{"id":"b3e3eff2-fc25-4642-86e0-2ddfd0d15051","arxiv_id":"2502.12176","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A survey that defines ten challenges in Federated Foundation Models and unifies them into a single weighted optimization objective.","lead":"This paper maps out ten open research problems in Federated Foundation Models and expresses each one as a formal optimization equation. It is a survey and roadmap, not a new algorithm, aimed at researchers choosing problems in this emerging field.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (17) does not reproduce the per-problem formulations: privacy constraints δ_k are dropped and Problem 4 is mislabeled as privacy-loss minimization, so the 'first mathematical unification' claim is unsupported.","rationale":"The reader's verdict of CONDITIONAL is appropriate, and my stress test does not move it. The reader's weakest assumption, the deterministic modeling of privacy loss, is a real simplification, but it is not the most load-bearing weakness of the central claim. The more direct problem is that Eq. (17), which the paper explicitly claims is the first mathematical unification, is internally inconsistent with the formulations it is supposed to unify: hard privacy constraints are dropped, the stated reduction to Problem 4 does not match Eq. (7), and the minimization over mechanisms F and C is not a well-posed optimization program. These issues are internal to the paper, not matters of external consensus, and they directly undercut the strongest claim. The survey content and taxonomy retain value, so the paper should not be rejected, but the Section XII claim should be revised or supported by explicit derivations of each reduction. Therefore the correct verdict remains CONDITIONAL, unchanged from the reader's verdict.","tokens_in":31536,"tokens_out":2844,"duration_ms":31162,"concrete_test":"Instantiate each of the ten problems as a special case of Eq. (17) by writing down the exact substitution, including the values of α_i, the remaining variables, and any constraints. Then compare each instantiation against the original formulation in the corresponding section. Particularly verify: (1) with α1=1 and all other α_i=0, does Eq. (17) become Eq. (4) including the constraint ℓ_{p,k}(w) < δ_k for all k? (2) with α5=1, does Eq. (17) become Eq. (7), i.e., min dist(w_un, w_re), rather than a privacy-loss minimization? (3) what is the well-defined feasible set for the variables F and C in Eq. (17)? If any of these checks fails, the claimed unification is not faithful and the Section XII claim should be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in Section XII is that Eq. (17) is the first mathematical formulation unifying all ten FedFM problems. This claim fails on internal consistency, not merely on external scope. First, Eqs. (4), (9), and (16) encode privacy as a hard constraint ℓ_{p,k}(w) < δ_k, but Eq. (17) contains no δ_k term and places ℓ_{p,k} only as a weighted objective component. A weighted objective cannot represent a hard constraint without an additional constraint set, so Eq. (17) does not subsume Problems 2, 5, and 10 as stated. Second, the claimed reduction 'α5 = 1 optimizes the privacy loss (i.e., problem 4)' is incorrect: Problem 4's formulation, Eq. (7), minimizes dist(w_un, w_re), the distance between an unlearned model and a retrained model, and no privacy-loss term ℓ_p appears in that objective. Thus the mapping from Eq. (17) to Problem 4 is not a reduction. Third, Eq. (17) writes 'min' over wg, {a_k}, F, and C, where F is an aggregation mechanism and C is a contribution evaluation mechanism; these are functions, not ordinary decision variables, and no feasible set, regularity condition, or definition of the loss arguments is supplied. As a result, Eq. (17) is not a well-defined optimization program. These are internal inconsistencies between the unified equation and the per-problem equations it claims to unify, so the 'first mathematical perspective' claim is not supported unless Eq. (17) is amended and the reductions are actually carried out.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript is a survey/position paper on federated foundation models (FedFMs). It identifies ten problems—foundational theory, private data utilization, continual learning, unlearning, Non-IID/graph data, bidirectional knowledge transfer, incentive and contribution evaluation, game mechanisms, model watermarking, and efficiency—and organizes them under five themes. For each problem it gives a short optimization formulation (Eqs. (1)–(16)), reviews representative methods, and lists challenges and potential solutions. Section XII then proposes Eq. (17) as a single weighted objective over utility, efficiency, watermark, contribution, and privacy losses, and claims that this is the first mathematical perspective unifying all ten problems. The per-problem formulations and the survey material are the paper's main content.","tokens_in":31896,"tokens_out":8194,"duration_ms":77516,"significance":"The survey is timely and useful as a map of a rapidly growing area: it assembles a large reference set, compares against existing reviews in Table II, and attempts to give each of the ten challenges a crisp mathematical statement. If Eq. (17) were a faithful unification, it would provide a valuable common formalism for reasoning about trade-offs across privacy, utility, efficiency, watermarking, and contribution. However, the central unification claim is not supported by the equations as written: several per-problem formulations are informal or inconsistent, and Eq. (17) drops the privacy thresholds and mislabels Problem 4. The value of the manuscript lies mainly in its taxonomy and literature coverage; the mathematical contributions need substantial revision before they can serve as a foundation.","major_comments":[{"comment":"The unified equation does not subsume the privacy-constrained problems it claims to cover. In Eqs. (4), (9), and (16), privacy appears as a hard constraint ℓ_{p,k}(w) < δ_k, while Eq. (17) contains no δ_k and places ℓ_{p,k} only in the weighted objective. A weighted objective term cannot enforce a hard constraint without an explicit constraint set, so the reductions α1=1 and α5=1 do not reproduce Problems 2, 5, or 10. The additional claim that α5=1 'optimizes the privacy loss (i.e., problem 4)' is also incorrect: Problem 4 is defined in Eq. (7) as minimizing dist(w_un, w_re), with no ℓ_p term in the objective. The mapping from Eq. (17) to Problem 8 is likewise unexplained, since Eq. (12) is a utility maximization rather than a utility-loss minimization. Until these reductions are actually carried out, the 'first mathematical perspective' claim in Section XII is unsupported.","section":"Section XII, Eq. (17)"},{"comment":"Eq. (17) is not a well-defined optimization program. The minimization is over wg, {a_k}, F, and C, where F is an aggregation mechanism and C a contribution evaluation mechanism; these are function-valued objects, and no function space, feasible set, or regularity condition is given for them. The notation F_{k∈[K],t∈[T]} appearing before the objective is undefined, and the loss terms are written with arguments (ws, wk, ak, C, {D_t^k}, G) even though C is simultaneously an optimization variable and an argument of the losses. A reader cannot evaluate, solve, or instantiate Eq. (17) for any of the ten problems.","section":"Section XII, Eq. (17)"},{"comment":"The 'no free lunch' constraint ℓu(w)+ℓp(w)+ℓe(w)>0 is trivial for any nonnegative loss functions and therefore does not express the privacy-utility-efficiency trade-off the text describes. A trade-off requires either a lower bound relating the three losses, as in Eq. (2), or a Pareto-front characterization. The same trivial inequality is reused as the constraint in Eq. (17), so the foundational-theory Problem 1 is not captured by the unified equation.","section":"Section II-A, Eq. (1)"},{"comment":"The contribution-evaluation objective depends on an Oracle that knows the true contribution C*_k for each client. Such ground-truth contributions are unobservable in any real federated system, and the manuscript does not define C*_k or explain how existing schemes (Shapley value, leave-one-out, least core) relate to it. As written, Eq. (11) turns contribution evaluation into supervised regression against an unknown target, which cannot be computed or verified. The formulation needs either an operational definition of C*_k or a restatement in terms of the fairness properties that contribution evaluation is meant to satisfy.","section":"Section VIII-A, Eq. (11)"},{"comment":"The game-mechanism formulation is internally inconsistent. The text writes that attackers minimize their utility Uatt, but in a defense game an attacker should be maximizing its own utility subject to costs; the equilibrium condition in Eq. (13) is the standard Nash condition U_k(a*_k,a*_-k) ≥ U_k(a_k,a*_-k), which presupposes maximization. The subsequently 'unified' objective in Eq. (14), max Σ_k U_k & min Σ_k Cost_k, also leaves the trade-off between utility and cost unspecified, and it does not follow from the two preceding displayed objectives. Problem 8 therefore lacks a coherent mathematical statement, and Eq. (17) cannot inherit one from it.","section":"Section IX-A, Eqs. (12)–(14)"},{"comment":"Privacy loss is treated throughout as a deterministic function ℓ_{p,k}(w, D_k) of the model and local data. In real federated systems, privacy leakage depends on the training algorithm, the number of communication rounds, the adversary's prior knowledge, and the randomness of the mechanism. Because Eq. (17) and the per-problem formulations do not include these variables, the unified framework cannot faithfully represent privacy guarantees such as differential privacy or membership-inference risk. A formal treatment should either define ℓ_{p,k} through a specific privacy mechanism or state explicitly that the formulation abstracts away from mechanism-dependent leakage.","section":"Throughout, Eqs. (1), (4), (9), (16), (17)"}],"minor_comments":[{"comment":"The notation min_w F_{k=0,1,...,K} αℓu,k(ws)+... is ambiguous: F appears both as an aggregation mechanism and as part of the minimization expression, and the relationship between the outer weight α and the stated constraint α+β+γ=1 is unclear. Please rewrite with explicit aggregation and separate index notation.","section":"Section X-A, Eq. (15)"},{"comment":"In Eq. (11), k is used both for the client on the left-hand side and as the summation index on the right-hand side; the sum should run over a different index (e.g., Σ_{j=1}^K ||C(ws,D_j)-C*_j||²). The current notation also defines ℓc,k as a global sum rather than a per-client loss, which conflicts with the surrounding text.","section":"Section VIII-A, Eq. (11)"},{"comment":"In Eq. (5), the old-task arguments are written as {D^t_k}_{t=1}^T inside ℓu,k(wg, {D^t_k}_{t=1}^T), but the aggregation F then takes a comma-separated list of old-task and new-task losses. The syntax suggests two different function types and should be made uniform.","section":"Section IV-A, Eq. (5)"},{"comment":"The server-side loss is written as ℓu,s(ws|wk,Dp) in one line and as ℓu,k(ws|wk,Dp) in the adjacent displays; the subscripts should be standardized to avoid confusion between the server and client indices.","section":"Section VII-A, Settings 2 and 3"},{"comment":"There are numerous typos and inconsistent terms: 'mechanishm' in Eqs. (6), (8), and (9), 'receptively' in Eq. (15), 'Tand' in Section IV-A, 'How to to improve' in the Problem 10 heading, and 'foundation theory' versus 'foundational theory' are used interchangeably. A careful copy-edit is needed.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper is a survey; the main risk is that the unified equation in Section XII is presented as a formal contribution but currently fails on internal consistency. The authors should either repair Eq. (17) and actually carry out the claimed reductions, or explicitly downgrade it to a schematic summary. The per-problem literature review is valuable, and the problems are worth fixing within the scope of a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The punchline: this is a broad survey of ten FedFM problems with a reasonable taxonomy, but the paper's headline contribution—the unified equation (17) that supposedly gives \"the first mathematical perspective\" on all ten problems—does not hold up. The stress-test note is right: Eq. (17) drops the privacy thresholds δ_k that appear in Problems 2, 5, and 10, and its claimed reduction to Problem 4 is wrong, since Problem 4 minimizes a distance to a retrained model rather than a privacy loss. So the central formal claim is internally inconsistent, not just ambitious.\n\nWhat is genuinely useful: the ten-problem structure is well organized. Each section defines a problem, surveys existing methods, and lists challenges and potential solutions. The coverage is broad—219 references—and Table II's comparison with prior surveys is helpful. As a map of the literature for someone new to FedFMs, it has value.\n\nThe soft spots are mostly in the math. Eq. (17) minimizes over aggregation mechanism F and contribution evaluator C as if they were decision variables, but no feasible sets or regularity conditions are given. The \"no free lunch\" constraint in Eq. (1) is trivial, just stating that losses are positive. Eq. (11) relies on an oracle ground-truth contribution C*_k that is not observable. Eq. (15) has confusing notation for F. There are also many typos. None of this sinks the survey sections, but it does sink the paper's claim to be a rigorous mathematical treatment.\n\nWho should read it: students or researchers wanting a structured overview of open problems in FedFMs. They should use it as a starting point and not rely on the unification equation. The paper deserves a serious referee—the taxonomy is worth publishing—but it needs major revision. The authors should either fix Eq. (17) so that the reductions actually work, or drop the \"first mathematical perspective\" claim and present the taxonomy on its own terms.\n\nMy recommendation: send to peer review with a clear request for major revision on the formal parts. The survey content is salvageable; the math needs real work.","headline":"Useful FedFM taxonomy undermined by a unified equation that misrepresents its own per-problem formulations.","tokens_in":32526,"tokens_out":3663,"would_cite":false,"duration_ms":32886,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A single weighted-loss equation unifies the ten hard problems of federated foundation models, the paper argues.","keywords":["federated foundation models","federated learning","foundation models","multi-objective optimization","privacy-utility trade-off","machine unlearning","model watermarking","contribution evaluation"],"falsifier":"Train two FedFM systems to identical final model weights using different numbers of communication rounds or different aggregation randomness, then measure membership-inference or gradient-inversion leakage on the same target client. If leakage differs while the model weights and local data are the same, privacy loss is not a function of weights and data alone, and the unified objective cannot faithfully represent the privacy constraints it claims to encode.","tokens_in":31330,"feed_emoji":"🧩","tokens_out":4523,"duration_ms":42252,"temperature":0.7,"pith_summary":"This survey identifies ten open problems that block practical Federated Foundation Models (FedFMs), systems in which a large server-side foundation model and many small client-side domain models train together without sharing raw data. For each problem it states an optimization objective, reviews existing methods, and lists open challenges. The paper's central claim is that all ten problems can be written as one weighted objective over five losses — utility, efficiency, watermark, contribution, and privacy — subject to a no-free-lunch constraint. If that unification holds, algorithm design for FedFMs becomes a single multi-objective optimization problem, and progress on one problem can be compared against the others in common terms. The paper also asserts that this is the first mathematical treatment of the FedFM problem set.","feed_headline":"Ten federated-AI problems collapse into one loss equation","feed_subtitle":"A unified objective ties privacy, utility, efficiency, watermarking, and contribution into a single weighted trade-off.","key_machinery":"The unifying object is Eq. (17), a five-term weighted objective with a no-free-lunch constraint. Each term is a loss evaluated on the server model, client models, client actions, aggregation mechanism, contribution evaluation mechanism, the sequence of tasks, and graph data; the five weights are nonnegative and sum to 1. The machinery works by coefficient selection: setting one weight to 1 collapses the general objective into the specific formulation of one of the ten problems, so the same equation serves both as a common framework and as a generator of individual problem statements.","core_discovery":"The paper's central claim is Eq. (17): minimize, over the server model, client models, client actions, aggregation mechanism, and contribution evaluation mechanism, a weighted sum of five loss terms — utility loss, efficiency loss, watermark loss, contribution loss, and privacy loss — subject to the constraint that the sum of utility, privacy, and efficiency losses is strictly positive. By setting one weight to 1 and the others to 0, the equation is claimed to recover the individual problems: utility-only covers private data utilization, continual learning, Non-IID and graph data, bidirectional knowledge transfer, and game mechanisms; efficiency-only covers Problem 10; watermark-only covers Problem 9; contribution-only covers Problem 7; privacy-only covers unlearning; and the constraint encodes the foundational-theory trade-off of Problem 1. The paper presents this as the first unified mathematical formulation of the key problems of FedFMs.","pith_inferences":["If the unification is taken literally, an immediate test is whether real FedFM systems admit a single set of weights that reproduces observed trade-offs; the paper does not propose how to choose those weights, so weight selection becomes a key open design problem.","A consequence the authors leave implicit is that the same equation could serve as a benchmark generator: each problem's objective can be turned into a synthetic task suite where methods are scored by how much they reduce the corresponding term without inflating the others.","Because the privacy term is written as a function of weights and local data alone, the framework will likely need an extended version that conditions privacy loss on communication rounds, randomness, and adversary knowledge — a natural next step beyond the paper.","The five-loss structure connects naturally to multi-objective federated learning formalizations, and the no-free-lunch constraint could be sharpened into concrete information-theoretic bounds if privacy loss were replaced by a mutual-information or differential-privacy based quantity."],"forward_implications":["If Eq. (17) is accepted, FedFM research can be organized as a single multi-objective optimization, with each of the ten problems a special case obtained by choosing weights.","The no-free-lunch constraint makes explicit that utility, privacy, and efficiency cannot all be minimized simultaneously, so every FedFM design must state which trade-off it accepts.","Problems 2, 3, 5, 6, and 8 are claimed to share the utility-loss term, meaning methods for private-data utilization, continual learning, Non-IID and graph learning, bidirectional transfer, and game mechanisms can be evaluated within one utility objective.","The framework folds incentive design (contribution loss) and security (watermark and privacy losses) into the training objective itself, making fairness and protection part of optimization rather than post-hoc add-ons.","The per-problem formulations in Eqs. (1) through (16) give concrete objective functions that future FedFM algorithms can instantiate, compare, and test against."],"supporting_citations":[{"why":"Supplies FedAvg, the aggregation mechanism F used throughout the problem formulations in Eqs. (1), (4), (5), and (17).","marker":"[112]"},{"why":"Provides the no-free-lunch theorem for privacy and utility in federated learning, which grounds the constraint in Eq. (1) and the constraint in Eq. (17).","marker":"[108]"},{"why":"Frames multi-objective optimization for privacy, utility, and efficiency, the basis for the foundational-theory trade-off in Problem 1.","marker":"[119]"},{"why":"Gives efficient and fair data valuation for horizontal federated learning, the source for the contribution-evaluation objective in Problem 7 and the contribution loss in Eq. (17).","marker":"[74]"},{"why":"Introduces FedIPR, the watermarking scheme whose white-box and black-box formulations anchor the watermark loss in Problem 9 and Eq. (17).","marker":"[92]"},{"why":"Defines spatial-temporal catastrophic forgetting in federated continual learning, which shapes the continual-learning objective in Problem 3 and informs Eq. (5).","marker":"[109]"},{"why":"Surveys federated unlearning and frames the distance-to-retrained-model objective used in Eq. (7) for Problem 4.","marker":"[110]"},{"why":"Provides the broader survey of advances and open challenges in federated learning with foundation models, establishing the scope and terminology of FedFMs that the paper builds on.","marker":"[4]"}],"fun_headline_variants":["All ten FedFM challenges fit one equation","Unified loss equation tames ten FedFM problems","One equation to rule them: FedFM challenges","Single loss unifies ten federated model problems","One weighted sum captures all FedFM difficulties"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The unified objective treats privacy loss as a fixed function of the model weights and the local dataset, but in real federated systems privacy leakage also depends on the training algorithm, the number of communication rounds, the adversary's prior knowledge, and randomness — none of which appear as variables in Eq. (17).","fun_headline_variants_meta":{"raw":{"variants":["All ten FedFM challenges fit one equation","Unified loss equation tames ten FedFM problems","One equation to rule them: FedFM challenges","Single loss unifies ten federated model problems","One weighted sum captures all FedFM difficulties"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000533,"raw_usage":{"total_tokens":2588,"prompt_tokens":991,"completion_tokens":1597,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":607,"completion_tokens_details":{"reasoning_tokens":1527}},"tokens_in":607,"tokens_out":1597,"duration_ms":10920,"temperature":1.0,"reasoning_tokens":1527,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T20:07:47.533331+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train two FedFM systems to identical final model weights using different numbers of communication rounds or different aggregation randomness, then measure membership-inference or gradient-inversion leakage on the same target client. If leakage differs while the model weights and local data are the same, privacy loss is not a function of weights and data alone, and the unified objective cannot faithfully represent the privacy constraints it claims to encode.","supporting_citations":[],"review_version":1}