acceptodds
Withdrawn from ICLR 2027

Hq-Recycle: Efficient Bilevel Optimization through Adjoint Recycling

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

We introduce Hq-Recycle, a quasi-Newton method for bilevel optimization that reduces repeated work in implicit hypergradient estimation by recycling adjoint estimates across outer iterations. The method tracks the lower-level solution using limited-memory quasi-Newton updates and retains the previous adjoint estimate. A fresh directional finite-difference curvature probe produces a one-dimensional Galerkin initialization, followed by preconditioned corrections only when the estimated residual exceeds a prescribed tolerance. This procedure limits refinement to a small probe budget while accounting for changes in the current adjoint system. For exact symmetric positive-definite systems, we establish an energy-minimization property of the recycled initialization and an error bound governed by the previous residual and inter-iteration system drift. We evaluate the four-probe configuration on three label-corrupted linear classification tasks and grouped ridge regression against qNBO, SHINE, PZOBO, BOME, HOAG, and AID-CG, using validation-based tuning and five fresh evaluation seeds. In classification ablations, recycling reduces the mean number of adjoint probes from four to approximately 1.07 per outer iteration. On grouped ridge regression, Hq-Recycle achieves a 2.44-fold speedup over tuned qNBO in mean time to a prespecified validation target and lowers mean test loss by 20.2% under a 0.5-second solver budget.

Primary area:
optimization
Keywords:
bilevel optimization, hypergradient estimation, adjoint recycling, quasi-Newton methods, implicit differentiation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_TDs1OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies how to reduce the cost of computing hypergradients in bilevel optimization. The main idea is to reuse the adjoint from the previous outer step. The method first rescales the old adjoint with a cheap one-dimensional step.

Critical strengths

The idea is simple and easy to implement, with little extra memory cost.

Critical weaknesses
  1. Reusing the previous adjoint as a warm start is already common in bilevel optimization. The main new part is a one-dimensional rescaling step, which seems to be a small change to existing methods. More importantly, the ablation results do not clearly support the main idea. Removing the rescaling step makes almost no difference. The version without adjoint recycling even performs better and faster on ridge regression. This makes me question whether recycling is really the reason for the improvement. The gains may come from early stopping or other implementation choices.

  2. The claim of low cost is not fully convincing. The method needs very few steps mainly after it has already reached a good solution. During the early stage of optimization, the cost saving seems much smaller. The experiments are limited to small-scale linear problems with strongly convex lower objectives. There are no tests on neural networks or non-convex lower-level problems. The paper also lacks comparison with related Krylov recycling methods. With more time, other methods seem to catch up. The theory only provides a local error analysis based on standard results, without a convergence guarantee for the full bilevel algorithm. The paper also introduces too many variants and names, making the main contribution hard to follow.

Questions for authors

See wakenesses.

Reviewer_vgXkOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Hq-Recycle, a quasi-Newton bilevel optimizer that reuses the previous adjoint estimate when computing implicit hypergradients. A fresh curvature action supports one-dimensional Galerkin rescaling, followed by residual-controlled preconditioned corrections within a four-probe budget. The method combines this adjoint solver with L-BFGS lower-level tracking and several implementation optimizations. The paper analyzes local approximation properties and evaluates short-budget performance on grouped ridge regression and three label-corrupted linear classification tasks.

Critical strengths
  1. The method is compact and practically motivated. Adapting a previous adjoint to the current system is a reasonable way to reduce repeated work. The recycling component retains one vector beyond the existing L-BFGS memory, and the probe budget explicitly includes the refresh. The complete solver demonstrates useful short-budget gains against the evaluated established baselines, including a 2.42-fold improvement over qNBO in mean time to the fixed ridge validation target in the ten-seed confirmation study.

  2. The empirical investigation is careful and transparent. The confirmation study uses separate development and evaluation seeds, repeated synchronized timing, paired uncertainty estimates, and independent lower-response evaluation. The authors distinguish the three experimental cohorts, report failures, and include matched warm-start and zero-start controls. Real-system replay and controlled drift experiments help explain the mechanism, while unfavorable results and the distinction between throughput and time to a quality target are explicitly acknowledged.

Critical weaknesses
  1. The end-to-end incremental value of the proposed recycling and rescaling mechanism remains insufficiently established. In Table 1, Hq-Recycle and the no-rescaling variant have almost identical mean ridge validation losses at 0.5 seconds (0.04793 versus 0.04791), and Section 4.3 reports a paired difference interval that includes zero. Table 3 likewise shows very similar classification results. More importantly, after independent outer-rate selection, the no-adjoint-reuse variant achieves a lower ridge validation loss (0.04685) and reaches the 75% development-fixed target faster (0.144 versus 0.209 seconds). This does not negate the complete solver's gains over established baselines, but it weakens the evidence that the proposed adjoint adaptation is responsible for its practical advantage. The local accuracy improvements and controlled amplitude-drift savings are informative; a clearer demonstration of when they translate into better full-optimization performance is still needed.

  2. The most directly related recycling approaches are not included in the empirical comparison. The paper cites LancBiO (Yang et al., 2025) and Ehrhardt et al. (2024), which exploit subspace information across related hypergradient systems, but leaves such comparisons to future work. Since the main contribution is a compact recycling strategy, comparison with these approaches is important for establishing its contribution beyond existing reuse methods. A single retained vector is attractive, but the resulting runtime, accuracy, and total-memory trade-off has not been demonstrated against methods that retain a richer subspace.

Questions for authors
  1. Incremental benefit: Can the authors demonstrate an application setting in which Galerkin rescaling provides a reproducible time-to-quality advantage over independently tuned Warm-PCG, and adjoint reuse improves over independently tuned Zero-PCG? How should the favorable zero-start ridge result affect the recommended choice of initialization?

  2. Direct recycling comparisons: Can Hq-Recycle be compared with LancBiO or an established recycled Krylov solver under comparable tuning effort? Please report runtime, solution quality, and total peak memory, including both lower-level curvature storage and adjoint workspace, to establish the proposed storage–accuracy trade-off.

Reviewer_M7rKOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes Hq-Recycle, a bilevel optimization method that reuses adjoint estimates across outer iterations to reduce hypergradient computation. It combines L-BFGS lower-level updates with rescaling of the previous adjoint and residual-controlled corrections under a fixed curvature-probe budget. The authors provide a local approximation-error analysis and evaluate the method on data reweighting and grouped ridge regression.

Critical strengths
  1. Reducing repeated computation in implicit hypergradient estimation is a practically relevant problem, and reusing previous adjoint estimates could help.

  2. The paper includes component ablations and controlled experiments examining when recycling and rescaling help.

Critical weaknesses
  1. Presentation and background. The background, problem formulation, and related-work discussion are not sufficiently developed. The paper should explain the problem more clearly and distinguish the proposed approach from existing methods.

  2. Insufficient motivation. The paper moves into algorithmic details before adequately motivating the design. Why should previous adjoint information remain useful across outer iterations? What benefits are expected, and how do these considerations lead to the proposed refresh, rescaling, and correction procedure?

  3. Convergence and complexity. The analysis establishes local adjoint-error bounds without convergence analysis. Can the authors provide a convergence guarantee? Since efficiency is the main motivation, a complexity analysis of computational cost against related works would also be valuable.

  4. Use of paper space. I strongly suggest reorganizing and better utilizing the main-paper space to improve accessibility.

Questions for authors

Please check the weakness.

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

Related papers

Discussion (0)

Sign in to comment.