{"id":"2c2ed5b1-2d49-4c0d-b84d-08a97c9eac4e","arxiv_id":"2412.12092","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"NMT optimizes lower-priority tasks under a Lagrangian penalty that keeps the primary task loss near its pre-trained optimum, with no manual balancing weights in the loss combination.","lead":"This paper proposes NMT, a multi-task learning framework that treats the primary task as a constraint and optimizes secondary tasks with a Lagrangian multiplier, claiming no weight tuning is needed. The authors report AUC gains on two public recommendation datasets and online gains in Taobao search, but the no-tuning claim conflicts with their own grid-search appendix.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's strong-duality claim is unsupported: Proposition 2's proof derives a two-sided Lipschitz bound but then relies on a one-sided inequality that does not follow, so the guarantee that the primary task is preserved is not established.","rationale":"The paper's central claim is that NMT guarantees the primary task is preserved while secondary tasks improve, with no tuning. That guarantee rests on Theorem 1's strong-duality result, which depends critically on the approximate convexity of the perturbation function established in Proposition 2. The proof of Proposition 2 is the load-bearing step, and it contains a clear logical gap: a two-sided Lipschitz bound is used to justify a one-sided inequality that does not follow. Without that inequality, the convex combination of optimal parameters need not be feasible, so the perturbation function need not be convex even approximately. Moreover, the epsilon in Assumption 2 is never quantified or linked to gradient descent, so the key hypothesis is not verifiable in practice. The online training curves show the primary loss exceeding its optimal value, which is consistent with the constraint being violated. These observations directly undermine the theoretical guarantee and, with it, the 'primary task preserved' promise. The reader's weakest-assumption point is essentially the same concern, and the REJECT verdict remains appropriate.","tokens_in":12963,"tokens_out":10854,"duration_ms":92983,"concrete_test":"Independently re-derive the appendix proof of Proposition 2: starting from bound (16), determine whether inequality (18) is a logical consequence. Concretely, evaluate (16) and (18) for a nonconvex primary loss with two separated wells, e.g., f1(θ)=(θ^2-1)^2 with θ*=-1 and r small, choosing ξ1 and ξ2 so that θ*(ξ1) and θ*(ξ2) lie in different wells; if the convex combination θt violates the constraint, (18) is false and Proposition 2 and Theorem 1 are unproven.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central promise of NMT is that secondary tasks can be optimized while the primary task's optimal value is preserved, with no task-weight tuning. The only mechanism offered for this guarantee is Theorem 1, which asserts strong duality between (CO) and (DO) under Assumptions 1 and 2. Theorem 1 depends on Proposition 2, whose appendix proof is invalid. From the Lipschitz bound (16), the authors obtain an absolute error bound |fi(θt) - (t fi(θ*(ξ1)) + (1-t) fi(θ*(ξ2)))| ≤ Lε/2. They then assert the one-sided inequality fi(θt) ≤ t fi(θ*(ξ1)) + (1-t) fi(θ*(ξ2)) in (18) and use it to claim that the convex combination θt is feasible. An absolute-value bound does not imply that one-sided inequality; the error can be positive, so θt may violate the constraint. Feasibility would require positive slack in the constraints, t s1 + (1-t) s2 ≥ Lε/2, but the proof never establishes any slack, and when constraints are active the slack is zero. Thus the approximate convexity of P(ξ) is not established, and even approximate convexity would not give the exact convexity needed for strong duality. Additionally, Assumption 2 essentially assumes that the feasible region is small (parameter differences bounded by ε) and is never connected to the actual optimization path of a deep network. The online training plot in Fig.3.b shows positive deviations of Lpay(θ)-Lpay(θ*) up to about 0.06, indicating that the constraint is violated during training. Therefore, the theoretical guarantee that the primary task is preserved is unsupported, and the empirical successes are not explained by the stated theory.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes No More Tuning (NMT), a multi-task learning framework that formulates prioritized optimization as a sequence of constrained problems: the highest-priority task is minimized first, and each lower-priority task is then minimized subject to constraints that keep higher-priority losses at or below their previously obtained values. The constraints are handled by a Lagrangian, with gradient descent on parameters and gradient ascent on Lagrange multipliers. The paper claims that NMT requires no additional hyper-parameters, provides theoretical guarantees through a strong-duality result (Theorem 1) under Assumptions 1 and 2, and reports offline gains on TikTok and QK-Video and online gains in a Taobao search A/B test. The central promise is that secondary tasks can be improved while the primary task's performance is preserved, without manual loss-weight tuning.","tokens_in":13312,"tokens_out":4638,"duration_ms":41530,"significance":"If the theoretical claim and experiments were reliable, the paper would offer a practically attractive way to enforce priority in MTL without manual loss weighting, and the Taobao deployment gives useful evidence of applicability. The framework's integration with existing MTL architectures is a strength, and the A/B test reports positive business metrics for the primary task. However, the central theoretical argument is not established, the no-tuning claim is contradicted by the algorithm's own inputs and the appendix's grid search, and the experimental reporting lacks error bars. As it stands, the paper's main contribution is not supported.","major_comments":[{"comment":"The proof of Proposition 2 is invalid at the step from Eq. (16) to Eq. (18). Eq. (16) is an absolute-value bound |f_i(θ_t) - (t f_i(θ*(ξ1)) + (1-t) f_i(θ*(ξ2)))| ≤ Lε/2, but Eq. (18) asserts the one-sided inequality f_i(θ_t) ≤ t f_i(θ*(ξ1)) + (1-t) f_i(θ*(ξ2)) without establishing that the error term has the correct sign. An absolute bound does not rule out f_i(θ_t) exceeding the convex combination, and when the constraints at ξ1 and ξ2 are active there is no slack to absorb a positive error. Therefore the feasibility of θ_t is not proved, the approximate convexity of P(ξ) is not established, and Theorem 1, which rests on Proposition 2, is unsupported.","section":"Appendix, Proof of Proposition 2"},{"comment":"Assumption 2 assumes that parameter differences inside the feasible region are bounded by an arbitrarily small ε and that this makes the perturbation function approximately convex, but the paper never connects ε to the optimization path of the deep networks used in the experiments. Figure 3b shows L_pay(θ) - L_pay(θ*) fluctuating up to about 0.06 above zero during training, i.e., the primary-task constraint is violated; the stated guarantee therefore does not apply to the reported training runs. The small-ε condition effectively assumes the conclusion that optimization stays within a region where the primary loss is preserved.","section":"Theoretical Analysis, Assumption 2 and Figure 3"},{"comment":"The central claim that NMT requires 'no additional hyper-parameters' is contradicted by Algorithm 1's inputs: the parameter learning rate η, the multiplier learning rate τ, the initial multiplier λ_init, and per-stage convergence thresholds are all user-set quantities, and the tolerance r_i appears in the theoretical problem (CO). The appendix confirms this by describing a grid search over nine combinations of η and τ. At minimum the claim must be weakened to 'no task-weight hyper-parameters', and the sensitivity to these remaining hyper-parameters must be reported.","section":"Algorithm 1 and Appendix: Training Detail"},{"comment":"The implemented algorithm uses the re-scaled loss L = 1/(1+Σ_j λ_j) (f_k(θ) + Σ_j λ_j (f_j(θ) - f_j(θ*_j))), while the theoretical analysis in Eq. (7) and Theorem 1 applies to the unscaled Lagrangian. The scaling factor is a function of λ and therefore changes the gradient dynamics of both θ and λ; the paper does not show that the strong-duality argument, even if valid, transfers to the re-scaled objective actually optimized in all experiments. This is a gap between theory and implementation.","section":"Re-scaling method and Eq. (7)"},{"comment":"The experimental claims are reported as point estimates without error bars, standard deviations, or significance tests, and the online A/B results are percentage deltas without confidence intervals. Given that many reported differences are on the order of 0.01%–0.5%, it is not possible to judge whether NMT improves the primary task or whether the differences are within run-to-run variation. This is particularly important for the paper's central claim that the primary task is never compromised.","section":"Tables 1–3 and Figure 2"}],"minor_comments":[{"comment":"The text says task prioritization is embedded in 'equality constraints', but the formulation throughout the paper uses inequality constraints; the terminology should be made consistent.","section":"Introduction, bullet 'No Need for Parameter Adjustments'"},{"comment":"The notation L_pay(θ*) is used in Eq. (11) without defining θ* in the context of the online experiment; it should be clarified whether θ* is the single-task pay-optimized parameter vector.","section":"Online Experiments, Eq. (11)"},{"comment":"Assumption 1 ('free from over-fitting') is not a formal mathematical condition; it should be replaced with a precise statement about the relationship between training loss and target performance.","section":"Theoretical Analysis, Assumption 1"},{"comment":"There are typographical and formatting issues, including 'involxd' in the Introduction, 'Y uan' in the references, and a duplicated parenthesis in the statement of problem (CO).","section":"Throughout"},{"comment":"The main text says the detailed proof of Proposition 2 will be provided in 'the Appendix of the extended version, which has been published on arXiv'; a journal or conference submission should be self-contained, and the proof should appear in the manuscript itself.","section":"Main text and Appendix"},{"comment":"The figure caption does not fully explain whether the colored lines are Pareto fronts or simple scans of two weights, nor whether the NMT points are obtained with the same computational budget; this should be clarified.","section":"Figure 2"}],"recommendation":"reject","confidential_remarks":"The paper's strongest asset is the industrial A/B test, but the theoretical guarantee is central to the 'no tuning' selling point and that guarantee is not proved. The no-tuning claim will also draw scrutiny because Algorithm 1 has multiple hyper-parameters and the appendix reports a grid search over them. I would not consider acceptance until the proof of Proposition 2 is corrected, the small-ε condition is connected to the actual optimization path, and the experiments include uncertainty quantification. The manuscript may be resubmitted after major revision addressing these points."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The algorithm is simple and the industrial data point is interesting, but the two central promises—no tuning and a theoretical guarantee—are unsupported by the paper's own content.\n\nWhat is new: the NMT wrapper, which first trains the primary task and then sequentially trains secondary tasks under a Lagrangian constraint, re-scaling the combined loss to avoid explosion. It is easy to implement and plugs into any gradient-descent MTL architecture. That is a legitimate, if modest, algorithmic contribution. The Taobao A/B test comparing pay+relevance against weight adjustment is a genuine industrial data point; the +0.26% order volume with +0.72% relevance, versus -0.35% order volume with weight adjustment, is worth paying attention to.\n\nThe soft spots are proportionally large. The proof of Proposition 2 in the appendix derives an absolute error bound (16) and then asserts a one-sided inequality (18). An absolute-value bound does not give you the sign, so the approximate convexity of the perturbation function is not established, and Theorem 1 collapses. Assumption 2 essentially assumes the parameters stay in a small epsilon-neighborhood—the conclusion—and Figure 3b shows the pay loss drifting above its optimum by up to 0.06 during training. So the 'primary task is preserved' guarantee is not delivered.\n\nThe 'No More Tuning' claim is directly contradicted by the appendix grid search over learning rates for both θ and λ. That is not a quibble; it is the main selling point. The experiments also lack error bars, and hyperparameters are selected on the evaluation datasets.\n\nWhat remains is a heuristic that may work well in practice. The public-dataset gains are consistent but small, and without error bars they are hard to weigh. A serious referee could help the authors strip the overclaims and either fix the proof or present the method as a heuristic. I would send this to review and expect major revision, not desk reject, because the industrial A/B result and the simplicity of the method give the community something to test. But in its current form, I would not accept.","headline":"A simple Lagrangian wrapper with an interesting industrial A/B result, but the no-tuning claim is refuted by the paper's own appendix and the duality proof has a sign error, so it overclaims as written.","tokens_in":13873,"tokens_out":4591,"would_cite":false,"duration_ms":38338,"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":"A Lagrange multiplier method gives multi-task models explicit priorities without loss-weight tuning.","keywords":["multi-task learning","task prioritization","constrained optimization","Lagrangian multiplier methods","strong duality","hyperparameter tuning","recommendation systems","learning to rank"],"falsifier":"On a two-task problem where the secondary task has its global optimum far from the primary task's optimum in parameter space, run the NMT algorithm and check the constraint f1(θ) ≤ f1(θ*) at the final iterate. If it is violated, the strong-duality guarantee does not hold for that setting; a more direct check is to numerically evaluate the perturbation function P(ξ) near the optimum and test whether it is convex over the ξ range the optimizer actually visits.","tokens_in":12722,"feed_emoji":"🎯","tokens_out":5946,"duration_ms":47947,"temperature":0.7,"pith_summary":"This paper claims that task priorities in multi-task learning can be enforced without manually balancing loss weights. Instead of weighting tasks, it solves a sequence of constrained problems: each lower-priority task is minimized subject to the constraint that all higher-priority tasks keep their previously achieved loss. The method, named NMT, uses a Lagrangian differential multiplier update (gradient descent on parameters, gradient ascent on multipliers) and the paper proves that, under a Lipschitz-smoothness and small-parameter-change assumption, strong duality holds, so the constrained problem can be solved as an unconstrained min-max problem. If correct, any gradient-based multi-task model can be given a priority ordering with no task-weight hyperparameters, and the primary task's performance is protected by construction. Experiments on two recommendation datasets and an online Taobao search A/B test report gains in the priority task and secondary tasks without the degradation seen with loss weighting.","feed_headline":"No More Tuning: A way to prioritize tasks without loss weights","feed_subtitle":"A primal-dual algorithm optimizes lower-priority tasks while holding the primary task at its best loss, with a duality guarantee.","key_machinery":"The mechanism is the Lagrangian differential multiplier method applied to each constrained subproblem. For stage k, the objective is L(θ, λ) = f_k(θ) + Σ_{j<k} λ_j (f_j(θ) − f_j(θ*)), with the constraint f_j(θ) ≤ f_j(θ*) for higher-priority tasks. Parameters θ are updated by gradient descent and multipliers λ by gradient ascent, giving an unconstrained min-max problem whose strong-duality guarantee (Theorem 1) rests on the perturbation function P(ξ) being approximately convex when the optimal parameters for neighboring tolerances are within ε. A re-scaling of the loss by 1/(1+Σλ_j) keeps the combined objective a normalized convex combination, avoiding loss explosion.","core_discovery":"The central discovery claim is that a lexicographic-style prioritization in multi-task learning—minimize f1, then minimize f2 subject to f1 ≤ f1*, then f3 subject to f1 ≤ f1* and f2 ≤ f2*, and so on—can be carried out by a simple primal-dual gradient algorithm with no task-weight hyperparameters. The constrained problem (CO) and its Lagrangian dual (DO) are shown to have strong duality under two assumptions: no overfitting, and Lipschitz continuity of the losses together with bounded parameter changes inside the feasible region. Consequently, the dual ascent-descent updates converge to a solution that respects the priority constraints, which the authors verify by integrating NMT with Shared-Bottom, OMoE, MMoE, PLE, and FAMO on TikTok and QK-Video, and by deploying it in Taobao search with order volume as top priority, where it improved lower-priority metrics without sacrificing order volume.","pith_inferences":["\"No tuning\" should be read as \"no task-balancing weights\": the method still relies on hyperparameters such as learning rates for θ and λ, and the appendix reports a grid search over those; a fair test of the no-tuning claim would fix these values without search.","The strong-duality proof is local: it requires ε small, meaning the secondary task's best parameters must stay near the primary optimum; for tasks whose optima are far apart in parameter space, the constraint may be violated in practice even if the algorithm appears to converge.","The sequential scheme suggests a natural extension to safety-constrained learning: any constraint that can be expressed as an inequality on a loss can be enforced by the same multiplier update, potentially replacing reward shaping in constrained reinforcement learning.","An empirical check of the theory would be to compute the perturbation function P(ξ) on a small two-task network; if it is non-convex across the ξ range the optimizer visits, Theorem 1's premise is not satisfied."],"forward_implications":["Any gradient-descent multi-task model can be given an explicit priority order by running NMT in stages; no task-balancing weights are introduced.","The tuning cost for m-task priorities drops from exponential grid search O(p^m) to m sequential optimization steps.","The primary task's loss is kept at its pre-optimized level (within the assumptions) while secondary tasks improve.","In the Taobao search deployment, adding lower-priority tasks under NMT improved their business metrics without the drop in order volume observed when the same tasks were added by loss weighting."],"supporting_citations":[{"why":"Supplies the Lagrangian multiplier method for constrained differential optimization that the NMT primal-dual updates are built on.","marker":"(Platt and Barr 1987)"},{"why":"Provides the strong-duality conditions (Slater's condition and convexity of the perturbation function) that Theorem 1 invokes.","marker":"(Rockafellar 1970)"},{"why":"Provides the MMoE architecture used as an experimental backbone to demonstrate NMT integration.","marker":"(Ma et al. 2018)"},{"why":"Provides the PLE architecture used as another experimental backbone for the offline multi-task recommendation evaluations.","marker":"(Tang et al. 2020)"},{"why":"Supplies the Tenrec benchmark datasets (TikTok and QK-Video) used for the offline performance comparisons.","marker":"(Yuan et al. 2023)"},{"why":"Provides FAMO, the gradient-based multi-task method that NMT is integrated with in the appendix to show compatibility.","marker":"(Liu et al. 2024)"}],"fun_headline_variants":["Prioritized multi-task learning without task-weight tuning","Primal-dual MTL: set priorities, skip manual tuning","Automatic task priorities via Lagrangian multiplier methods","Hold top task, free the rest: MTL without loss weights","Lexicographic MTL from a simple primal-dual algorithm"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The guarantee that the primary task stays uncompromised assumes the optimal parameters for lower-priority tasks remain within a very small neighborhood of the primary task's optimum, where the loss functions are smooth and the constraint set is effectively convex.","fun_headline_variants_meta":{"raw":{"variants":["Prioritized multi-task learning without task-weight tuning","Primal-dual MTL: set priorities, skip manual tuning","Automatic task priorities via Lagrangian multiplier methods","Hold top task, free the rest: MTL without loss weights","Lexicographic MTL from a simple primal-dual algorithm"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000334,"raw_usage":{"total_tokens":1879,"prompt_tokens":993,"completion_tokens":886,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":609,"completion_tokens_details":{"reasoning_tokens":818}},"tokens_in":609,"tokens_out":886,"duration_ms":7939,"temperature":1.0,"reasoning_tokens":818,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:16:40.774663+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a two-task problem where the secondary task has its global optimum far from the primary task's optimum in parameter space, run the NMT algorithm and check the constraint f1(θ) ≤ f1(θ*) at the final iterate. If it is violated, the strong-duality guarantee does not hold for that setting; a more direct check is to numerically evaluate the perturbation function P(ξ) near the optimum and test whether it is convex over the ξ range the optimizer actually visits.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Lagrangian multiplier method for constrained differential optimization that the NMT primal-dual updates are built on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the strong-duality conditions (Slater's condition and convexity of the perturbation function) that Theorem 1 invokes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the PLE architecture used as another experimental backbone for the offline multi-task recommendation evaluations."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides FAMO, the gradient-based multi-task method that NMT is integrated with in the appendix to show compatibility."}],"review_version":1}