acceptodds
Under review as a conference paper at ICLR 2027

TrACE: Trajectory-Adaptive Error Mitigation for Variational Quantum Eigensolvers

Abstract

Variational quantum eigensolvers (VQEs) use repeatedly measured energy estimates as in-loop optimization signals, making the parameter trajectory itself vulnerable to hardware noise and finite-shot fluctuations. Existing mitigation methods are limited in this regime: physics-based schemes require costly per-query executions, while offline learning-based mitigators are usually frozen pointwise predictors and can fail under the trajectory-dependent distribution shift induced by VQE. We propose **TrACE**, a trajectory-aware pretrain-and-adapt framework for in-loop VQE error mitigation. For each target Hamiltonian, TrACE pretrains a reusable mitigator within a problem-scoped ansatz pool, learning a structured mitigation prior from circuit, parameter, and noisy-measurement features. During deployment, it serves as the optimizer-facing objective and absorbs trajectory context. It is intermittently adapted without ideal-energy labels under self-supervision, updating only the fusion and output modules through repeated noisy views, trajectory stability, and weak prior anchoring. We also introduce a trajectory-level evaluation protocol that measures the quality of the parameters and convergence of VQE, rather than rewarding overly optimistic pointwise energy estimates. Across TFIM, molecular-energy, and MaxCut benchmarks, including a 100-qubit TFIM instance, TrACE improves solution quality, convergence stability, and optimization speed over representative baselines under various noise settings.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Quantum Error Mitigation, Graph Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_E98QOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper propose TrACE, a trajectory-aware pretrain-and-adapt framework for in-loop VQE error mitigation. The trained model can provide corrected energy predictions to classical optimizers to enhance the training. The paper considers multiple use cases including TFIM, molecular energy and MaxCut benchmarks. The paper claims improved solution quality, convergence stability and optimization speed.

Critical strengths
  1. The problem is well motivated. QEM should not consider the deviation in single expectation values but how these deviations' effects on VQE optimizer.
  2. Evaluation metrics are better than conventional QEM benchmarks, especially for the purposr of VQE.
  3. Good collections of ablations including TFIM, MaxCut, H4 as well as noise settings.
Critical weaknesses
  1. The authors should elaborate whether the self-supervised adaptation can actually mitigate systematic error. (Eq. (14)–(17))
  2. Please explain why the \kappa_{t} is set to 0 during training and can expect that the model can learn the correct representation of trajectory context.
  3. Baseline comparisons are not entirely fair. For example, TrACE use 3 random seed while other baselines only have one run.
Questions for authors

1.Please clarify how the 100 qubit simulation is carried out. In the paper it said the MPS is used. It is atraightforward in noise free case, but how the "noisy" version is simulated. 2. Please provide the details of MPS bond dimension and truncation tolerance? 3. Please provide some perspectives on variational quantum circuits which are not simulable on shallow MPS as mentioned in this paper. This is important for the general applicatility of the proposed method.

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

The authors introduce TrACE, a learning-based error mitigation framework that provides corrected energy estimates during variational quantum eigensolver (VQE) optimization. For each target Hamiltonian, the model is pre-trained on noisy and ideal energy pairs using circuit structure, variational parameters, and measurement features. During VQE, the encoder remains frozen while the fusion module and output head are updated in an unsupervised manner with consistency, trajectory stability, and anchor loss. Experiments on five problems evaluate the resulting optimization trajectories and report improvements in accuracy and convergence over other baselines.

Critical strengths

1. Lightweight adaptation without ideal energy labels

The framework adapts only the fusion module and output head during VQE, using self-supervised losses without requiring ideal energy labels. The ablation study in Figure 5 and Table 6 shows that online adaptation improves convergence compared with the frozen pre-trained model.

2. Shared mitigation and model across ansatz families

TrACE supports a shared pre-trained mitigator across multiple ansatz families for a fixed Hamiltonian. The TFIM experiments evaluate the shared pre-trained mitigator on both HEA and HVA circuits with separate online adaptation for each ansatz.

Critical weaknesses

1. The benefit of the online adaptation is not sufficiently isolated.

The ablations of frozen versus adapted explains that adaptation helps, but it also uses additional measurements to make independent noisy data. There is no comparison with a frozen model that uses the same measurement budget for averaging, or with an MLP baseline adapted using the same self-supervised losses. It remains unclear how much of the improvement requires the proposed adaptation framework rather than additional sampling or a simpler adaptation scheme.

2. Insufficient description of model architecture

The fusion module is not specified sufficiently as well as the output head. Since the fusion module and output head are the components updated during online adaptation, the missing details make it difficult to reproduce.

Questions for authors
  1. Does online adaptation still improve performance over frozen TrACE when the additional measurements are instead used to average the frozen model's predictions? How does TrACE compare with an MLP adapted using the same self-supervised losses and measurement budget? These comparisons would help clarify whether the adaptation strategy generalizes across backbone architectures and whether simple averaging or smoothing operations can achieve similar improvements.

  2. Please specify the internal operations of the fusion module and output head.

I would be willing to reconsider my score if the responses clarify these points.

Reviewer_LfT6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

TrACE combines a pretrained energy predictor with online adaptation during VQE to reduce the impact of noise on both energy estimates and the optimization trajectory.

Critical strengths

The problem is well motivated. Adaptation without ideal-energy labels during deployment is useful. Evaluating the quality of the final parameters is a strength, and the frozen-versus-adapted comparison provides encouraging evidence.

Critical weaknesses

The baseline comparisons need matched seeds, settings, horizons, and total measurement budgets.

It is also unclear how much benefit comes from correcting online measurements versus learning the energy landscape from ideal-energy training labels.

A fully measurement-free control and clearer GPR implementation details would help isolate the contribution.

Questions for authors

My evaluation score will be increased if the weaknesses and the follow-up questions are addressed and discussed:

Can you provide comparisons using matched initial parameters and multiple seeds for TrACE and the strongest baselines? Please also account for total measurement shots, including adaptation and calibration overhead.

How does TrACE compare with a predictor trained on the same ideal-energy labels using only circuit structure and parameters?

Does the GPR baseline implement the ansatz-specific prior of the cited method? It is important to show the comparison to establish the benefit over existing trajectory-aware mitigation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.