{"id":"cb63fc64-72cb-46a2-a6dc-1d6066c83753","arxiv_id":"2501.10945","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A comprehensive survey categorizes gradient-based multi-objective deep learning by output type, with theory, applications, datasets, libraries, and open challenges.","lead":"A survey organizes gradient-based multi-objective deep learning methods by what they produce: one balanced model, a finite set of trade-off models, or a continuous Pareto set of models. It also collects theory, applications, and software resources for multi-task learning, fairness, and LLM alignment.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Taxonomy category (ii) is internally inconsistent: Section 4.2 places many-objective cover methods (F4M/SoM/MosT) under 'finite set of Pareto-optimal solutions,' although these solve a different covering problem.","rationale":"The reader's conditional verdict is supported, but the load-bearing concern is not the 'first survey' claim; it is an internal inconsistency in the central taxonomy. Section 1.2 defines category (ii) as finite sets of Pareto-optimal solutions for user selection, while Section 4.2 includes F4M, SoM, and MosT, which are cover methods for the m>n regime. This mismatch means the survey's organizing principle does not cleanly partition the methods it presents, weakening the 'comprehensive map' contribution. The concern is concrete and testable by reading the cited papers and, if needed, running a small computational check. Strengths of the survey, such as the broad reference list, the resource section, and the theory summary, remain intact, and the taxonomy flaw is fixable by reclassifying these methods or broadening the category definition, so the conditional verdict should stand unchanged.","tokens_in":36769,"tokens_out":10059,"duration_ms":111525,"concrete_test":"Inspect the F4M paper (ICLR 2025) and the Many-Objective Cover Problem paper (PPSN 2024, ref [118]) and check whether their stated objective is to 'cover' all objectives with a small set of solutions rather than to approximate the Pareto set for user selection. Then verify the survey's Section 1.2 definition: if the output of F4M with n<m is not offered as a set of Pareto-optimal trade-offs for a user to choose from, the taxonomy assignment in Figure 2 is contradicted. A minimal computational check is to run F4M with n=2 on the tri-objective problem f(theta)=(theta1,theta2,1-theta1-theta2) with theta on the standard 2-simplex, and compare the two returned objective vectors to the Pareto front theta1+theta2=1; if the returned set is a pair of covering points rather than a sample of the front, the 'finite set of Pareto-optimal solutions' label is misleading.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The survey's central organizing principle, stated in Section 1.2, is that category (ii) contains algorithms that 'obtain a finite set of Pareto-optimal solutions that allows users to select from multiple options based on their specific needs.' Yet Section 4.2 itself acknowledges a 'converse case, where objectives outnumber solutions (m>n)' and includes F4M (Eq. 31), SoM (Eq. 32), and MosT, which are framed in the cited literature as 'cover' methods (e.g., 'Many-Objective Cover Problem: Discovering Few Solutions to Cover Many Objectives' [118]). These methods optimize an aggregation of per-objective minima over an n-element set, not a discrete approximation of the Pareto front for user selection; the resulting solutions are not a menu of trade-offs but a coverage set. The paper thus violates its own output-based taxonomy, and a reader using Figure 2 to select methods would misplace a distinct subarea. This is a substantive flaw in the survey's central claim to systematically categorize the field, not a cosmetic issue.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper is a survey of gradient-based multi-objective optimization (MOO) methods for deep learning. It proposes a taxonomy based on the type of output produced by an algorithm: (i) a single Pareto-optimal solution, (ii) a finite set of Pareto-optimal solutions, and (iii) an infinite/continuous Pareto set. For each category, it reviews representative algorithms, summarizes convergence and generalization theory, describes applications (reinforcement learning, Bayesian optimization, computer vision, neural architecture search, recommender systems, LLMs), and lists benchmark datasets and software libraries. The paper also claims to be the first survey focused specifically on gradient-based MOO methods in deep learning.","tokens_in":36994,"tokens_out":8185,"duration_ms":81821,"significance":"If the proposed taxonomy holds, this survey provides a useful organizing framework and a practical entry point for researchers, especially through its maintained GitHub resource list, its structured comparison of algorithms, and an accurate-in-main-lines account of convergence results. The output-based organizing principle is more discriminating than a chronological or method-by-method listing. The main weakness is that the taxonomy is not applied consistently: several algorithms in the ``finite set'' category do not produce a menu of trade-off solutions for user selection, which undermines the central claim of systematic categorization. The novelty claim also needs qualification in light of the acknowledged competing survey [141].","major_comments":[{"comment":"The taxonomy in Section 1.2 defines category (ii) as algorithms that ``obtain a finite set of Pareto-optimal solutions that allows users to select from multiple options based on their specific needs.'' However, Section 4.2 explicitly includes F4M (Eq. 31), SoM (Eq. 32), and MosT, which the text itself describes as addressing the ``converse case, where objectives outnumber solutions (m > n)'' and minimizing an aggregation of per-objective minima over an n-element set. These are coverage methods, not a discrete approximation of the Pareto front for user selection; their output is a coverage set. This contradicts the paper's central organizing principle and would mislead a reader who uses Figure 2 to locate methods. The authors should either move these methods to a separate category (e.g., ``coverage sets'') or explicitly revise the category definition and Figure 2 to accommodate this distinct subarea.","section":"Section 1.2 vs. Section 4.2 (Eqs. 31-32)"},{"comment":"The claim that this is ``the first survey paper focusing on the gradient-based MOO methods in deep learning'' is difficult to sustain because the paper itself acknowledges Peitz and Hotegni [141] as a survey of MOO algorithms for deep learning, albeit one that covers only a limited selection of gradient-based methods and omits theory and applications. The novelty claim should be qualified, for example as ``the first comprehensive survey covering algorithms, theory, and applications,'' and the comparison with [141] should be made more explicit so that the distinguishing features are clear.","section":"Section 1.1"}],"minor_comments":[{"comment":"The closed-form solution for IMTL-G as printed has a dimension mismatch: the expression λ_(2,...,m) = g_1^T U (D U^T)^{-1} yields a 1×d row vector, whereas λ_(2,...,m) should be an (m-1)-dimensional column vector. The correct form is likely (U^T D)^{-1} U^T g_1 (up to transposition); please correct the formula.","section":"Eq. (17)"},{"comment":"The approximation G^T d ≈ (1/η)[f_1^(k)-f_1^(k+1), ..., f_m^(k)-f_m^(k+1)]^T is terse; adding one sentence that invokes the first-order Taylor expansion f_i(θ-ηd) ≈ f_i(θ) - η g_i^T d would make the derivation much easier to follow.","section":"Section 3.2.3, Eq. (24)"},{"comment":"The ACM Reference Format line contains ``Received 20 February 2007; revised 12 March 2009; accepted 5 June 2009'' and ``© 2018,'' which are evidently template artifacts; these should be updated or removed before publication.","section":"Page 1, ACM Reference Format"},{"comment":"The sentence ``Haghtalab et al. are the first to show the sample complexity lower bound of Ω̃((v+m)/ε^2)'' is missing a citation for the specific work; please add the reference so the reader can locate the result.","section":"Section 6.2"},{"comment":"The notation g̃_α(ℓ(...)) is unclear because g̃_α is described as an MOO algorithm that produces a single solution, not a scalar loss function; please define how an algorithm is used inside the expectation, or rewrite the objective in a more standard operator form.","section":"Section 5.2, Eq. (37)"},{"comment":"The box ``Methods without Using Preference Vectors'' lists GradHV, MOO-SVGD, F4M, SoM, and MosT together; until the taxonomy issue raised in Major Comment 1 is resolved, consider adding a footnote or visual distinction to indicate that the last three are coverage-based methods with a different goal.","section":"Figure 2"}],"recommendation":"major_revision","confidential_remarks":"The reference list contains a noticeable concentration of work by the authors' own research groups (especially Chen, Lin, Zhang, Zhao, and Kwok). This is natural in a nascent field but should be monitored so the survey reads as a field-wide map rather than a portfolio of the authors' own contributions. The ``first survey'' claim is likely to attract criticism; a softened formulation would reduce that risk. The taxonomy inconsistency in Section 4.2 is substantive and should be resolved before publication; the rest of the survey is solid and provides a valuable resource."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: send this to review. The output-based taxonomy is genuinely useful, and the survey is mostly accurate; the 'first survey' claim is too strong and one taxonomy bucket is sloppy, but neither is load-bearing.\n\nWhat's actually new is the organization. Grouping methods by output type—single solution, finite set, infinite set—gives practitioners a real decision aid, and the paper backs it with clean descriptions of the algorithms, a convergence table, a generalization section, applications (including LLM alignment), and a resource list. The math summaries match the cited papers closely; I spot-checked MGDA, CAGrad, Nash-MTL, GradHV, and the stochastic convergence results and found no misstatements. The Section 4.2 inclusion of F4M, SoM, and MosT is the one real taxonomy problem. Those methods solve a coverage problem—choose n solutions so every objective is near some solution—not the 'menu of trade-offs for user selection' that Section 1.2 promises. So the stress-test is partially right. But it is not fatal: they do output finite sets, so the output-based taxonomy can absorb them if the definition is widened to 'finite set of solutions for diverse needs, including coverage.' Right now Figure 2 will mislead readers who assume category (ii) is homogeneous. That is a revision-level fix.\n\nThe other soft spots are minor. The claim to be the first survey is overstated: Peitz and Hotegni [141] covers overlapping ground, and the authors cite it. The template metadata is stale (J. ACM 2018, 35 pages), which is sloppy but irrelevant to content. The derivation aid around Eq. 24 is terse, and there are notation slips, e.g. in the GradHV equation the bracketed terms are described as 1x1 and 1xd but written as gradients. The citation pattern is self-heavy—many entries are from the authors' own groups—but the survey also covers enough outside work that I read this as field concentration rather than exclusion.\n\nWho is this for? Someone entering multi-task learning, fairness, or LLM alignment who wants a map of the method space and a starting point for convergence results. A specialist will not learn much they did not know, but the survey is a solid reference. I would cite it, and I would bring it to a reading group for students. Deserves a serious referee; the revision should fix the taxonomy definition and tone down the first-survey claim.","headline":"Useful, mostly accurate survey whose output-based taxonomy is a real contribution; the 'first survey' claim is overstated and the finite-set category needs cleanup, but it deserves peer review.","tokens_in":37503,"tokens_out":2673,"would_cite":true,"duration_ms":31147,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"One taxonomy sorts gradient-based multi-objective deep learning into three output types.","keywords":["Multi-Objective Optimization","Multi-Task Learning","Pareto Set Learning","Deep Learning","Gradient Balancing","Scalarization","Hypernetwork","LLM Alignment"],"falsifier":"Find a published gradient-based MOO algorithm that produces a single solution during training but at evaluation time can return an arbitrary preference-specified Pareto-optimal model without retraining or an explicit learned preference map; if no category in the three-way taxonomy accommodates it, the taxonomy is not exhaustive. For the first-survey claim, a counter-example would be any earlier survey that already covers gradient-based MOO algorithms together with their theory and applications across all three output regimes.","tokens_in":36576,"feed_emoji":"🧭","tokens_out":8608,"duration_ms":77141,"temperature":0.7,"pith_summary":"The paper sets out to prove that the scattered set of gradient-based methods for multi-objective deep learning is actually one field with a simple organizing question: what does the method give you back? It groups the literature into methods that return a single balanced Pareto-optimal model, methods that return a finite set of trade-off models, and methods that learn a continuous mapping from user preferences to models. Under this taxonomy it reviews convergence and generalization theory, applications from multi-task vision to large-language-model alignment, benchmark datasets and software libraries, and a list of open problems. A sympathetic reader would take away that the field has matured enough to be mapped, and that the map itself is the paper's central contribution.","feed_headline":"One taxonomy sorts multi-objective deep learning into three outputs","feed_subtitle":"The field's methods are mapped by what they return: one model, a finite menu, or any preference on demand.","key_machinery":"The central organizing device is the output-type taxonomy: single balanced solution versus finite set of Pareto-optimal solutions versus infinite continuous Pareto set learned as a preference-to-parameter map. The taxonomy carries the entire survey — every algorithm is placed into one of the three categories, the discussion of trade-offs (memory, cost, controllability) is organized around the categories, and the open problems are stated as gaps within or between them.","core_discovery":"The paper's central claim is that gradient-based multi-objective deep learning is best understood by the type of output an algorithm produces, not by its internal optimization mechanism. Single-solution methods, including loss balancing and gradient balancing, aim for one well-balanced Pareto-optimal model. Finite-set methods, whether they decompose the problem by preference vectors or directly optimize for diversity, produce a discrete approximation of the Pareto front. Infinite-set methods learn a preference-conditioned structure, such as a hypernetwork, a FiLM-conditioned network, or a parameter combination of base models, so that a user can generate a model for any preference on demand. The paper asserts this is the first survey devoted specifically to gradient-based multi-objective optimization (MOO) in deep learning, and it uses the three-way split to unify coverage of algorithms, theory, applications, resources, and open challenges.","pith_inferences":["The output-type taxonomy might apply equally well to gradient-free MOO, suggesting a unified comparison criterion across evolutionary and gradient-based methods: what kind of output the search returns, not how the search moves.","If infinite-set methods mature, the unit of deployment for a model could shift from a single checkpoint to a preference-conditioned artifact that is reconfigured at inference time, a change with serving and personalization consequences the paper only begins to explore through LLM alignment.","A directly testable prediction follows from the theory section: on the same benchmark (for example NYUv2 or QM9), single-solution gradient-balancing methods with Pareto-stationarity guarantees should show lower run-to-run variance in objective trade-offs than loss-balancing methods; the paper does not run this comparison.","The paper treats user preferences exclusively as a simplex vector; richer preference models such as constraints, rankings, or natural-language descriptions would likely require a new axis in the taxonomy, a direction the authors flag as open."],"forward_implications":["A practitioner facing a new multi-objective problem can first ask what output is needed — one balanced model, a menu of trade-offs, or on-demand preference control — and then select from the corresponding family of methods.","Gradient-balancing methods that converge to Pareto stationarity are available at stochastic rates comparable to single-objective optimization, so the practical bottleneck is computational cost, not lack of convergence guarantees.","Generalization theory for multi-objective deep learning is younger than convergence theory; near-optimal sample complexity for Tchebycheff scalarization exists, and the paper calls for extending stability-based and architecture-aware generalization analyses.","The paper's open-problem list implies that the field's next steps are reducing gradient-balancing cost, handling many objectives, distributed training, and carrying multi-objective methods beyond RLHF into other stages of the LLM lifecycle."],"supporting_citations":[{"why":"Supplies the MOO definitions — Pareto dominance, Pareto set, Pareto front, preference vectors — that the whole survey builds on.","marker":"[127]"},{"why":"Introduces the multiple-gradient descent algorithm whose Pareto-stationarity convergence is the basis of the gradient-balancing family.","marker":"[38]"},{"why":"Brings MGDA into deep multi-task learning and anchors the single-solution gradient-weighting family.","marker":"[157]"},{"why":"The overlapping survey the paper contrasts itself with when claiming to be the first focused on gradient-based MOO in deep learning.","marker":"[141]"},{"why":"Provides the smooth Tchebycheff scalarization used across the single-solution, finite-set, and infinite-set categories.","marker":"[102]"},{"why":"Defines the exact Pareto-optimal search used as a training solver for preference-aligned finite-set methods.","marker":"[123]"},{"why":"Introduces hypernetwork-based Pareto front learning, the archetypal infinite-set method.","marker":"[134]"},{"why":"Gives a geometric, hypervolume-based view of Pareto set learning that supports the theory of infinite-set methods.","marker":"[227]"},{"why":"Supplies the multi-task dense prediction survey that anchors the computer-vision application section.","marker":"[179]"}],"fun_headline_variants":["Multi-objective deep learning: one model, a menu, or any preference","How to balance many goals in deep learning: pick your output type","Three outputs define multi-objective deep learning methods","Gradient-based MOO in deep learning: a taxonomy by output"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The survey's usefulness rests on the assumption that categorizing methods by the type of output they produce is the right and exhaustive way to organize the field, so that every real algorithm fits cleanly into one of the three boxes.","fun_headline_variants_meta":{"raw":{"variants":["Multi-objective deep learning: one model, a menu, or any preference","How to balance many goals in deep learning: pick your output type","Three outputs define multi-objective deep learning methods","Gradient-based MOO in deep learning: a taxonomy by output"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000564,"raw_usage":{"total_tokens":2667,"prompt_tokens":932,"completion_tokens":1735,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":548,"completion_tokens_details":{"reasoning_tokens":1662}},"tokens_in":548,"tokens_out":1735,"duration_ms":11545,"temperature":1.0,"reasoning_tokens":1662,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T18:47:42.330884+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Find a published gradient-based MOO algorithm that produces a single solution during training but at evaluation time can return an arbitrary preference-specified Pareto-optimal model without retraining or an explicit learned preference map; if no category in the three-way taxonomy accommodates it, the taxonomy is not exhaustive. For the first-survey claim, a counter-example would be any earlier survey that already covers gradient-based MOO algorithms together with their theory and applications across all three output regimes.","supporting_citations":[{"cited_title":"Multi-objective Deep Learning: Taxonomy and Survey of the State of the Art","cited_arxiv_id":"2412.01566","evidence_quote":"The overlapping survey the paper contrasts itself with when claiming to be the first focused on gradient-based MOO in deep learning."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives a geometric, hypervolume-based view of Pareto set learning that supports the theory of infinite-set methods."}],"review_version":1}