{"id":"a1a6c977-06c0-45ff-8a07-2508390b5654","arxiv_id":"2412.01566","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A survey and taxonomy of multi-objective deep learning methods, organized by training algorithm and by when the decision maker selects a trade-off.","lead":"This preprint surveys methods for training deep networks against several competing objectives, from weighted losses to multi-gradient descent and evolutionary algorithms. It proposes a taxonomy based on when the decision maker chooses a trade-off and maps recent work across supervised, unsupervised, generative, and reinforcement learning.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Taxonomy Figure 6 omits adaptive/dynamic weighting, which Section IV-B2 treats as a distinct class; the central completeness claim is internally inconsistent.","rationale":"The paper's strongest claim is that the Section III-A taxonomy organizes multi-objective deep learning. The load-bearing assumption is completeness: every method class relevant to the field should map cleanly to Figure 6, and the survey text should not introduce categories that the taxonomy omits. The reader flagged completeness as the weakest assumption. The audit I propose confirms that the concern is real and internal: Section IV-B2 explicitly separates adaptive weighting from scalarization, yet Figure 6 has no node for adaptive weighting. This is not a matter of external literature coverage or a debatable boundary case; it is a direct mismatch between the taxonomy and the survey's own presentation. Because the taxonomy is the paper's primary contribution, this gap weakens the central claim unless the figure is revised or the survey text is adjusted to subsume adaptive weighting under an existing branch. The reader's CONDITIONAL verdict remains appropriate: the survey is still useful, but the taxonomy needs correction. I therefore keep the verdict unchanged rather than escalating to rejection, as the issue is fixable and does not invalidate the survey's descriptive value.","tokens_in":30380,"tokens_out":3785,"duration_ms":33836,"concrete_test":"Perform a mapping audit: for every method described in Sections IV-A through IV-C, assign it to a leaf of Figure 6 using the paper's own definitions. Determine whether the adaptive-weighting methods of Section IV-B2 ([31], [104], [121], [149]) fit any leaf; if not, the taxonomy requires either a new branch (e.g., 'adaptive/dynamic weighting') or a revised definition of scalarization that reconciles the Section IV-B2 statement.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central contribution is the taxonomy in Figure 6. Its two top-level branches ('decide-then-optimize' vs 'optimize-then-decide') and four algorithmic families (scalarization, MGDA, MOEA, continuation) are claimed to organize the state of the art. However, Section IV-B2 introduces 'adaptive weighting' as a separate category, stating: 'Unlike scalarization, adaptive weighting dynamically adjusts the importance of objectives throughout the training process.' Methods in that section (e.g., [121], [149], [104], [31]) are not fixed-weight scalarization, do not compute an entire Pareto front, and are not human-interactive, so none of the three top-level branches or four families in Figure 6 accommodates them without contradicting the text. This is an internal inconsistency, not just a coverage gap: the survey body contains a method class that the taxonomy cannot classify. The claimed gap ('no interactive methods') and the survey's conclusions are therefore less secure because the organizing principle is not exhaustive.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript presents a survey of multi-objective deep learning (MODL). Its central contribution is a taxonomy in Section III-A (Figure 6) that classifies methods by decision timing (decide-then-optimize, optimize-then-decide, interactive) and by algorithmic family (scalarization, MGDA, MOEA, continuation). The survey then reviews supervised, unsupervised/self-supervised, generative, and reinforcement learning settings, as well as neural architecture search and applications, and closes with a section on using deep learning for multi-objective optimization. The mathematical preliminaries (KKT conditions, Pareto critical sets, common descent directions, scalarization, continuation) are stated correctly, and the individual method descriptions are consistent with the cited literature in the spot-checks performed.","tokens_in":30514,"tokens_out":7062,"duration_ms":60450,"significance":"If the taxonomy is accepted as complete, the paper provides a useful organizing framework for a rapidly growing field, with broad coverage and technically sound preliminaries. The explicit algorithms for MGDA, MOEA, and continuation make the review self-contained, and the treatment of reinforcement learning as a special case is a reasonable structural choice. The survey's value hinges on the exhaustiveness of its taxonomy, which is stated as its main contribution; therefore, the omission of a method class that the survey itself treats as distinct weakens the central claim. This is fixable through a targeted revision rather than a fundamental rework.","major_comments":[{"comment":"The taxonomy in Figure 6 has no category that accommodates the 'adaptive weighting' class defined in Section IV-B2, creating an internal inconsistency in the survey's organizing contribution. In Section IV-B2 the authors write that adaptive weighting is 'unlike scalarization' and describe methods (e.g., [121], [149]) that dynamically adjust objective weights during training, do not compute an entire Pareto front, and are not interactive. These methods fit none of the three top-level branches or four algorithm families shown in Figure 6 without contradicting the text: they are not fixed-weight scalarization, not MGDA, not MOEA, not continuation, and not optimize-then-decide. The inconsistency is compounded by Section IV-A2, where FAMO [104] is described as a weighted sum with a dynamic weighting strategy, i.e., adaptive weighting is treated as a form of scalarization there. The authors should either subsume adaptive weighting under scalarization (and revise the 'Unlike scalarization' sentence and the placement of [104]) or add adaptive weighting explicitly to Figure 6 and discuss it in Section III-A; without this fix, the taxonomy cannot claim to organize the state of the art.","section":"IV-B2 and Figure 6"}],"minor_comments":[{"comment":"In the line 'the pupulation's fitness' and in the Algorithm 2 comment 'pupulation', 'pupulation' should be 'population'.","section":"Algorithm 2"},{"comment":"The word 'scalarizaiton' in the sentence 'A so-called conic scalarizaiton techinque' should be 'scalarization technique'.","section":"IV-A2"},{"comment":"The phrase 'realistic deep leraning applications' contains a typo; 'leraning' should be 'learning'.","section":"IV-A3"},{"comment":"In the introductory sentence, 'multi-objetive deep learning' should be 'multi-objective deep learning'.","section":"IV-E"},{"comment":"The phrase 'see also the survey [11] for for an extensive introduction' contains a duplicated 'for' and should be corrected.","section":"Remark 3"},{"comment":"In 'multi-task learning was used for for phoneme detection', the word 'for' is duplicated.","section":"IV-E2"}],"recommendation":"major_revision","confidential_remarks":"The only substantive concern is the adaptive-weighting taxonomy inconsistency; once the authors reconcile Figure 6 with Section IV-B2, the survey will be a solid contribution. The citation pattern appears balanced, and the authors' own prior work is cited where relevant without dominating the reference list."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: a useful survey with a clean presentation, but the taxonomy has a real seam around adaptive weighting, and the authors' own wording makes the seam visible. The stress-test note overstates the damage; the paper is still a solid map, but it needs a small fix.\n\nWhat's new: the survey organizes the deep learning side of multi-objective optimization by decision timing (decide-then-optimize, optimize-then-decide, interactive) and algorithmic family (scalarization, MGDA, MOEA, continuation). That framing is borrowed from MCDM and prior MORL surveys, but applying it to deep learning specifically is useful. Coverage spans supervised, unsupervised, generative, and reinforcement learning, and the authors honestly flag missing pieces: interactive deep MODL, non-convex deep benchmarks, and scarce continuation work. The mathematical preliminaries are stated correctly, and spot-checks of citations looked okay. That is real value for a newcomer.\n\nSoft spots: the completeness claim is not airtight. Section IV-B2 introduces 'adaptive weighting' with the phrase 'Unlike scalarization, adaptive weighting dynamically adjusts the importance of objectives...' but Figure 6 does not include adaptive weighting as a branch. Methods like FAMO are essentially scalarization with time-varying weights, so the taxonomy could absorb them without damage, but the text creates a contradiction that needs resolution. The stress-test note is too strong when it says the central claim collapses; the taxonomy still captures the main landscape, this is a seam rather than a collapse. Still, the authors should either fold adaptive weighting into the scalarization branch explicitly or add it as a sub-category and stop calling it 'unlike scalarization.' Second, the selection of included papers skews toward the authors' own group; many of the cited continuation and MGDA papers are their own. That is not disqualifying, but a stated inclusion criterion would be nice. Minor typos (e.g., 'pupulation') suggest light proofreading.\n\nWho it is for: grad students or ML researchers looking for a starting map of multi-objective deep learning. It is not a new-algorithms paper and should not be judged as one.\n\nRecommendation: send to peer review. It deserves a serious referee and likely a round of minor-to-major revisions to make the taxonomy consistent. I would probably read it and cite it as a survey reference if the revision tightens the adaptive weighting issue.","headline":"A useful map of multi-objective deep learning with a real but fixable seam in the taxonomy around adaptive weighting.","tokens_in":31031,"tokens_out":3045,"would_cite":true,"duration_ms":25587,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This survey argues that every multi-objective deep learning method can be classified by when the decision maker chooses a trade-off—before, after, or during optimization—and by whether it uses scalarization, multiple-gradient descent…","keywords":["multi-objective optimization","deep learning","taxonomy","Pareto front","multi-task learning","multi-objective reinforcement learning","scalarization","multiple-gradient descent algorithm"],"falsifier":"Find one published deep learning training method where a human decision maker changes preferences during optimization and the optimizer responds by moving along the Pareto front; that would fill the empty interactive branch and contradict the survey's main gap claim. Alternatively, find a method class that fits none of the two decision-timing branches and none of the four algorithm families, which would break the taxonomy's coverage.","tokens_in":30184,"feed_emoji":"🧭","tokens_out":4564,"duration_ms":42028,"temperature":0.7,"pith_summary":"The paper is a survey whose main contribution is a taxonomy of multi-objective deep learning methods. It organizes the field by two questions: when the decision maker chooses a trade-off, and which optimization strategy is used. The authors argue that nearly all current work falls into decide-then-optimize or optimize-then-decide, with only scalarization, multiple-gradient descent, evolutionary algorithms, and continuation as the relevant method families. They also claim that interactive multi-objective deep learning, where a human steers training by preferences, has not yet been explored. A sympathetic reader should care because this map makes a rapidly growing and fragmented literature navigable, and it identifies a concrete open research gap.","feed_headline":"Decision time splits all multi-objective deep learning methods","feed_subtitle":"A taxonomy sorts every training method by when the trade-off is chosen, and exposes an empty branch: interactive training.","key_machinery":"The load-bearing object is the taxonomy itself, presented in Figure 6 as a tree whose top decisions are decision timing and algorithm family, with dashed ellipses for hybrid extensions. It is backed by the Karush-Kuhn-Tucker condition $\\sum_{k=1}^K \\alpha_k^* \\nabla L_k(\\theta^*) = 0$, which defines Pareto-critical points and is the shared basis for MGDA and continuation methods. The taxonomy does the work of the paper: it partitions recent methods from supervised, unsupervised, generative, and reinforcement learning, reveals that MGDA is the practical workhorse for deep learning because it reuses standard gradient machinery, and makes the absence of interactive methods visible as an empty branch.","core_discovery":"The central claim is taxonomic: the state of the art in multi-objective deep learning can be organized by a two-level classification. The top level asks whether the decision maker chooses a trade-off before optimization, after optimization, or interactively during optimization. The second level asks which algorithmic family is used: scalarization, multiple-gradient descent algorithms (MGDAs), multi-objective evolutionary algorithms (MOEAs), or continuation methods. Under this taxonomy, the paper reports that the deep learning community has mostly chosen the decide-then-optimize branch, with MGDA as the dominant gradient-based tool, while optimize-then-decide methods split between MOEAs and weight-varying scalarization. The authors further claim that the interactive branch is empty: methods that alternate between optimization and a decision maker's preferences exist in classical multi-objective optimization but have not been transferred to deep learning training. They also observe that over-parameterization can collapse the Pareto front for some multi-task problems, so conflicts between objectives may vanish on very large networks.","pith_inferences":["An immediate testable next step would be to couple a preference-elicitation loop with continuation methods or MGDAs, giving a decision maker control over the output of deep multi-objective training while keeping gradient efficiency.","The over-parameterization observation implies that benchmark selection matters: on sufficiently large shared-parameter networks, multi-task objectives may not conflict, so studies of multi-objective deep learning should report network scale and verify that the Pareto front does not collapse.","The taxonomy's decision-timing axis could be extended to large language model alignment, where multiple reward models or judges are already combined by weighted sums; interactive preference steering during fine-tuning would fill the empty branch in a commercially relevant setting.","A deep learning benchmark with a known non-convex Pareto front would be valuable because the paper notes that no such benchmark currently exists, and it would separate scalarization methods that fail on non-convex frontiers from MGDA and other approaches that handle them."],"forward_implications":["A researcher encountering a new multi-objective deep learning method can classify it by two questions, and the classification immediately indicates which optimization machinery is relevant.","Gradient-based approaches, especially MGDA, are the practical choice for deep multi-objective training, while evolutionary algorithms remain mainly useful for architecture search and hybrid settings rather than direct network training.","The interactive branch is a concrete research gap: no published method currently lets a decision maker steer deep learning training by expressing preferences among objectives during optimization.","Because over-parameterized networks can resolve objective conflicts, multi-task deep learning on very large shared networks may not require full Pareto-front machinery.","Reinforcement learning is a special case of the taxonomy: single-policy algorithms correspond to decide-then-optimize, and multiple-policy algorithms correspond to optimize-then-decide, with the population being value functions rather than policies."],"supporting_citations":[{"why":"Supplies the definitions of Pareto optimality, Pareto sets, and the KKT condition that frame the entire survey.","marker":"[120]"},{"why":"Provides the second foundational introduction to multi-criteria optimization and scalarization concepts used in the preliminaries.","marker":"[43]"},{"why":"Defines the multiple-gradient descent algorithm whose common descent direction anchors the gradient-based branch of the taxonomy.","marker":"[42]"},{"why":"Introduces NSGA-II, the standard multi-objective evolutionary algorithm used to illustrate the MOEA family in the taxonomy.","marker":"[35]"},{"why":"Gives the prominent deep-learning application of MGDA to multi-task learning, which the survey treats as a reference point for supervised multi-objective training.","marker":"[161]"},{"why":"Provides an earlier taxonomy of multi-objective reinforcement learning that the paper builds on and contrasts with its own classification.","marker":"[106]"},{"why":"Offers a modern practical guide to multi-objective reinforcement learning that supports the paper's adaptation of the taxonomy to the RL setting.","marker":"[70]"},{"why":"Presents a decomposition-based MORL taxonomy whose decision-timing structure shares similarities with the survey's classification.","marker":"[48]"}],"fun_headline_variants":["Interactive multi-objective training: an empty branch in deep learning","Deep learning's missing piece: interactive multi-objective optimization","Taxonomy reveals deep learning skips interactive trade-offs","Over-parameterization may erase objective conflicts in deep nets","Multi-objective deep learning: all choices made before training"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The taxonomy's completeness is the load-bearing premise: if a major method class is missing, or if the surveyed papers are significantly misclassified, the survey's map and its claimed open gap for interactive methods collapse.","fun_headline_variants_meta":{"raw":{"variants":["Interactive multi-objective training: an empty branch in deep learning","Deep learning's missing piece: interactive multi-objective optimization","Taxonomy reveals deep learning skips interactive trade-offs","Over-parameterization may erase objective conflicts in deep nets","Multi-objective deep learning: all choices made before training"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000152,"raw_usage":{"total_tokens":1172,"prompt_tokens":884,"completion_tokens":288,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":500,"completion_tokens_details":{"reasoning_tokens":207}},"tokens_in":500,"tokens_out":288,"duration_ms":2903,"temperature":1.0,"reasoning_tokens":207,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:16:49.745053+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Find one published deep learning training method where a human decision maker changes preferences during optimization and the optimizer responds by moving along the Pareto front; that would fill the empty interactive branch and contradict the survey's main gap claim. Alternatively, find a method class that fits none of the two decision-timing branches and none of the four algorithm families, which would break the taxonomy's coverage.","supporting_citations":[],"review_version":1}