acceptodds
Under review as a conference paper at ICLR 2027

Hybrid Quantum Selection of Learning Rates in Language-Model Training: Termination Criteria and the Cost Against Classical Selection

Abstract

Learning-rate choices strongly affect language-model fine-tuning, and prior work shows that layer-wise learning rates can improve training; however, the best layer-group learning-rate profile is generally unknown. This paper studies quantum minimum finding as a layer-group learning-rate selector for a future coherent quantum training architecture. The search combines Dürr–Høyer minimum finding with BBHT search and considers two classes of stopping policies: fixed-budget and confidence-based termination. To simulate a coherent quantum search for the best learning-rate profile, measured profile scores at fixed training checkpoints are used to construct a lookup oracle that reproduces the required search behavior. The associated quantum-selection cost is evaluated under both a unit-cost model and IBM hardware-informed timing estimates. Experiments on a 1.5-billion-parameter language-model fine-tune evaluate when quantum selection can reduce the cost of exhaustive profile search, which becomes classically prohibitive as the learning-rate profile space grows.

Primary area:
optimization
Keywords:
quantum minimum finding, Grover search, learning-rate selection, language-model fine-tuning, termination criteria, hybrid quantum-classical training, cost accounting

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_JvE9OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates Grover-based quantum minimum finding for selecting layer-group learning-rate profiles during language-model fine-tuning. It combines the Dürr–Høyer minimum-finding algorithm with BBHT search, compares several termination criteria, and evaluates their behavior using classically precomputed loss tables and small-scale quantum hardware experiments. The intended motivation is to translate a quadratic reduction in search queries into lower learning-rate selection costs.

Critical strengths

No

Critical weaknesses

My main concern is that the paper does not establish a credible path from oracle-query savings to an end-to-end computational advantage. In the experiments, every candidate profile is evaluated classically to construct the score table, but this preprocessing cost is excluded from the quantum selection time, whereas the classical exhaustive baseline is charged for evaluating all candidates. Once these scores have been computed, the minimum can already be identified classically with negligible additional overhead relative to model evaluation. Consequently, the reported speedups do not demonstrate a practical advantage for the implemented workflow.

Questions for authors
  1. How would the coherent quantum loss oracle be implemented, and how would its cost affect the reported speedup?

  2. Would the reported speedup over classical exhaustive search remain if the cost of constructing the score table were included?

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

The paper proposes to use a quantum search procedure, founded in Grover search, to fine-tune learning rates for language model training. They discuss about different stratgies for cost accounting and termination criteria as well as simulate some scenarios that compare the different strategies.

Critical strengths

The authors tackle a challenge that practicioners face, thus providing a good motivation for the research question.

Critical weaknesses

As is, I do not see a convincing motivation on why quantum computing might be a good approach for this problem. The fact that Grover's algorithm allows searching this list alone does not make it appliable for this scenario. In particular, the authors simulate the calculation of f(z)f(z) and mention that this would ideally be done in a quantum routine, however, this very step is highly non-trivial. The training they propose to do for every possible combination of learning rates is to-date and also for the forseeable future (to the best of my knowledge, otherwise I would ask the authors to provide a reference) simply not possible on a quantum computer, thus, questioning the approach.

Further, even if we assumed that it would be possible to have the Language model on a quantum computer and train it, there would be a very significant overhead arising from the communication between classical and quantum computers since it is well-known that loading data into a quantum computer is inherently expensive, which is not discussed in the manuscript. It is for these reasons that I do not think there is a convincing motivation for employing quantum computers to the task at hand.

Further, the authors lack a proper baseline for their approach, which is partially also attributable to the missing background and related work section on learning rate selection. It is very unlikely that a standard way of learning rate selection today really is an exhaustive search over all possible combinations (if I am wrong please provide a reference). Rather, learning rate scheduling and adaptive learning rates are employed as, seemingly, the exact value of the learning rate is not very relevant, rather the adaptive nature.

The same goes for the results of Figure 4. I do see that their approach has a speed-up over an exhaustive search, but it is also evident that an exhaustive search can simply not be feasible to do for thousands of rows and the size of today's models. Further, I am not able to follow where the speed-ups come from. May I ask the authors how the training + Grover's search is simulated/executed?

Beyond content, the paper is not written as is common for academic publications. In particular, there is no background section with explanations of learning rate selection as well as the quantum routines and state-of-the-art work. Some of these explanations are mixed in with the methodology, making it difficult to distinguish between the contribution and background. Further, there is a lengthy appendix, that is hardly ever referenced. For these reasons the paper is difficult to follow.

Questions for authors
  • Could the authors provide a convincing motivation on why the approach is interesting and realistic? I am aware that this will not be in the near-term in any case (due to quantum HW limitations), but I think adressing what f(z)f(z) could be on a quantum computer and how a reasonable objective can be computed would be instructive.
  • Secondly, could the authors provide reasoning and any estimation about the overhead from classical-quantum communication and how this would impact the advantages they see from employing the quantum search?
Reviewer_TXV1OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper studied quantum minimum finding for choosing layer-group learning rates during language-model fine-tuning. Candidate losses were computed through short classical training replays and stored in a lookup oracle. Four stopping rules were compared using simulated searches. A small hardware experiment checks search behavior. The authors report success, regret, and conditional timing estimates. Full training runs compare selected profile sequences with fixed learning-rate choices, and larger search spaces were studied through extrapolation.

Critical strengths

The analysis of stopping overhead is useful. It shows finding better candidates can be much cheaper than establishing when to stop. The paper measures this distinction carefully. Using actual training losses give the search experiments some grounding. The stated oracle assumptioncalso makes the limitations visible to the reader.

Critical weaknesses

The paper fails in establishing a computational advantage for the learning problem. It computes every candidate score classically, then excludes that work from quantum selection cost. Once those scores exist, the best candidate can be already known classically. A future coherent training oracle may change the problem. But that oracle was neither implemented nor given a credible full resource estimate. The reported speedups therefore did not demonstrate a faster training method.

The ML motivation is also weakly validated. The comparison centers around exhaustive search. Practical hyperparameter selection rarely requires finding the exact best points in fixed tables. It requires finding a good setting within a budget. Strong constant profiles also perform almost identical to the selected sequences. This raises a basic question: is the proposed search solving a problem that matters here? For a top ML conference, the paper needs evidence of useful selection under realistic costs. The current results mainly compare quantum stopping policies.

Questions for authors

Questions:

  • Is the intended contribution a stopping-rule study or a practical learning-rate selector? The evidence supports the former more clearly. Please state which claim the paper asks reviewers to accept.
  • At the same total replay budget, does the selector beat random search or a simple classical adaptive method? Compare final training quality, not only exact recovery of the table minimum. Small score differences may have little practical value.
  • What resource estimate supports the assumed coherent oracle? Include training updates, loss evaluation, uncomputation, and numerical precision. Without that estimate, which speedup claims remain justified? Please also explain why repeated selection is needed when strong constant profiles perform nearly as well.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.