acceptodds
Withdrawn from ICLR 2027

EvoGrad: Automatically Evolving Gradient Aggregation Rules for Multi-Task Learning Using Genetic Programming

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Multi-task learning (MTL) aims to improve data efficiency and generalization by leveraging shared information across tasks. Many existing gradient balancing methods developed for MTL limit their aggregation rules (namely programs) to a linear search space spanned by task gradients, and determine the optimal programs merely based on a local decrease in training loss, which may not lead to model improvement. In this paper, we propose EvoGrad, an evolutionary framework based on multi-objective genetic programming (GP) that automatically discovers optimal programs. By composing nonlinear operations over task gradients, EvoGrad simultaneously explores more expressive search spaces beyond their linear span and identifies optimal programs. To accelerate the exploration, we adopt a hybrid selection strategy that first efficiently selects diverse and balanced candidate programs based on loss-decrease rates, and then identifies the best program using dataset-specific evaluation metrics. Experiments on four benchmark datasets with task numbers ranging from 2 to 40, together with three toy examples, demonstrate the effectiveness of EvoGrad in MTL.

Primary area:
optimization
Keywords:
Multi-task learning, Gradient aggregation, Genetic Programming

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_k97HOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes EvoGrad, which uses genetic programming to search for gradient aggregation expressions in multi-task learning. The method first screens candidate expressions based on the expected loss reduction for each task, then selects the final rule by temporarily updating the model and comparing task metrics on the current training batch. Experiments cover four multi-task datasets and synthetic problems, achieving good results on several metrics.

Critical strengths

The experiment in the paper covers different numbers of tasks and prediction types, and the appendix provides results for each task, facilitating observation of performance trade-offs behind the overall metrics. The authors also report the case where the model performs well on training but poorly on testing on NYU-v2, which helps understand the scope of applicability of the method.

Critical weaknesses
  1. The contribution of automatically discovered non-linear rules has not been sufficiently validated. Table 3 lacks a control group that only searches linear combinations, while Table 2 statistically reports the frequency of multiplication and division operations, yet it does not demonstrate that these operations bring performance gains. The paper discusses GradDrop and BayesAgg-MTL, but without experimental comparisons, it remains unclear what additional value automatic search offers over existing aggregation rules.

  2. The paper does not demonstrate whether the performance gains justify the additional training cost. The method requires evolutionary search and repeated updates, yet Table 3 only compares training times among its own variants. The advantages shown in Table 1 vary depending on the dataset and metric, and the performance trajectories are close to those of FairGrad and SVFair. If a comparison under identical training times were provided, it would verify whether the added complexity is justified.

  3. The scoring in Equation 5 is influenced by the magnitude of candidate directions. However, the actual updates in the code still undergo gradient clipping and Adam processing. A higher score does not necessarily imply a greater actual reduction. The paper needs to validate the relationship between scoring and actual loss changes, and verify whether the results still hold after controlling for direction scale.

Questions for authors
  1. When maintaining the same search process and candidate evaluation count, what results do we get from only searching linear weight combinations? How does the gain from automatic search compare to existing coordinate-wise gradient aggregation methods?

  2. Under the same training time, does EvoGrad still maintain its advantage over SVFair, FairGrad, and FAMO?

  3. How consistent is the proxy score with the actual loss change after update? Does normalizing the candidate direction's norm significantly affect the program's selection and final performance?

Reviewer_aZmFOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes EvoGrad, a genetic programming framework that automatically evolves gradient aggregation rules for multi-task learning. EvoGrad composes element-wise nonlinear operations over normalized task gradients and uses a multi-objective evolutionary strategy guided by three first-order loss-decrease surrogates, followed by a lightweight metric-guided selection of the best program per step. Experiments across four benchmarks and a toy example show that EvoGrad achieves competitive or superior results, particularly on the large-task CelebA dataset, and ablations support the contribution of the proposed objectives and selection strategy.

Critical strengths
  1. This paper presents a flexible approach to gradient aggregation using genetic programming. By combining linear and nonlinear operators, a series of aggregation strategies with greater expressive power and easier automatic discovery can be constructed.
  2. The proposed method achieves excellent results across four benchmarks ranging from 2 to 40 tasks. In particular, on the CelebA benchmark involving 40 tasks, the proposed evolutionary framework demonstrates its ability to identify effective aggregation rules in relatively complex multi-task settings.
Critical weaknesses
  1. This paper lacks sufficient evidence to support the necessity and effectiveness of nonlinear gradient aggregation. Although the authors argue that their contribution lies in exploring gradient aggregation directions beyond the linear space of the task, the current experiments do not sufficiently demonstrate or validate the advantages offered by this nonlinear search space. Specifically, neither the baseline comparisons nor the ablation experiments provide any validation of this claim. Moreover, this paper does not provide a theoretical analysis or derivation to demonstrate the necessity of nonlinear gradient aggregation.
  2. The proposed framework largely combines established techniques, including standard GP operators and NSGA-II, loss-decrease-based fairness objectives, and performance-aware gradient balancing. Although applying these techniques to nonlinear gradient aggregation is potentially interesting, the paper does not sufficiently clarify what constitutes a fundamentally new methodological contribution beyond their integration.
Questions for authors
  1. The proposed surrogate objectives are sensitive to the magnitude of the generated gradient directions. How does EvoGrad prevent genetic programming from exploiting gradient scaling rather than discovering genuinely better aggregation directions?
  2. Gradient aggregation targets suffer from scale bias, which may cause evolutionary search algorithms to optimize the magnitude of the gradient rather than the effective direction. Specifically, the objectives in Eq. (5) are not scale-invariant. When all task descent rates are positive, GP can improve fitness simply by amplifying the aggregated gradient without discovering a better direction. Moreover, metric-driven program selection carries the risk of adaptive overfitting. This concern is consistent with the results on NYU-v2, where Appendix B shows that EvoGrad achieves stronger training performance than SVFair but fails to maintain this advantage on the test set.
  3. Since PCGrad, CAGrad, and especially BayesAgg-MTL are missing from the empirical comparison, could the authors clarify why these widely used baselines were excluded? Including them would help verify whether EvoGrad's automatically evolved nonlinear aggregation rules truly outperform existing hand-designed nonlinear mechanisms.
  4. In half of the cases, EvoGrad’s actual performance does not consistently outperform the strong baseline. Based on the results in Table 1, existing gradient-balancing methods can achieve comparable or even better results without running large-scale GP searches. So why is it necessary to introduce such a complex evolutionary framework?
Reviewer_noGLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

EvoGrad uses genetic programming to discover gradient aggregation rules for multi-task learning. Candidate programs combine arithmetic operations over normalized task gradients. NSGA-II screens them using three objectives measuring overall predicted loss decrease, the number of benefiting tasks, and worst-task progress. A diversity-aware shortlist is then evaluated through temporary updates using task-specific metrics. Experiments cover four benchmarks with 2–40 tasks and three synthetic problems. EvoGrad achieves the best mean rank on three benchmarks, although improvements in aggregate relative performance are less consistent.

Critical strengths
  1. Interesting formulation and well-motivated selection strategy.
    Searching over compositional gradient programs offers a flexible alternative to scalar task weighting. Combining inexpensive first-order screening with direct task-metric evaluation addresses the potential mismatch between training-loss reduction and evaluation performance.
Critical weaknesses
  1. The performance–cost trade-off and robustness remain unclear.
    EvoGrad exceeds SVFair in aggregate relative improvement only on CelebA (1.01% versus 0.63%), while trailing it on the other three benchmarks (Table 1). Three runs are reported, but no standard deviations or confidence intervals are provided. Despite the population search and repeated candidate evaluations, the paper lacks runtime comparisons against the main baselines or results under equal wall-clock budgets. Moreover, candidates are scored on the same mini-batch used to compute gradients, and Appendix B documents a training–test performance gap on NYU-v2. These findings leave the reliability and practical value of the gains insufficiently established.
Questions for authors
  1. How does EvoGrad compare with a linear-only aggregation search under the same candidate-evaluation budget and update-scale controls? Do directions outside the task-gradient span contribute measurable improvements?
  2. Do the gains remain reliable across seeds and under matched wall-clock budgets? Please provide variability estimates and runtime comparisons with the baselines.

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

Related papers

Discussion (0)

Sign in to comment.