acceptodds
Under review as a conference paper at ICLR 2027

Two-Time-Scale Weight Adaptation and Evolutionary Genome Search for Pareto Front Coverage

Abstract

Many problems in science and engineering require minimising several conflicting objectives at once. In settings where gradients of the objectives are available, e.g. multi-task learning, preference alignment, and adjoint-based design, scalarisation reduces the problem to a gradient-friendly family of weighted sub-problems. However, it shifts the problem onto weight selection: the weight distribution that spreads solutions evenly over the Pareto front is governed by the solution map that is non-injective and often discontinuous. We propose TWARD (Two-time-scale Weight Adaptation and Repulsion Descent), a two-time-scale method for gradient-based Multi-objective optimisation. On the fast scale, the decision variables descend an entropic reference-point scalarisation, a smoothed Tchebycheff achievement function whose reference point is parametrised by the log of the weights, with a parameter which controls coverage. On the slow scale, entropic mirror descent moves the weights on the simplex under a repulsion measured between solutions in objective space, promoting spreading of solutions on the Pareto front, translated into a weight update in closed form by inverting the scalarisation rather than the solution map. Where a single gradient flow is confined to one connected piece of the front, an evolutionary outer loop varies the -dimensional scalarisation genome . At a matched evaluation budget, we see superior performance compared to established baseline approaches over a wide range of synthetic and real-world tasks.

Primary area:
optimization
Keywords:
multi-objective optimization, Pareto front coverage, scalarization, two-time-scale optimization, mirror descent, evolutionary search

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_REQTOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes TWARD, a two-time-scale weight-adaptation method for gradient-based multi-objective optimization. The fast time scale runs gradient descent on a soft achievement scalarization function (ASF) in the decision variable x, while the slow time scale updates the weight vector λ on the simplex via entropic mirror descent. The weight update combines an objective-space RBF repulsion term, which pushes nearby particles apart on the front, with a centering term that keeps the adaptive reference point near the ray from z. A key construction is that, because λ enters the ASF as logλ, a desired displacement in objective space can be mapped to a displacement in logλ in closed form, avoiding the solution-map Jacobian. An outer evolutionary loop searches the genome (λ,τ) with convergence and spacing tests, aiming to cross disconnected or degenerate fronts.

Critical strengths
  1. The coupling between the scalarization and the weight dynamics is clean and non-trivial. Making λ an additive shift of the reference point, so that logλ moves linearly with objective-space displacement, is a neat trick that avoids inverting the solution map and gives the slow-scale update a closed-form geometric meaning. This is the main technical idea that distinguishes the method from standard preference-based gradient solvers.

  2. The two-time-scale separation is well motivated and the fixed-λ analysis is honest about its scope. Lemma 1 and Theorem 1 quantify how the smoothness constant and the iteration complexity depend on τ, and the paper uses this to justify the temperature ladder rather than treating τ as an arbitrary knob. The observation that a single τ is sufficient for representability but not for reachability by descent is a useful clarification.

  3. The empirical study is broad and the budget matching is taken seriously. The paper evaluates on 33 problems across four benchmark families, compares against a reasonable set of gradient and aggregation baselines, and matches the Jacobian-evaluation budget between TWARD and LibMOON-style methods. The appendix also reports tuning-seed versus held-out-seed behavior and a spacing-radius sweep, which is more transparency than many papers provide.

Critical weaknesses
  1. The theory does not cover the algorithm that is actually run. Theorem 1 analyzes gradient descent on S_{λ,τ} at fixed λ, and the proof relies on the fast scale treating λ as frozen. The actual method alternates fast x-updates with slow λ-updates and an outer evolutionary loop with restarts, convergence tests, and spacing tests. There is no convergence guarantee for the coupled two-time-scale dynamics, no bound on the approximation error introduced by the repulsion-to-weight inversion on curved fronts, and no argument that the outer loop finds a well-spread subset of the front. The paper acknowledges that the inversion is exact only on locally flat patches and approximate on curved fronts, but it does not quantify how this approximation affects coverage or stationarity.

  2. The method is heavily hyperparameter-dependent, and the paper tunes per problem.

  3. The empirical comparison is not uniformly favorable, and the paper’s “matches or improves” framing is too generous. On WFG1, WFG3, WFG4, and several ZCAT instances, TWARD is not the best method; MOO-SVGD or HVGrad is better on WFG1/WFG3, and the paper’s own tables show gaps. On DTLZ1 and DTLZ3, the paper explains that all gradient methods struggle because of the multimodal distance term, which is fair, but this also means the benchmark does not discriminate between methods there. The main comparison lacks statistical tests.

  4. The Jacobian budget is matched, but TWARD uses many short trajectories with restarts while the baselines use long trajectories per ray; this is a legitimate budget match, but it changes the optimization dynamics and should be discussed. The paper also does not compare against a simple baseline that runs the same soft-ASF scalarization with fixed weights and the same number of restarts, which would isolate the contribution of the weight adaptation from the contribution of restarts and repulsion.

Questions for authors
  1. Can you provide a convergence or approximation analysis for the coupled two-time-scale dynamics, or at least bound the error introduced by the repulsion-to-weight inversion δ log λ=−δF/(τs) on curved fronts? How large is this error in practice, and does it explain any of the coverage gaps on WFG1/WFG3?

  2. Table 5 shows per-problem tuning of seven hyperparameters. What is the recommended default configuration for a new problem, and how sensitive is the method to these choices? Can you report a sensitivity study or a tuning protocol that does not require per-instance search?

  3. The comparison matches Jacobian evaluations, but TWARD uses G=4 generations with short inner trajectories and restarts, while the baselines use one long run per ray. Can you add an ablation with fixed weights but the same restart/repulsion budget, to isolate the effect of weight adaptation from the effect of restarts?

  4. On WFG1/WFG3 and some ZCAT instances, TWARD is not the best method. Can you clarify the scope of the claim “matches or improves” and identify the geometries where the method is expected to be weaker?

2 comments

We thank the reviewer for the thorough review and for highlighting the log⁡λ\log\lambda shift of the reference point and the transparency of the empirical study. W1 and Q1 are answered in C1 and C2 of the general response at the top of this page, and we summarise them here.

W1. The theory does not cover the algorithm that is actually run. Theorem 1 analyzes gradient descent on S(λ,τ)S_{(\lambda,\tau)} at fixed λ\lambda, and the proof relies on the fast scale treating λ\lambda as frozen. The actual method alternates fast xx-updates with slow λ\lambda-updates and an outer evolutionary loop with restarts, convergence tests, and spacing tests. There is no convergence guarantee for the coupled two-time-scale dynamics, no bound on the approximation error introduced by the repulsion-to-weight inversion on curved fronts, and no argument that the outer loop finds a well-spread subset of the front. The paper acknowledges that the inversion is exact only on locally flat patches and approximate on curved fronts, but it does not quantify how this approximation affects coverage or stationarity.

Thank you. We agree that Theorem 1 alone does not cover the algorithm that is run. C1 bounds the error of the lift (Eq. 14) on curved fronts, and Q1 reports its size. C2 shows that, with gradient descent, the particles follow the moving weights and the weights follow the bilevel loop, both up to O(ηλ/ηx)O(\eta_\lambda/\eta_x). C2 is local, does not show that the weights converge, and does not cover Adam, the running scale ss or the outer loop. We have stated this in the paper. [Revised: Sec. 4.2 and 6, App. E.1 and E.2]. The error of the lift changes where a point ends on the front, not whether it is stationary, since every archived point passes the convergence test. Its effect on coverage is measured in Q3, where freezing the weights worsens 30 of 33 problems. For the outer loop we have no coverage guarantee. The spacing test keeps archived points at least rr apart, which bounds the size of the archive. Since a share of every generation is drawn at random, any region farther than rr from the archive receives a new point with positive probability in each generation. Replacements within rr can sometimes remove a point that dominates its replacement. We measured this in Rcsj Q2, and it is rare. [Revised: Sec. 4.3, App. E.9]

W2. The method is heavily hyperparameter-dependent, and the paper tunes per problem.

Thank you. We agree this is a limitation and answer it in Q2.

W3. The empirical comparison is not uniformly favorable, and the paper's "matches or improves" framing is too generous. On WFG1, WFG3, WFG4, and several ZCAT instances, TWARD is not the best method; MOO-SVGD or HVGrad is better on WFG1/WFG3, and the paper's own tables show gaps. On DTLZ1 and DTLZ3, the paper explains that all gradient methods struggle because of the multimodal distance term, which is fair, but this also means the benchmark does not discriminate between methods there. The main comparison lacks statistical tests.

Thank you. We agree that "matches or improves" is too generous and have restated the scope as in Q4. [Revised: Sec. 1 and 6] On WFG1 and WFG3, both TWARD and TWARD-S are ahead of every baseline in IGD and HV in our tables (e.g. WFG3 IGD: TWARD-S 0.022, TWARD 0.035, best baseline HVGrad 0.043). The gap there is between our two variants, which Q4 explains. We agree that DTLZ1/DTLZ3 do not discriminate between gradient methods, which is why they are reported separately (Appendix B.7). Two-sided Wilcoxon rank-sum tests over 5 seeds, between our best variant and the best baseline on each problem, show ours significantly better at the 5% level on 25 of 33 problems and worse on two (ZDT3 vs. HVGrad, WFG4 vs. MOO-SVGD). [Revised: Sec. 5.4, App. E.7, Table 13]

W4. The Jacobian budget is matched, but TWARD uses many short trajectories with restarts while the baselines use long trajectories per ray; this is a legitimate budget match, but it changes the optimization dynamics and should be discussed. The paper also does not compare against a simple baseline that runs the same soft-ASF scalarization with fixed weights and the same number of restarts, which would isolate the contribution of the weight adaptation from the contribution of restarts and repulsion.

Thank you. This choice is deliberate and we have added a discussion of it [Revised: Sec. 5.4, App. E.4]. Appendix D.10 shows that, at fixed cardinality, more steps per trajectory mostly refine regions already found, while restarts with new genomes reach new ones. Where convergence rather than finding new regions is the bottleneck, the long-trajectory variant TWARD-S is better (5nES Q6). The control baseline is in Q3 below (Sec. 5.1, Table 1, App. E.3).

Q1. Can you provide a convergence or approximation analysis for the coupled two-time-scale dynamics, or at least bound the error introduced by the repulsion-to-weight inversion δlog⁡λ=−δF/(τs)\delta\log\lambda=-\delta F/(\tau s) on curved fronts? How large is this error in practice, and does it explain any of the coverage gaps on WFG1/WFG3?

Thank you. C2 bounds the lag of the coupled dynamics and Fig. 87 of the revised paper measures it. C1 bounds the error of the lift by τα/(1−τα)\tau\alpha/(1-\tau\alpha), and in practice the median cosine between the actual and predicted move is 0.97 (Fig. 86 of the revised paper). The lift does not explain the WFG1/WFG3 gaps. On WFG3 it is as accurate as elsewhere (cosine 1.00, length ratio 1.00), and on WFG1 convergence is the bottleneck, since no TWARD-S particle passes the convergence test even before the update. [Revised: Sec. 4.2, App. E.1 and E.2]

Q2. Table 5 6 now shows per-instance tuning of seven hyperparameters. What is the recommended default configuration for a new problem, and how sensitive is the method to these choices? Can you report a sensitivity study or a tuning protocol that does not require per-instance search?

Thank you. As a default we take, for each hyperparameter, the median of its tuned values over the 33 problems: τ∈{0.2,0.7}\tau\in\lbrace 0.2,0.7\rbrace, γ=7.5\gamma=7.5, h=0.35h=0.35, ηx=0.05\eta_x=0.05, ηλ=0.005\eta_\lambda=0.005, κ=0.5\kappa=0.5, ρ=0.01\rho=0.01. Leaving any single problem out changes only the smallest temperature (0.2 to 0.25), so the default is not fitted to the problem it is tested on. Run unchanged on all 33 problems, it raises the median IGD by 18% relative to per-problem tuning, stays within 25% on 20 problems, and still beats the best tuned baseline on 18. It transfers to WFG (+26%) and ZCAT (+15%) but not to ZDT3 and DTLZ2/5/6, whose tuned step sizes and temperatures are much smaller (on DTLZ6 no run converges). One-at-a-time sweeps on DTLZ2, DTLZ5, DTLZ7, WFG7 and ZCAT2 show that γ\gamma is forgiving (at most 40% within a factor of 2), while hh and the τ\tau ladder cost far more when too small than when too large. Halving them costs up to 92% and 5×5\times, doubling them at most 32% and 36% (ηλ\eta_\lambda is in 5nES Q5). We therefore recommend starting from the default and tuning only the τ\tau ladder and the step sizes, erring towards larger τ\tau and hh. Both hh and the spacing radius rr act on normalised objectives, so they need no rescaling across problems. The per-problem values are in Table 6 (Table 5 of the submission). [Revised: Sec. 5.4, App. E.5]

Q3. The comparison matches Jacobian evaluations, but TWARD uses G=4G=4 generations with short inner trajectories and restarts, while the baselines use one long run per ray. Can you add an ablation with fixed weights but the same restart/repulsion budget, to isolate the effect of weight adaptation from the effect of restarts?

Thank you. Yes, see the ablation table in our reply to 5nES Q3. With the same restarts, archive and budget, fixed weights are 30% worse in median IGD (24% with random genomes), while keeping the weight adaptation with random genomes costs only 1.7%. [Revised: Sec. 5.1, Table 1, App. E.3]

Q4. On WFG1/WFG3 and some ZCAT instances, TWARD is not the best method. Can you clarify the scope of the claim "matches or improves" and identify the geometries where the method is expected to be weaker?

Thank you. We have restated the scope as follows [Revised: Sec. 1, 5.4, App. E.7]. Our methods have the best IGD on 27 of 33 problems at the matched budget, with the largest margins on disconnected, degenerate and curved-Pareto-set geometries (DTLZ5–7, ZCAT). They are not best on six. On ZDT3, WFG2, ZCAT14 and ZCAT17, HVGrad, which ascends the hypervolume directly, is better. On WFG4 MOO-SVGD is better, and on WFG5 Agg-PBI is ahead by 0.001. On degenerate connected fronts (WFG3, DTLZ6) the solver alone (TWARD-S) is better than TWARD, because convergence rather than finding new regions is the bottleneck. On the strongly biased WFG1 the two are within 0.3%.

Reviewer_5nESOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TWARD, a hybrid gradient-based and evolutionary framework for improving Pareto front coverage in multi-objective optimization. The method employs a two-time-scale optimization mechanism in which decision variables are updated through gradient descent on an entropic reference-point scalarization, while preference weights are adapted more slowly using entropic mirror descent driven by objective-space repulsion. A key component is the derivation of a closed-form relationship between changes in objective values and logarithmic preference weights, which avoids explicitly computing the Jacobian of the solution map. To further improve coverage of disconnected and irregular Pareto fronts, an evolutionary outer loop searches over scalarization parameters and maintains an archive of approximately stationary and well-separated solutions. The authors provide theoretical results concerning scalarization smoothness and convergence to approximate Pareto-stationary points, and evaluate the proposed method on ZDT, DTLZ, WFG, ZCAT, and a multi-objective fairness-learning task. The experimental results demonstrate competitive performance in terms of inverted generational distance and hypervolume compared with several gradient-based multi-objective optimization methods.

Critical strengths
  1. The paper addresses a relevant challenge in gradient-based multi-objective optimization, namely obtaining diverse solutions across complex Pareto fronts. The combination of gradient-based optimization and evolutionary search is well motivated, particularly for disconnected and degenerate Pareto fronts.

  2. The proposed objective-space repulsion mechanism and its translation into preference-weight updates are interesting. The approach avoids explicitly computing the solution-map Jacobian and provides a potentially efficient mechanism for improving solution diversity.

  3. The experimental evaluation is relatively comprehensive, covering several established benchmark families and comparing against multiple gradient-based multi-objective optimization approaches. The reported results demonstrate competitive performance, particularly on challenging DTLZ and ZCAT problems.

Critical weaknesses
  1. The theoretical justification of the proposed weight adaptation mechanism requires further clarification. The closed-form relationship between objective displacement and logarithmic preference weights is derived under a fixed front-normal assumption and is acknowledged to be approximate on curved fronts. However, the actual algorithm operates on potentially highly nonlinear, disconnected, and degenerate Pareto fronts. It remains unclear under what conditions the proposed weight updates produce the intended objective-space movement and whether they consistently improve Pareto front coverage.

  2. The convergence analysis primarily concerns gradient descent on the scalarized objective with fixed preference weights. It does not establish convergence or stability of the complete two-time-scale dynamics, where the decision variables and preference weights are updated simultaneously. Furthermore, the theoretical analysis assumes gradient descent, whereas the implementation employs Adam. The relationship between the theoretical results and the practical algorithm should be clarified.

  3. The individual contributions of the proposed components are not sufficiently established. Although the paper investigates the effect of removing repulsion, a more systematic ablation study is needed to distinguish the contributions of adaptive weights, two-time-scale updates, temperature adaptation, evolutionary genome search, and archive selection. In particular, the improvements obtained by the evolutionary outer loop may partly result from repeated random restarts and selection mechanisms rather than genome evolution itself.

  4. The fairness of the experimental comparisons requires further investigation. The methods are compared using matched Jacobian evaluation budgets, which is a reasonable starting point. However, the proposed approach introduces additional computations for pairwise repulsion, weight adaptation, evolutionary operations, and archive management. These computational costs are not fully reflected in the evaluation budget. Moreover, comparisons against evolutionary multi-objective optimization algorithms would help establish the practical benefits of the proposed hybrid framework.

  5. The scalability and practical applicability of the method remain insufficiently demonstrated. Most benchmark problems involve relatively low-dimensional decision spaces and two or three objectives. Although the UCI Adult experiment provides an additional application, it uses a relatively small neural network. Further experiments on higher-dimensional problems, more objectives, and larger learning models would strengthen the claimed advantages of combining gradient-based optimization with evolutionary search.

Questions for authors
  1. The derivation of the preference-weight update relies on holding the Pareto front normal fixed. How accurate is this approximation on highly curved or degenerate Pareto fronts? Can the authors provide a quantitative analysis of the discrepancy between the desired objective-space displacement and the actual displacement induced by the weight update?

  2. The theoretical analysis establishes approximate stationarity for the fast-scale gradient descent under fixed preferences. Can the authors provide additional convergence or stability analysis for the coupled two-time-scale dynamics, particularly when the preference weights are continuously adapted through repulsion?

  3. The proposed framework combines several mechanisms, including adaptive weights, repulsion, temperature variation, evolutionary search, and archive selection. Could the authors provide a systematic ablation study to quantify the individual contributions of these components? In particular, how does the evolutionary outer loop compare with an equivalent number of independent random restarts using the same inner solver and archive-selection strategy?

  4. The experiments compare the methods primarily using matched Jacobian evaluation budgets. Could the authors additionally report computational time, memory consumption, and scaling behavior with respect to the number of particles, objectives, and decision variables? Comparisons with representative evolutionary multi-objective optimization algorithms would also help clarify the practical advantages of the proposed framework.

  5. The theoretical analysis assumes gradient descent, while the practical implementation uses Adam. How sensitive is the proposed method to the relative learning rates of the fast and slow updates, and how is the required separation of time scales maintained during optimization?

  6. The results in Tables 1 and 2 indicate that TWARD does not consistently outperform TWARD-S or other baselines across all performance metrics. For example, TWARD-S performs better on DTLZ6 and WFG3 in terms of IGD, while HVGrad achieves better results on ZDT3. Could the authors further analyze when the evolutionary outer loop is beneficial and when it may negatively affect convergence or coverage?

  7. The proposed method is evaluated mainly on problems with two or three objectives. Could the authors investigate its scalability to many-objective optimization problems and larger-scale machine learning tasks, where the computational overhead of population-based repulsion and evolutionary search may become more significant?

4 comments

We thank the reviewer for the detailed review and for finding the combination of gradient-based optimisation and evolutionary search well motivated. W1, W2, Q1, Q2 and the Adam part of Q5 are answered in C1 and C2 of the general response at the top of this page, and we summarise them here.

W1. The theoretical justification of the proposed weight adaptation mechanism requires further clarification. The closed-form relationship between objective displacement and logarithmic preference weights is derived under a fixed front-normal assumption and is acknowledged to be approximate on curved fronts. However, the actual algorithm operates on potentially highly nonlinear, disconnected, and degenerate Pareto fronts. It remains unclear under what conditions the proposed weight updates produce the intended objective-space movement and whether they consistently improve Pareto front coverage.

Thank you. C1 gives the conditions. The move is exact on flat patches, has relative error at most τα/(1−τα)\tau\alpha/(1-\tau\alpha) on curved ones, happens only along the front on degenerate ones, and cannot move past the edge of a component of a disconnected front, which is left to the outer loop. Q1 measures the error. On coverage, freezing the weights worsens 30 of 33 problems (+30% median IGD, Q3), so the updates improve coverage consistently, but not on every problem. [Revised: Sec. 4.2 and 5.1, Table 1, App. E.1 and E.3]

W2. The convergence analysis primarily concerns gradient descent on the scalarized objective with fixed preference weights. It does not establish convergence or stability of the complete two-time-scale dynamics, where the decision variables and preference weights are updated simultaneously. Furthermore, the theoretical analysis assumes gradient descent, whereas the implementation employs Adam. The relationship between the theoretical results and the practical algorithm should be clarified.

Thank you, we agree this needed clarifying. C2 analyses the coupled dynamics and measures them with gradient descent and with Adam. The analysis is local, holds for gradient descent, and does not show that the weights converge. For Adam we rely on the measurements and on the convergence test, which checks the tolerance of Theorem 1 on every returned point. We have stated these limits in the paper. [Revised: Sec. 4.2 and 6, App. E.2]

W3. The individual contributions of the proposed components are not sufficiently established. Although the paper investigates the effect of removing repulsion, a more systematic ablation study is needed to distinguish the contributions of adaptive weights, two-time-scale updates, temperature adaptation, evolutionary genome search, and archive selection. In particular, the improvements obtained by the evolutionary outer loop may partly result from repeated random restarts and selection mechanisms rather than genome evolution itself.

Thank you. We answer this in Q3.

W4. The fairness of the experimental comparisons requires further investigation. The methods are compared using matched Jacobian evaluation budgets, which is a reasonable starting point. However, the proposed approach introduces additional computations for pairwise repulsion, weight adaptation, evolutionary operations, and archive management. These computational costs are not fully reflected in the evaluation budget. Moreover, comparisons against evolutionary multi-objective optimization algorithms would help establish the practical benefits of the proposed hybrid framework.

Thank you. The comparison assumes that evaluating ff and its Jacobian is the main cost per step, and we have stated this explicitly [Revised: Sec. 5, App. E.6]. The additional operations do not evaluate ff or its Jacobian, and their cost does not depend on the decision dimension dd. The Jacobian count therefore reflects the total cost better as Jacobians become more expensive. The comparison with EMO algorithms is in Rcsj W3.

Q1. The derivation of the preference-weight update relies on holding the Pareto front normal fixed. How accurate is this approximation on highly curved or degenerate Pareto fronts? Can the authors provide a quantitative analysis of the discrepancy between the desired objective-space displacement and the actual displacement induced by the weight update?

Thank you. On curved fronts, C1 shows that whenever τα<1\tau\alpha<1 the relative error is at most τα/(1−τα)\tau\alpha/(1-\tau\alpha), with α\alpha the rate at which the normal turns, and that it does not depend on the step length. On degenerate fronts the normal is not unique, and only the part of the requested move along the front happens. Since the repulsion acts between neighbours on the same curve, the requested move is mostly along it, and on DTLZ5/6 the particles move almost exactly in the predicted direction. To measure the error, we apply the update with step size 0.01 to converged particles of all 33 problems and compare the actual move vs intended move (Table 12 and Fig. 86 of the revised paper). Over all problems the median cosine is 0.97.

Problem   Front               cosine   length ratio   not moving
DTLZ2     concave             0.998    1.12           11%
DTLZ5     degenerate          1.000    1.09           21%
DTLZ6     degenerate          1.000    1.11           0%
DTLZ7     disconnected        0.971    0.34           35%
WFG3      degenerate          1.000    1.00           22%
WFG7      concave             0.995    1.07           10%
ZCAT3     curved Pareto set   1.000    1.00           6%

Particles that do not move sit at the edge of a component, past which they cannot move (e.g. DTLZ7), or belong to problems tuned with τ≥0.5\tau\ge0.5. [Revised: Sec. 4.2, App. E.1, Table 12, Fig. 86]

Q2. The theoretical analysis establishes approximate stationarity for the fast-scale gradient descent under fixed preferences. Can the authors provide additional convergence or stability analysis for the coupled two-time-scale dynamics, particularly when the preference weights are continuously adapted through repulsion?

Thank you. C2 covers this case. The bound there holds for any weight motion of bounded speed, and the repulsion step is bounded because Eq. 11 divides it by the mean force norm rˉ\bar r. Adapting the weights continuously therefore does not destabilise the fast scale. After a transient the particles stay within O(ηλ/ηx)O(\eta_\lambda/\eta_x) of the solutions for their current weights, and the weights follow the bilevel loop to the same order. Fig. 87 of the revised paper measures this lag and shows it growing with ηλ/ηx\eta_\lambda/\eta_x, as predicted. The analysis is local (it needs a non-degenerate inner minimiser), does not cover jumps between components of a disconnected front, which the outer loop handles, and does not show that the weights converge. [Revised: Sec. 4.2, Eq. 16, App. E.2, Fig. 87]

Q3. The proposed framework combines several mechanisms, including adaptive weights, repulsion, temperature variation, evolutionary search, and archive selection. Could the authors provide a systematic ablation study to quantify the individual contributions of these components? In particular, how does the evolutionary outer loop compare with an equivalent number of independent random restarts using the same inner solver and archive-selection strategy?

Table 1 and Fig. 88 of the revised paper change one component at a time on all 33 problems at the same cost (160k Jacobians).

Variant                         median IGD   worse >5%   better >5%   p
All genomes random (restarts)   +1.7%        13          6            0.10
Frozen weights                  +30%         30          1            4e-8
Frozen weights and random       +24%         26          2            6e-8
No repulsion                    +43%         29          2            5e-8
Single tau                      +0.3%        8           3            0.11
No archive tests                -7.1%        3           20           0.001

Adaptive weights and repulsion give most of the gain, and frozen weights are the slow limit of the two time scales (the fast limit is in Q5). A single τ\tau is neutral on most problems but fails on some, e.g. ZDT3 (0.010→0.2560.010\to0.256). Removing the convergence and spacing tests archives every child. IGD then falls by 7%, mostly because more points are returned and IGD does not penalise points that fail the stationarity test. On a few problems it is worse (WFG1, ZCAT17). For the outer loop, each generation already draws a share of its children at random and breeds the rest from the archive. Drawing all of them at random changes little, while the same random restarts with frozen weights are far worse, so the gain comes from the weight adaptation. [Revised: Sec. 5.1, Table 1, App. E.3, Fig. 88]

Q4. The experiments compare the methods primarily using matched Jacobian evaluation budgets. Could the authors additionally report computational time, memory consumption, and scaling behavior with respect to the number of particles, objectives, and decision variables? Comparisons with representative evolutionary multi-objective optimization algorithms would also help clarify the practical advantages of the proposed framework.

Thank you, we have added the following. On the same machine a full TWARD run takes a median of 19 s per problem (8–47 s), against 10 to 19 minutes for the LibMOON baselines. The implementations differ (JAX vs. PyTorch), so this is a practical cost, not an algorithmic speed-up. Fig. 90 of the revised paper varies NN, mm and dd one at a time on DTLZ2. The time per step grows sublinearly in NN up to our N=100N=100 and faster beyond, since the repulsion costs O(N2m)O(N^2m). It grows about linearly in dd once d≳1000d\gtrsim1000, where the Jacobian dominates, and only mildly in mm. Memory is O(Nmd+N2)O(Nmd+N^2) per step for the Jacobians and the kernel, plus O(∣A∣(d+m))O(|\mathcal A|(d+m)) for the archive. The comparison with EMO algorithms is in Rcsj W3. [Revised: App. E.6, Fig. 90]

Q5. The theoretical analysis assumes gradient descent, while the practical implementation uses Adam. How sensitive is the proposed method to the relative learning rates of the fast and slow updates, and how is the required separation of time scales maintained during optimization?

Thank you. Gradient descent vs. Adam is answered in C2 (Fig. 87 of the revised paper). For the sensitivity, we scaled ηλ\eta_\lambda from 0.1×0.1\times to 10×10\times its tuned value with ηx\eta_x fixed (Fig. 89 of the revised paper). Within a factor of 3, IGD changes by at most 16% on DTLZ7 and ZCAT2 and by up to 2.5×2.5\times on DTLZ2, DTLZ5 and WFG7. Further out, the weights either barely move (IGD approaches the frozen-weight result) or move too fast for xx to converge (on DTLZ2, 61% instead of 97% of runs pass the convergence test). The separation is kept by fixing the ratio ηλ/ηx\eta_\lambda/\eta_x per problem (Table 6) and by dividing the repulsion by rˉ\bar r (Eq. 11), so the weight step does not grow with it. Runs in which xx has not converged are discarded by the convergence test. [Revised: Sec. 5.4, App. E.5, Fig. 89]

Q6. The results in Tables 1 and 2 indicate that TWARD does not consistently outperform TWARD-S or other baselines across all performance metrics. For example, TWARD-S performs better on DTLZ6 and WFG3 in terms of IGD, while HVGrad achieves better results on ZDT3. Could the authors further analyze when the evolutionary outer loop is beneficial and when it may negatively affect convergence or coverage?

Thank you. The outer loop helps when finding new regions is the bottleneck: on curved Pareto sets (ZCAT) and, given enough generations, on disconnected fronts (on DTLZ7 the IGD falls from 0.090 at G=4G=4 to 0.036 at G=25G=25, with all four patches covered, Section 5.3). It helps less when the front is connected and degenerate or strongly biased (DTLZ6, WFG3). There convergence is the bottleneck, and the longer trajectories of TWARD-S refine better than many short restarts. HVGrad ascends the hypervolume directly and is strongest on the two-objective ZDT3. At the matched budget TWARD improves on TWARD-S on 27 of 33 problems. We have added this discussion. [Revised: Sec. 5.4, App. E.4]

W5. The scalability and practical applicability of the method remain insufficiently demonstrated. Most benchmark problems involve relatively low-dimensional decision spaces and two or three objectives. Although the UCI Adult experiment provides an additional application, it uses a relatively small neural network. Further experiments on higher-dimensional problems, more objectives, and larger learning models would strengthen the claimed advantages of combining gradient-based optimization with evolutionary search.

Thank you. We have added a larger experiment with more objectives. We train N=15N=15 LoRA adapters (rank 16, d≈2.2×106d\approx2.2\times10^6 parameters each) on a frozen Qwen2.5-0.5B-Instruct model, with DPO-style preference losses on five HelpSteer2 attributes (m=5m=5) and mini-batch gradients. All methods get the same budget of 72,000 sequence backward passes and are tuned on a separate split. We run TWARD-S, since we do not expect disconnected fronts here. Test results over 3 seeds:

Method             HV (higher better)   In box (of 15)   Spacing CV (lower better)
TWARD-S            5.61 ± 0.78          13.3             0.53
soft-Tche          4.69 ± 0.19           8.7             2.88
Agg-LS             4.38 ± 0.57           9.0             3.05
soft-Tche + rep.   4.23 ± 0.45           7.3             2.25
Agg-Tche           4.02 ± 0.94           8.0             2.94
EPO                2.72 ± 0.49          11.7             1.72
HVGrad             1.91 ± 0.07          15.0             1.61

TWARD-S has the highest mean test hypervolume. A paired bootstrap over the test pairs shows it significantly better than Agg-LS, EPO and HVGrad, but with three seeds it does not separate TWARD-S from the Tchebycheff baselines. The clearer difference is where the solutions lie. The scalarisation baselines place several adapters far outside the reference box, on extreme verbosity or complexity, and the answers these adapters generate are judged much less helpful than those of the base model. No TWARD-S adapter loses more than 0.21 points of judged helpfulness, while 6 to 11 of the 45 adapters of each scalarisation baseline lose more than 0.5. Scaling in NN, mm and dd is in Q4. [Revised: Sec. 5.5, Table 4, App. F]

Q7. The proposed method is evaluated mainly on problems with two or three objectives. Could the authors investigate its scalability to many-objective optimization problems and larger-scale machine learning tasks, where the computational overhead of population-based repulsion and evolutionary search may become more significant?

Thank you. The language-model experiment in W5 has m=5m=5 objectives and d≈2.2×106d\approx2.2\times10^6 parameters per particle. It uses TWARD-S, so it includes the repulsion but not the outer loop. There the weight step and the repulsion use only stored losses and no backward pass, so they cost nothing in the matched budget. In the benchmark solver the repulsion costs O(N2m)O(N^2m) per step, the weight update O(Nm)O(Nm), breeding O(m)O(m) per child and the archive tests O(∣A∣m)O(\vert\mathcal A\vert m) per child. None of these evaluates ff or its Jacobian, and none depends on dd. How this cost grows with NN, mm and dd is measured in Q4. [Revised: Sec. 5.5, Table 4, App. E.6 and F.4, Fig. 90]

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

This paper proposes a two-time-scaled weight adaptation technique, along with an outer evolutionary search, for multi-objective optimization with gradient information available.

Critical strengths
  1. An interesting research problem - indeed, maintaining a set of uniformly-distributed solutions for multi-objective problems is very useful but challenging, particuarly for scalarization methods that need a number of weights and a scalarization function to decompose the problem into a few single-objective problems.

  2. The idea is intuitive - adapting weights can naturally lead to better diversity.

  3. Compared to existing techniques, the proposed method perform well (but I have some reservation on this, which can be seen in the weakness section).

Critical weaknesses
  1. It is not very clear to me why you would like to use such a two-time-scaled weight adaptation method. I can clearly see what they are, but struggle to understand why this is needed, particularly in Abstract. Why a fast scale and a slow scale would help the optimization process? To sum up, the motivation of using the method is not very clear.

  2. Also, in Introduction, it is not very clear why an objective-space repulsion that pushes nearby particles’ weights apart on the simplex can ensure better uniformity.

  3. In the evolutionary multi-objective optimization area, weight adaptation techniques can make decomposition-based MOEAs work very well, even for problems with disconnected, degenerate, and inverted simplex Pareto fronts, see e.g., https://arxiv.org/pdf/2502.16481. However, in your experiment, it is clear to see that your method doesn't perform particularly well on those problems, such as DTLZ7. I am wondering what advantage of your algorithm, compared to those weight adaptation methods in the EMO area.

Questions for authors

I would like the authors to respond to the three weaknesses above, and also to answer the following questions.

  1. I wonder whether it is possible to incorporate weight adaptation methods in the EMO area into your method for even better uniformity.

  2. I noticed you use an archive to store solutions. I wonder whether it is unbounded or bounded. If it is unbouned, then can it become very large on some cases? If it is bounded, can it deteriorate over time (i.e., have a chance to preserve an inferior solution that is dominated by some solutions eliminated previously).

  3. Why is the method called evolutionary genome search?

5 comments

We thank the reviewer for finding the problem interesting and the weight adaptation intuitive and effective. Our answer to W1 also draws on C2(above).

W1. It is not very clear to me why you would like to use such a two-time-scaled weight adaptation method. I can clearly see what they are, but struggle to understand why this is needed, particularly in the Abstract. Why would a fast scale and a slow scale help the optimization process? To sum up, the motivation of using the method is not very clear.

Thank you. Choosing the weights is a bilevel problem. The outer problem places the weights so that their solutions spread over the front (repulsion, Eq. 12). The inner problem finds those solutions, x∗(λ)∈arg⁡min⁡xSλ,τ(x)x^{\ast}(\lambda)\in\arg\min_x S_{\lambda,\tau}(x). The outer update only makes sense on converged points, since a point lies on the front only once its inner problem is solved. Solving this as a bilevel loop (solve every inner problem, then update the weights, repeat) is exact but needs a full inner solve per weight update. Two time scales replace it with a single loop. If the weights move slowly compared with the inner solver, xx follows x∗(λ)x^{\ast}(\lambda) and the weights follow the bilevel loop (Borkar, 1997, 2008), up to an error proportional to ηλ/ηx\eta_\lambda/\eta_x (C2). We have revised the Abstract and Introduction accordingly. [Revised: Abstract, Sec. 1 and 4.2, App. E.2]

W2. Also, in the Introduction, it is not very clear why an objective-space repulsion that pushes nearby particles' weights apart on the simplex can ensure better uniformity.

Thank you. The repulsion is computed between points in objective space, not between weights. When two particles are close in objective space, i.e. they converge to similar solutions, their weights are pushed apart so that they converge to different parts of the front. The push comes from the kernel on the standardised objectives, ωkj=exp⁡(−∥F~k−F~j∥2/h2)\omega_{kj}=\exp(-\Vert \tilde F_k-\tilde F_j\Vert ^2/h^2) (Eq. 12), which is strong for close pairs and negligible for distant ones. Crowded points are therefore separated while isolated ones are left alone, and a point stops moving when the pushes from its neighbours balance, i.e. when its gaps are locally equal. We measure closeness in objective space and not on the simplex, as evenly spaced weights generally give unevenly spaced points (Eq. 1). Two particles with nearly equal weights can converge to distant points, e.g. on different components of a disconnected or non-convex front, and should not repel each other. Conversely, a range of weights can select the same point (kinks, edges or degenerate fronts), and those particles should repel. Uniformity is therefore encouraged at the scale of the bandwidth hh, not guaranteed. Switching the repulsion off raises the median IGD by 43% and worsens 29 of 33 problems (5nES Q3). We have clarified this sentence in the Introduction. [Revised: Sec. 1]

W3. In the evolutionary multi-objective optimization area, weight adaptation techniques can make decomposition-based MOEAs work very well, even for problems with disconnected, degenerate, and inverted simplex Pareto fronts, see e.g. https://arxiv.org/pdf/2502.16481. However, in your experiment, it is clear to see that your method doesn't perform particularly well on those problems, such as DTLZ7. I am wondering what advantage your algorithm has compared to those weight adaptation methods in the EMO area.

Thank you for the reference. We agree that on DTLZ7 EMO weight adaptation does better. The cited paper reports IGD 0.0530.053–0.0540.054 for AdaW and ATM-MOEA/D with 100k function evaluations, and in our runs MOEA/D-AWA reaches a median of 0.059 against our 0.087. Among gradient methods DTLZ7 is one of our strongest results. TWARD and TWARD-S are first and second of the twelve methods in our tables (mean IGD 0.090 and 0.092, against 0.223 for the best baseline). The gradient baselines barely improve or get worse with more steps on DTLZ7 (App. D.10), while TWARD keeps improving with more generations and covers all four patches, with IGD 0.036 at G=25G=25 and 0.024 at G=100G=100 (Section 5.3), at a larger budget.

We also ran MOEA/D, MOEA/D-AWA, NSGA-III and RVEA on our 33 problems, with the standard settings of the original papers and 100k function evaluations. The budgets are not comparable, since a Jacobian costs more than a function evaluation. By mean IGD the EMO methods are best on 5 of these 33 problems [Revised: Sec. 5.4, App. E.8, Table 14]. On the 20 ZCAT problems, whose Pareto sets are curved in 30 dimensions, our IGD is 1.6–7.7$\times$ lower on every instance. The advantage of our method is that it brings weight adaptation to gradient-based MOO. Its evolutionary search acts only on the mm-dimensional genome (λ,τ)(\lambda,\tau) and not on xx, so its cost does not depend on dd, and every returned point is certified ε\varepsilon-stationary. We have added AdaW and ATM-MOEA/D to the related work. [Revised: Sec. 2]

Q1. I wonder whether it is possible to incorporate weight adaptation methods in the EMO area into your method for even better uniformity.

Thank you for the suggestion. The outer loop does not depend on how new genomes are proposed, so we replaced our breeding step with the rule MOEA/D-AWA uses to add weights: new genomes are placed near the archive points with the largest gaps to their neighbours. Our spacing test already removes crowded points, as AWA does, so we kept it. With the same solver, archive and budget (5 seeds, the 33 problems of W3), the two rules give similar IGD. AWA is lower by more than 3% on 12 problems, within 3% on 9 and higher by more than 3% on 12. It is better on 15 of the 20 ZCAT problems and worse on the degenerate fronts DTLZ5, DTLZ6 and WFG3, where the gap to the nearest neighbours is a poor measure of sparsity. A Wilcoxon signed-rank test over the 33 problems still finds no significant difference, so neither rule is better overall. [Revised: App. E.8, last two columns of Table 14]

Q2. I noticed you use an archive to store solutions. I wonder whether it is unbounded or bounded. If it is unbounded, can it become very large in some cases? If it is bounded, can it deteriorate over time (i.e., have a chance to preserve an inferior solution that is dominated by some solutions eliminated previously)?

Thank you. The archive has no hard cap, but its size is bounded by the spacing radius rr. A child that passes the convergence test is added only if no archived point lies within distance rr of it (in range-normalised objective space). Otherwise it replaces those points if it is more stationary than all of them, and is discarded if not (spacing test, Section 4.3 and Def. 5). Archived points are therefore always at least rr apart, so the archive holds at most as many points as fit on the front at that spacing, of order (1/r)m−1(1/r)^{m-1}. The radius thus sets the density. A larger rr gives a smaller archive and a smaller rr a denser one, as the sweep in Appendix D.1 shows.

The archive can deteriorate as the reviewer describes, because replacement is decided by stationarity, not by dominance. We measured how often this happens on all problems. On average 2.2 of the 24 replacements per run remove a point that dominates its replacement, and 0.9% of the returned points are dominated by a point removed earlier. A dominance guard, under which a child may not replace a point that dominates it, removes this at no cost in IGD (median change −0.5%), and we have added it as an option [Revised: Sec. 4.3, App. E.9]. All reported indicators use the non-dominated subset of the archive.

Q3. Why is the method called evolutionary genome search?

Thank you. The genome is the scalarisation (λ,τ)(\lambda,\tau), which the gradient solver turns into a Pareto point, as a genotype is expressed as a phenotype. Only genomes are recombined and mutated, and xx is always obtained by gradient descent.

I would llike to thank the authors for providing a detailed reply to my commetns and questions. Most of them are well addressed, particularly W1 and Q2.

However, for the reply to W3 and Q1, I would like to point out two things.

First, IGD obtained in different papers are not comparable as 1) they may use different reference sets (a reference set with more solutions will lead to better IGD values) and 2) solution sets to be considered may have different sizes (a larger set will in general have a smaller IGD). In addition, the IGD value also depends on the connection between solution sets to be evaluated and the reference set. So, IGD of 0.053-0.054 obtained in the mentioned papers is not comparable to 0.036 and 0.024 in your paper.

Second, the EMO algorithms used have rather poor performance on problems with irregular Pareto front. MOEA/D, NSGA-III and RVEA are not weight adaptation methods. MOEA/D-AWA performs poorly in terms of uniformity on such problems (this can be seen in related EMO papers).

I mean it is absolutely OK if your algorithm performs worse than advanced EMO algorithms since they are not the same type of approaches and each has its own advantage, but it is desirable to use state-of-the-arts in the experiments.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Common Questions

We thank the three reviewers for their careful reading. All new experiments use the same problems, reference fronts and seeds as the paper (5 seeds, matched budget of 160k Jacobians). They are in a new Appendix E, except the language-model experiment, which is Appendix F. Numbers in our replies refer to the revised PDF. Eqs. 1–14 are unchanged from the submission. The result tables keep their numbers (Tables 1 and 2). The tuned configurations are Table 6 (Table 5 in the submission). The new tables are Table 3 (UCI Adult), Table 4 (language model), Table 12 (lift error, App. E.1), Table 13 (ablation, App. E.3), Table 14 (rank-sum tests) and Table 15 (EMO comparison). Pointers such as [Revised: App. E.1] give the location of each change. Two questions were raised by more than one reviewer. We answer them once in C1 (below) and C2 (in the comment "General response (2/2): C2"), each of which lists the points it answers, and the replies under each review cover the rest. When a weakness and a question ask for the same analysis, we answer it at the question.

C1. Does the weight update produce the intended move?

5nES W1 and Q1 (when the update produces the intended move, and how large the error is on curved and degenerate fronts) and REQT W1 and Q1 (a bound on the error of the lift on curved fronts, and its size in practice). The effect on coverage, the WFG1/WFG3 gaps and the outer loop are answered in the individual replies.

We thank the reviewers for this question. The update produces the intended move up to a relative error set by the temperature and the curvature of the front, which we bound below and measure in Fig. 86 of the revised paper.

Take one particle converged under the weights λ\lambda, with scaled objectives g=(F−z)/sg=(F-z)/s (Section 3) and softmax weights π\pi (Eq. 5). The repulsion asks for the move v=r^/sv=\hat r/s. The lift (Eq. 14) sets log⁡λ′=log⁡λ−v/τ\log\lambda'=\log\lambda-v/\tau, and after xx converges again under λ′\lambda' the particle has moved by g′−gg'-g. We work to first order in vv, and the single loop follows this point up to O(ηλ/ηx)O(\eta_\lambda/\eta_x) (C2). At a converged point ∑iπi∇gi(x)=0\sum_i\pi_i\nabla g_i(x)=0, so π⊤(g′−g)=0\pi^\top(g'-g)=0, i.e. the particle moves only along the front, whose normal is π\pi. On a flat patch Section 4.2 gives g′−g=Pvg'-g=Pv with P=I−1π⊤P=I-\mathbf{1}\pi^\top. For a curved front, Eq. 5 can be rewritten as g=−τlog⁡λ+τlog⁡π+τlog⁡Z1g=-\tau\log\lambda+\tau\log\pi+\tau\log Z \mathbf{1}. Writing this at both converged points, subtracting, and applying PP, which removes the part along 1\mathbf{1} and leaves g′−gg'-g unchanged, gives

g′−g=Pv+τP(log⁡π′−log⁡π).g'-g=Pv+\tau P(\log\pi'-\log\pi). (∗)

On a flat patch π′=π\pi'=\pi and the move is exactly PvPv. On a curved front the last term is the whole error, and it appears only because the normal turns. If α\alpha is the rate at which the normal turns, ∥P(log⁡π′−log⁡π)∥≤α∥g′−g∥\Vert P(\log\pi'-\log\pi)\Vert \le\alpha\Vert g'-g\Vert, so whenever τα<1\tau\alpha<1,

e=∥g′−g−Pv∥∥Pv∥≤τα1−τα.e=\frac{\Vert g'-g-Pv\Vert }{\Vert Pv\Vert } \le \frac{\tau\alpha}{1-\tau\alpha}.

The error does not depend on vv, so smaller steps do not reduce it. This assumes that xx converges again to the nearby point, and it holds inside each component of a disconnected front. A particle at the edge of a component that is pushed outwards stays there or jumps to another component, and the outer loop covers this with new genomes. On degenerate fronts α\alpha is not defined, but the particle still moves along the front. Every archived point passes stat(x)≤ε\mathrm{stat}(x)\le\varepsilon.

Fig. 86 of the revised paper measures this. We apply one update to converged TWARD-S particles and let them converge again. The median cosine between g′−gg'-g and PvPv is 0.97. It is below 0.9 only on the problems marked red, which are tuned with τ≥0.5\tau\ge0.5 or have sharp curvature, so τα\tau\alpha is large and the bound gives no guarantee. Particles that do not move sit at the edge of the front or of a component (e.g. DTLZ7), or belong to problems tuned with τ≥0.5\tau\ge0.5. Halving or doubling the step leaves the error unchanged, as predicted. The effect on coverage is in 5nES Q3. [Revised: Sec. 4.2, Eq. 15, App. E.1, Fig. 86, Table 12]

AuthorsOpenReview ↗

Common Questions

C2. Does the single loop behave like the bilevel loop, and what about Adam?

Rcsj W1 (why a single loop can replace the bilevel loop), 5nES W2, Q2 and the Adam part of Q5 (stability of the coupled dynamics, and gradient descent vs. Adam), and REQT W1 and Q1 (an approximation analysis of the coupled dynamics). The step-size sensitivity (5nES Q5) and the outer loop (REQT W1) are answered in the individual replies.

We thank the reviewers for this question. Theorem 1 holds with the weights frozen. The two-time-scale theorem (Borkar, 1997, and Borkar, 2008, Ch. 6) states that the single loop follows the bilevel loop as ηλ/ηx→0\eta_\lambda/\eta_x\to0. Intuitively, if the weights move slowly relative to xx, then between two weight updates the particles have many steps to catch the current solutions x∗(λ)x^{\ast}(\lambda), so when the repulsion is evaluated they are already close to the points a bilevel loop would have used. What is left is the lag at the fixed ratio we use, i.e. how far xnx_n is from x∗(λn)x^{\ast}(\lambda_n), the solution for its current weights. If the inner minimiser is non-degenerate (∇x2S⪰μI\nabla^2_xS\succeq\mu I, the only new assumption on the problem) and the weight step is bounded (Eq. 11 divides the repulsion by rˉ\bar r), each xx-step contracts the distance to the current x∗(λ)x^{\ast}(\lambda) while that target moves by at most cηλc \eta_\lambda per step, so the distance settles at

∥xn−x∗(λn)∥≤cμηληx.\Vert x_n-x^{\ast}(\lambda_n)\Vert \le \frac{c}{\mu} \frac{\eta_\lambda}{\eta_x}. (∗∗)

The weights follow the bilevel loop to the same order. The result is local. It says nothing where x∗(λ)x^{\ast}(\lambda) jumps, which is where the outer loop acts, and it does not show that the weights converge. It holds for gradient descent, not for Adam. [Revised: Sec. 4.2, Eq. 16, App. E.2, Fig. 87]

Fig. 87 of the revised paper measures the lag. We run TWARD-S with ηλ/ηx\eta_\lambda/\eta_x from 10−310^{-3} to 1, then freeze the final weights and solve for xx again. Gradient descent and Adam both reach the solver's floor with frozen weights. As the ratio grows the lag rises by two to four orders of magnitude, close to proportionally to ηλ/ηx\eta_\lambda/\eta_x for Adam, as (∗∗) predicts, and with more scatter for gradient descent.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.