personalization without clipping and clipping without a grace period. The results show that grace periods can prevent collapse under moderate heterogeneity, while personalization becomes important under severe heterogeneity. The experiments also report the accuracy cost of enforcing contribution constraints and the limited benefit when local training already performs well. These comparisons make the findings more informative than a simple performance comparison against existing FL methods.

Critical weaknesses

The additional benefit of contribution clipping remains unclear. While personalization improves participation utility, the contribution of the clipping mechanism itself is less convincing. In the difficult CIFAR-10 setting ( alpha= 0.1), the co-design and heads-only baseline have the same number of cumulative-bound violations (23 out of 180), but the co-design achieves lower accuracy and per-round participation IR. This raises questions about whether clipping provides meaningful additional protection, particularly when it can prevent some clients from reaching positive surplus later in training. Since enforcing cumulative participation constraints is one of the paper's main objectives, this limits the practical significance of the proposed contract.

Questions for authors
  1. What additional protection does contribution clipping provide beyond personalization? In Tables 3 and 4, the co-design and the heads-only baseline have the same number of agent-seed pairs whose post-grace ledger crosses −Δᵢ (23 out of 180) on CIFAR-10 at α = 0.1. Meanwhile, heads-only achieves higher accuracy (64.7% vs. 59.4%) and per-round participation IR (91.2% vs. 87.7%). Does clipping reduce the magnitude or duration of cumulative deficits, even though the number of violations is unchanged? It would be helpful to clarify what additional benefit the clipping mechanism provides and whether this benefit justifies the observed accuracy cost.

  2. How does the mechanism-preservation result apply when clients consider their personalized utility? Theorem 3 assumes that clients continue to best-respond on the original strategic utility vᵢ, even though their participation utility is evaluated using personalized heads. The paper acknowledges that a client best-responding on its personalized utility plays a different game, which is not analyzed. Since the paper considers self-interested clients, what justifies the assumption that they would not adjust their contributions based on the utility they actually experience?

3 comments

We thank the reviewer for taking the time to review our paper.

Weakness and Q1: what the clip adds beyond personalization.

What additional protection the clip provides. The clip is how the contract acts on its bound, ΨiTg(t)≥−Δi\Psi_i^{T_g}(t)\geq-\Delta_i. With a fixed set of agents and the base mechanism's payments unchanged, the one thing the contract can limit is the agent's own contribution: when the agent's unconstrained update would take its ledger below −Δi-\Delta_i, the clip lowers the contribution to the largest that keeps the ledger at or above −Δi-\Delta_i, and to zero if none does. In the heads-only run nothing limits the contribution: an agent whose ledger is past its tolerance keeps contributing. The clip cannot repair a shortfall that is not in the agent's own cost. Under Mβ∗M_{\beta^*} an agent is refunded about 95% of its cost, so stopping saves it under 5% of that cost. Its accuracy gap to autarky and its share of the others' costs remain. This is why, at CIFAR-10 α=0.1\alpha=0.1, 23 of the 180 agent-seed pairs (30 agents, six seeds) cross with the clip and 23 without it, 22 of them the same, and why the protection has to come from the heads, which act on the accuracy gap. Proposition 2, which we added as Appendix A.9, states exactly what the clip can and cannot do in each round (our reply to Reviewer r5Jt, Q1).

Magnitude and duration. The clip reduces the deficit of agents that already reach near-perfect accuracy alone, and deepens and lengthens it for most of the others. Stopping a contribution lowers the costs that all agents share, and it takes data from the model. Of the 30 agent-seed pairs the clip stops at α=0.1\alpha=0.1, 12 already reach at least 98.7% accuracy alone (7 at MNIST, 5 at CIFAR-10), and all 12 end with a smaller deficit. The other 18 are at CIFAR-10, where the model needs their data: its accuracy ends 5.3 points lower, and their deficit is deeper with the clip for 17 and lasts longer for 14. The revision adds the depth of each agent's worst deficit, with and without the clip, to Appendix C.7.

Whether the benefit justifies the accuracy cost. As Section 7.3 already reports, the price of the clip is set by the tolerance, which the designer chooses. The heads-only run is actually a version of co-design at a large enough tolerance where there won't be any clips along the path: at CIFAR-10 α=0.1\alpha=0.1 the co-design runs at Δi=20\Delta_i=20 are identical to CIFAR-10 α=0.1\alpha=0.1 heads-only runs, agent by agent and round by round (Appendix C.8). So Tables 3 and 4 compare two tolerances. At Δi=1\Delta_i=1 the clip stops those 23 agents. At Δi=20\Delta_i=20 it stops none, and no ledger falls below the tolerance. In that configuration the larger tolerance is the better choice. In the other five configurations the clip changes accuracy by at most 0.08 points at Δi=1\Delta_i=1 (Appendix C.6), and in 11 of the 21 co-design runs it never engages.

Q2: agents that adjust contributions to the utility they experience.

How the preservation result applies. Only in part. Theorem 3's mechanism-preservation claim rests on one fact: fitting the heads changes only the heads, not the shared model, so nothing the base mechanism computes reads them, neither the strategic utility vv on which agents best-respond nor the payments (Appendix A.8). An agent that best-responds on its personalized utility reads its own head, so its update, the path of contributions and the equilibrium then depend on the heads, and Theorem 3 makes no claim about them. The payments still depend only on the contributions, so the payment rule is unchanged and budget balance still holds.

What justifies the assumption. It is a design choice: the co-design computes each update from the shared model's accuracy, and the heads enter only the participation ledger and the clip (Algorithm 1). We added the case you describe, with all agents deciding on their deployed head (Appendix C.10; full results in our reply to Reviewer r5Jt, W2 and Q2). Contributions then fall in every configuration and every seed, because on average an agent's deployed accuracy responds far less to its own contribution than the shared model's does. And 579 of the 630 agent-seed pairs end the run with a lower ledger than in the paper's runs. So the improvements depend on contributions being decided on the shared model, and Section 6 now says so.

Thank you for the detailed response and the additional analyses.

Q1. The role of the clip is now clearer. However, on CIFAR-10 (α = 0.1), the clip does not reduce the number of violations. For most agents, it actually deepens the deficit and lowers accuracy. The protection seems to come mainly from the personalized heads, and I think the paper should state this clearly.

Q2. Thank you for adding the experiment in Appendix C.10. When agents decide based on their deployed head, both contributions and ledgers decrease. Since the paper assumes self-interested clients, I see this as a substantive limitation.

The additional analyses are helpful, but my main concerns remain, so I will keep my score.

Thank you for the follow-up.

Q1. We will add this to Section 7.3: "At this configuration, the deficits mainly come from accuracy below autarky, which the personalized heads will raise; the clip can only lower the agent's own cost, mostly refunded by the base mechanism, so it does not reduce the number of agents that cross −Δi-\Delta_i (Table 4)."

Q2. The clients in the base mechanism are already self-interested: each chooses its contribution to raise its own utility (Section 3). The co-design keeps this, and the personalized head provides a better personal lens through which each client measures what they get from the federation against training alone (Section 6). The two serve the client differently. The personalized head already tunes the shared model to the client's own data. Anything more the client gains comes from the other clients' data, which reaches it only through the shared model. Deciding on its deployed head, the client sees little return from contributing, since the head has already adapted to its data: its deployed accuracy responds far less to its own contribution, and as a result, the client contributes less. Doing so works against the clients themselves: when every client decides on its own head, contributions fall, the shared model under every head gets worse, and most clients end up worse off by their own heads (Appendix C.10).

Reviewer_xvkGOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper focuses on participation incentives along FL trajectories rather than only at equilibrium. It defines a pathwise individual rationality through a lower bound on cumulative utility versus autarky baseline. Under a uniform warm-up deficit assumption, the authors prove that sufficiently restrictive contribution clipping causes contributions to collapse. The authors also introduce a grace period and derives sufficient conditions under which the unconstrained trajectory satisfies the post-grace bound. Personalized heads improve the utility used to assess participation while leaving the stipulated global-head strategic utility unchanged. Experiments on CIFAR-10 and MNIST investigate clipping, grace periods, personalization, heterogeneity, and participation rates.

Critical strengths
  1. The paper identifies a meaningful issue of Individual Rationality in FL: equilibrium-only analysis, and emphasizes that the eventual outcome need not compensate participants for an unattractive learning trajectory.
  2. The empirical analysis is solid: multi-failure modes are reported, including severe heterogeneity failures, substantial seed variation, autarky saturation, and the performance of personalization without clipping. These disclosures are valuable for practice: they show where the proposed mechanism applies and where its limitations lie.
Critical weaknesses
  1. I feel the the paper is not enough establish the participation incentives suggested by the co-design framing. The proposed grace contract would reset the ledger and excludes the initial deficit. Announcing such exclusion does not establish that a self-interested participant through the grace period would accept or remain. The result additionally assumes that agents continue updating according to the original strategic utility, which is a strict assumption.

  2. in theorem 2, the authors establishe that a sufficient-tolerance regime where clipping never activates, while guarantee transfer uses an optimal head and an unbounded approximation gap separates that head from the deployed one. These are legitimate conditional results, but they do not certify the operative regime motivating the empirical co-design.

Questions for authors
  1. Is there any real-world participation or commitment model makes joining and remaining through the grace period rational?
  2. Is there a guarantee for cumulative utility from once entry, or is the current claim restricted to a conditional post-grace accounting constraint?
  3. Can you separate data used for derivative estimation and clipping from an untouched evaluation set?
3 comments

We thank the reviewer for taking the time to review our paper.

W1 (first part) and Q1: is there a real-world participation or commitment model that makes joining and staying through the grace window rational?

Yes: it is the commitment that the line of work we build on already uses, and we are now including it in Section 5 of the revision draft. An agent joins for what the whole run gives it compared with training alone. In Karimireddy et al. (2022), individual rationality is one comparison of the model an agent receives with the model it would train alone, and the base mechanism of Murhekar et al. (2023) fixes the set of agents and gives equilibrium guarantees, with no condition on the path. A real cross-silo federation set up this way is MELLODDY: in a three-year EU co-funded research project, ten pharmaceutical companies ran one federated run per year, all ten contributed to the shared model in every learning iteration, and each company realized aggregate improvements on its own classification or regression models (Oldenhof et al., 2023; Heyndrickx et al., 2024). Our model keeps this commitment and adds terms for the path: the unprotected period has an announced length TgT_g, and after it the mechanism acts on a tolerance Δi\Delta_i for each agent's cumulative loss from then on. We know of no federation with an announced window followed by a guarantee; that part is what our contract proposes. The announcement is not what makes the run worth joining. The outcome is, and the revision now reports it from round 1 (Appendix C.7):

Counted from round 1CIFAR-10 α=10\alpha=10α=1\alpha =1α=0.1\alpha =0.1MNIST α=10\alpha =10α=1\alpha =1α=0.1\alpha =0.1
Base mechanism (FedBR-BG): agents that end the run ahead of autarky90/9090/905/18025/9012/900/90
Grace contract without heads: same count90/9090/901/18025/9012/900/90
Co-design: same count90/9090/90158/18067/9058/9017/90
Without heads: net loss over the grace window, median / largest16.2 / 24.317.6 / 31.155.2 / 87.34.1 / 7.35.3 / 7.812.0 / 26.2
Co-design: same13.5 / 21.25.3 / 15.32.6 / 24.41.7 / 3.51.5 / 3.31.1 / 5.3

Δi=1\Delta_i=1, Tg=100T_g=100; 600 rounds on CIFAR-10 and 250 on MNIST. Counts are agent-seed pairs (30 agents, 3 seeds; 6 seeds at CIFAR-10 α=0.1\alpha=0.1). An agent ends ahead if its per-round surplus, summed from round 1 to the last round, is at least zero. Net losses are the median / largest over agent-seed pairs, in units of Δi\Delta_i. The co-design rows use the deployed head's accuracy, the other rows the shared model's. Before TgT_g the grace contract's runs are identical to the base mechanism's, so the two share the net-loss row.

Counted from round 1, 480 of the 630 agent-seed pairs end the run ahead of autarky under the co-design, against 222 under the base mechanism: more in four configurations, and every agent under both in the other two. Staying is covered by the same comparison: for each of these 480, the rounds still to come add up to a gain at every round of the window. The contract without heads gives nearly the base mechanism's counts (218), so the difference comes from the personalized heads. The contract sets the terms of the commitment, and the heads leave more agents ahead. On MNIST at α=10\alpha=10 and 11, 23 and 32 of 90 are still behind under the co-design when the run ends at round 250, and 22 and 29 of them are gaining on average over the final 50 rounds. MNIST at α=0.1\alpha=0.1 is the configuration where 14 to 19 of the 30 agents already reach at least 99% accuracy alone, above what the federated model attains (Section 7.3). Our model fixes the set of agents for the run, and Section 8 names relaxing this as the first open question.

Q2: a guarantee from entry, or a post-grace constraint?

The guarantee is for the rounds after the window, and it was stated in Section 5: the contract "bounds the post-grace ledger, not the ledger from round 1". This restriction is forced by Theorem 1. Clipping that also covers the warm-up, at a tolerance below 2κi2\kappa_i, drives every contribution to zero within 21 rounds with probability at least 0.95 (n=30n=30, m=10m=10), provided the warm-up lasts that long. In our experiments, 56 of the 57 bare-clipping runs with Δi≤2\Delta_i\leq2 collapse, each within 20 rounds. Counted from entry, the contract's constraint reads Ψi(t)≥−Δi−Wi\Psi_i(t)\geq-\Delta_i-W_i for t≥Tgt\geq T_g, where WiW_i is the net loss over the grace window (table above). For the whole run, we give the measurement above instead of a bound.

W1 (second part): agents keep updating on the original strategic utility.

Theorem 3 assumes this, and we added the case where they do not (Appendix C.10 and Section 6; results in our reply to Reviewer r5Jt, W2, and Q2). When all agents decide on their deployed head, contributions, accuracy, and participation IR fall in every configuration, and most agents end the run worse off than in the paper's runs, so the improvements depend on the assumption.

W2: The theorems do not certify the regime of the experiments.

We added a per-round result for that regime. Proposition 2 (Appendix A.9) needs neither condition the review names: it holds when the clip engages, and it is stated on the deployed head's ledger. It covers the clip as implemented, on any accuracy path. It states what the clip guarantees in each round, up to its estimation error (at most 0.006 in our runs), and shows that a larger breach can build up only in rounds in which no contribution passes the clip's test and in rounds in which the agent is not sampled. Table 4 counts the agents whose ledger crosses the tolerance. The full statement is in our reply to Reviewer r5Jt (Q1).

The theorems themselves are stated in the model in which the base mechanism's equilibrium guarantees are proved, where accuracy depends only on the contributions, because both build on that equilibrium: Theorem 3 preserves it, and Theorem 2 bounds the ledger along the path to it. This is also why Theorem 2 needs a tolerance at which the clip never engages: once the clip acts, the path leaves the base mechanism's (Appendix A.8), and the convergence argument behind the bound no longer applies. The experiments run the federated protocol itself, one best-response step per round on a model that is still training, as Murhekar et al. (2025) do in their FL protocol; there, the paper measures the bound instead of claiming it, and Section 7.3 says where the clip engages. The revision states this in Section 7.1, and the scope of Theorems 2 and 3 in the Limitations, and the contribution list now says "under a sufficient tolerance".

On the optimal and the deployed head, Section 6 already covers this point. The identity of Corollary 2 holds for the deployed head as it is: its ledger equals the mechanism's ledger plus the gain that head accumulates. Only the sign of the gain needs the optimal head, and Section 6 reports it in the runs: the gain is non-negative on at least 86% of post-grace agent-rounds in every configuration, and Table 1 gives its mean.

Q3: an evaluation set untouched by derivative estimation and clipping.

Yes. We reran the co-design, all 21 runs, with each agent's test data split once, at random, into two halves. Every decision reads one half: the derivative estimates, the clip's accuracy reading and the ledger the clip checks. No decision reads the other half.

Decision half / untouched halfCIFAR-10 α=10\alpha=10α=1\alpha=1α=0.1\alpha=0.1MNIST α=10\alpha=10α=1\alpha=1α=0.1\alpha=0.1
Accuracy of the shared model85.0 / 85.186.8 / 86.859.6 / 59.799.2 / 99.199.1 / 99.298.6 / 98.4
Accuracy of the deployed heads84.8 / 85.088.2 / 88.193.0 / 92.999.1 / 99.099.1 / 99.199.4 / 99.3
Participation IR99.6 / 99.6100.0 / 99.984.5 / 87.795.2 / 95.384.6 / 89.739.9 / 35.9

The decisions shape the trained model, so the direct test is its accuracy on the half they read, and on the half they do not. The two agree within 0.2 points for the shared model and within 0.3 points for the deployed heads, in every configuration and with no consistent sign. Participation IR compares the deployed head's accuracy on each half with the agent's autarky accuracy. That accuracy is measured once on the agent's whole test split before the run, as in the paper, so no decision can change it. Participation IR differs between the halves by at most about 5 points, again with no consistent sign. It is more sensitive than accuracy because it counts the sign of the per-round surplus, and on MNIST, 23% to 78% of these surpluses are within 0.01 of zero. These are new runs whose decisions read half as many test images, so we compare the two halves within each run and not with Table 1.

References

  • Sai Praneeth Karimireddy, Wenshuo Guo, and Michael I. Jordan. Mechanisms that Incentivize Data Sharing in Federated Learning. arXiv:2207.04557, 2022.
  • Aniket Murhekar, Zhuowen Yuan, Bhaskar Ray Chaudhury, Bo Li, and Ruta Mehta. Incentives in Federated Learning: Equilibria, Dynamics, and Mechanisms for Welfare Maximization. Advances in Neural Information Processing Systems 36 (NeurIPS), 2023.
  • Aniket Murhekar, Jiaxin Song, Parnian Shahkar, Bhaskar Ray Chaudhury, and Ruta Mehta. You Get What You Give: Reciprocally Fair Federated Learning. Proceedings of the 42nd International Conference on Machine Learning (ICML), PMLR 267, 2025.
  • Martijn Oldenhof et al. Industry-Scale Orchestrated Federated Learning for Drug Discovery. Proceedings of the AAAI Conference on Artificial Intelligence 37(13), 2023.
  • Wouter Heyndrickx et al. MELLODDY: Cross-pharma Federated Learning at Unprecedented Scale Unlocks Benefits in QSAR without Compromising Proprietary Information. Journal of Chemical Information and Modeling 64(7), 2024.

Comments on the paper

1 comment

We thank all reviewers again for taking the time to review our paper carefully and for providing rigorous comments. As noted in our comments, we have uploaded a revised version of the paper, with the added text in blue. The main text changes only in a few places. Most additions are in the appendix and reflect the new experiments and analysis that the reviewers asked for: a per-round guarantee for the clip as implemented (Proposition 2, Appendix A.9), runs in which agents choose their contributions on their deployed head (Appendix C.10), and, in Appendix C.7, how many agents end the run ahead of autarky counted from round 1 and each agent's worst deficit with and without the clip. If you feel that some of this material belongs in the main paper, we are happy to reorganize the paper accordingly. We are glad to answer any further questions or take further feedback until the end of the discussion period.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.