acceptodds
Under review as a conference paper at ICLR 2027

Privileged Guidance for Learning Fitness Landscape Representations from Limited Observations

Abstract

Learning informative fitness landscape representations from limited observations is challenging because sparse function evaluations provide only partial and sampling-sensitive views of the underlying landscape. Although richer observations, function structures, and optimizer behaviors can provide complementary landscape information, they are costly or unavailable when representing a new problem. To address this issue, we propose Privileged Guidance for Fitness Landscape Learning (PGFL), which improves low-budget representation learning by combining consistency across different limited observations with richer information available only during pretraining. Specifically, PGFL employs a landscape encoder that combines pointwise relational encoding with dimension-aware latent aggregation to obtain fixed-length representations applicable across dimensions. Multi-view alignment promotes representation consistency across different sampling designs and transformed function instances. PGFL further exploits higher-budget observations through cross-budget distillation, while function structures and optimizer behaviors provide complementary structural and behavioral supervision for the learned representations. After pretraining, all privileged information and training-only components are removed, and the frozen encoder requires only limited observations of a new function for downstream use. Experiments on the 24 noiseless BBOB functions evaluate PGFL on high-level landscape-property prediction and automated algorithm selection, together with cross-dimensional generalization, component ablations, and representation analyses. The **source code** is provided in the **supplementary material**.

Primary area:
optimization
Keywords:
Fitness landscape learning, black-box optimization, pretraining, privileged guidance, contrastive learning, alignment learning

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_y5RfOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces a framework for learning fitness landscape representations from limited function evaluations. During pretraining, it combines multi-view alignment with guidance from higher-budget observations, function structure, and optimizer behavior. Only the frozen encoder is needed at deployment. The method is evaluated on the BBOB benchmark.

Critical strengths

The training setup is well motivated for black-box optimization, where function evaluations can be expensive. PGFL uses richer information offline while keeping the evaluation budget low on new problems. The algorithm-selection results also include the sampling cost, which makes the reported gains more meaningful.

Critical weaknesses
  • The ablations do not sufficiently separate the paper’s main methodological contribution from the other design choices. The overall benefit of privileged guidance is not directly isolated. Table 4 removes one component at a time, but every variant still uses some privileged supervision. A comparison using the same features, encoder, and training data with multi-view learning alone is missing. This matters because removing multi-view alignment causes the largest performance drop.

  • Both downstream tasks train and test on different instances of the same 24 BBOB functions. Recognizing familiar function types may already help predict their landscape labels and solver performance. It remains unclear whether these results carry over to unseen function types.

Questions for authors

Q1. How much does privileged guidance add over the same model trained with multi-view alignment alone?

Q2. How does PGFL perform when entire function types are held out?

Q3. Which Deep-ELA model variants and pretrained checkpoints were used for the 25d and 50d comparisons, and how were they selected?

4 comments

Thank you for your comments and suggestions on our paper. We address each question in turn below.

Q1: The multi-view-only baseline uses the same input descriptors and encoder architecture as PGFL. It retains both alignment objectives and disables all three privileged objectives.

VariantHLP Overall ↑Multi. ↑Global ↑Funnel ↑RF relERT ↓kNN relERT ↓
PGFL0.97 ± 0.010.95 ± 0.010.95 ± 0.011.00 ± 0.0111.80 ± 1.3713.39 ± 2.36
Multi-view only0.91 ± 0.010.90 ± 0.000.85 ± 0.010.99 ± 0.0013.56 ± 1.4915.36 ± 2.13
PGFL − Multi-view only+0.06+0.05+0.10+0.01−1.76−1.97

The comparison supports the added benefit of privileged guidance for both HLP and AAS.

Q2: We evaluated HLP and AAS using five folds grouped by BBOB function ID. In each fold, we used the frozen PGFL encoder to extract features and fitted the downstream models using only the training function types. For HLP, we used instances 0–99 of the training types and instances 100–119 of the held-out types. For AAS, we used the available instances of the respective types. PGFL and ELA used the same folds and a sampling budget of 25d25d. The original PGFL results from Table 4 are included for reference.

MethodEvaluation splitHLP Overall ↑Multi. ↑Global ↑Funnel ↑RF relERT ↓kNN relERT ↓
PGFLOriginal instance split0.97 ± 0.010.95 ± 0.010.95 ± 0.011.00 ± 0.0111.80 ± 1.3713.39 ± 2.36
PGFLFunction-type holdout0.60 ± 0.010.45 ± 0.010.38 ± 0.010.97 ± 0.0114.12 ± 2.1019.96 ± 1.15
ELA (25d25d)Function-type holdout0.54 ± 0.000.36 ± 0.010.33 ± 0.010.93 ± 0.0117.23 ± 1.1124.56 ± 4.75

Under function-type holdout, PGFL outperforms ELA on both HLP and AAS.

HLP is harder under this split because the classifiers must predict landscape properties for unseen function types. In particular, some Global labels occur in only one BBOB function type, so the corresponding training fold has no examples of those labels when that type is held out.

Q3: We selected Deep-ELA Large to use the same model variant across all four BBOB dimensions in our main evaluation (d∈{2,3,5,10}d \in \{2, 3, 5, 10\}). Deep-ELA's dimensionality limit counts both decision variables and objectives (Seiler et al., Sections 3.2.1 and 5.1). For a 10-dimensional single-objective problem, d+m=10+1=11d + m = 10 + 1 = 11, which exceeds Medium's limit of 6 and fits within Large's limit of 12. We used the released large_25d_v1 and large_50d_v1 checkpoints as frozen feature extractors for the 25d25d and 50d50d comparisons, respectively.

Weakness 2 · Limited evaluation of generalization

We agree that the original instance-based split provides limited evidence for generalization to unseen function types. We chose BBOB as a standard ELA benchmark with established landscape labels, standardized protocols, and extensive public optimizer records, enabling comparisons with existing methods.

To broaden the evaluation, we added an algorithm-selection experiment on UAV path planning in 6, 12, and 30 dimensions, using 400 problem instances per dimension and eight candidate optimizers. The PGFL and Deep-ELA encoders remained frozen, and the selectors, including PCT, were fitted using only the training folds. We report normalized fixed-budget objective scores (×100; lower is better), summarized as mean ± SD across ten sampling repeats.

DSBSELA 25d RFELA 25d kNNELA 50d RFELA 50d kNNDeep-ELA Large 25d RFDeep-ELA Large 25d kNNDeep-ELA Large 50d RFDeep-ELA Large 50d kNNPCT (100)PGFL RFPGFL kNN
67.42 ± 0.007.93 ± 0.359.58 ± 0.677.70 ± 0.479.65 ± 0.637.09 ± 0.329.98 ± 0.857.12 ± 0.319.99 ± 0.566.87 ± 0.486.69 ± 0.759.75 ± 0.32
129.04 ± 0.009.96 ± 0.3813.36 ± 0.819.71 ± 0.3712.67 ± 1.46----10.13 ± 0.359.42 ± 0.3612.61 ± 0.36
306.20 ± 0.006.20 ± 0.007.26 ± 0.236.20 ± 0.007.15 ± 0.30----6.50 ± 0.185.67 ± 0.157.13 ± 0.21

PGFL uses 25d25d observations, averaging results from three pretrained encoder seeds within each sampling repeat. “-” denotes unsupported Deep-ELA dimensions.

PGFL remains competitive with the evaluated methods on this application benchmark, providing additional evidence for transferring representations pretrained on generated functions to applications beyond BBOB.

Appendix E.1 and E.4 already discuss dependence on the pretraining distribution and the limited empirical scope. We will clarify these boundaries in the main text. Future work will broaden pretraining through more diverse function generators and application-derived optimization traces, and further evaluate higher-dimensional, noisy, constrained, and mixed-variable problems.

I thank the authors for their response, which has addressed my concerns. I am happy to maintain my positive score.

Thank you very much for the reply. We also appreciate your recognition of our work and your constructive comments, which further improve our paper. Thank you for your contribution to the research community.

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

The paper proposes PGFL for learning fitness landscape representations from limited observations. It combines multi-view alignment with training-only guidance from larger observation sets, function structures, and optimizer behavior. The frozen encoder is evaluated on landscape-property prediction and algorithm selection, with improvements over several existing representations.

Critical strengths
  1. The method addresses a clear practical need. Rich information is used during offline pretraining, while a new black-box function only requires a small observation budget.

  2. The evaluation includes both landscape-property prediction and algorithm selection. Comparisons with ELA and Deep-ELA use matched observation budgets, and algorithm-selection results include the cost of collecting observations.

Critical weaknesses
  1. The main benefit of privileged guidance is not fully separated from the new encoder and multi-view training. Table 4 removes each guidance source separately, leaving the other privileged signals active. A baseline using the same encoder and pretraining functions, with multi-view alignment but no privileged guidance, is needed to establish the central contribution.

  2. The basis for aligning different function instances needs clearer justification for algorithm selection. Sharing an expression structure does not imply the same optimization difficulty or best solver. The behavior loss itself treats these differences as important by excluding the transformed instance. Although separate projection heads are used, the paper should explain which differences alignment should remove and which the representation should retain.

Questions for authors

Refer to the weaknesses above.

3 comments

We appreciate the reviewer’s thoughtful assessment and constructive comments. Below, we address the questions raised in the review.

Q1: We evaluated a multi-view-only baseline with the same input descriptors and encoder architecture as PGFL. It retains sampling and instance alignment, with budget distillation, structural guidance, and behavior guidance disabled.

VariantHLP Overall ↑Multi. ↑Global ↑Funnel ↑RF relERT ↓kNN relERT ↓
PGFL0.97 ± 0.010.95 ± 0.010.95 ± 0.011.00 ± 0.0111.80 ± 1.3713.39 ± 2.36
Multi-view only0.91 ± 0.010.90 ± 0.000.85 ± 0.010.99 ± 0.0013.56 ± 1.4915.36 ± 2.13

PGFL achieved higher HLP scores and lower relERT with both selectors. This comparison supports the added benefit of privileged guidance beyond multi-view alignment alone.

Q2: We agree that related instances can differ in optimization difficulty and their best solver. We therefore use basic, controlled transformations. Alignment aims to reduce representation sensitivity to input translation, signed coordinate permutations, and positive affine objective changes. Meanwhile, the encoder should retain landscape properties and optimizer-response information, including progress and search dynamics.

To balance these goals, alignment operates in a projection space, while behavior guidance supervises the encoder representation used downstream. Behavior labels apply only to the original function’s Anchor and Sampling views, since the transformed instance may have a different optimizer response.

The following ablations retain all three privileged objectives.

VariantHLP Overall ↑Multi. ↑Global ↑Funnel ↑RF relERT ↓kNN relERT ↓
PGFL0.97 ± 0.010.95 ± 0.010.95 ± 0.011.00 ± 0.0111.80 ± 1.3713.39 ± 2.36
w/o instance alignment0.92 ± 0.000.87 ± 0.000.89 ± 0.000.99 ± 0.0012.87 ± 0.1914.82 ± 1.27
w/o multi-view0.86 ± 0.010.80 ± 0.020.78 ± 0.020.99 ± 0.0012.79 ± 1.2516.23 ± 4.04

PGFL achieves the best mean performance across HLP Overall and both AAS selectors, supporting the benefit of instance alignment within this joint training framework. We will clarify these design motivations in the main text, including which variations alignment aims to reduce and which task-relevant information the encoder is encouraged to retain.

Thanks to the author's answer to my questions, I will maintain my positive score.

We greatly appreciate the reviewer’s time and valuable comments. Thank you for the positive evaluation.

Reviewer_u7CNOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper learns a representation of a continuous black-box objective from a small observation set and freezes it for downstream use. Each sampled point is described by twenty hand-designed relational descriptors, encoded by a shared MLP, then aggregated by dimension-conditioned latent queries through cross-attention and latent Transformer blocks. Pretraining uses generated expression-tree functions. Three low-budget views of the same function, differing in sampling design and in the function instance, are aligned contrastively. Three further objectives exploit information available only offline: distillation from a higher-budget teacher view through an EMA encoder, regression onto eight structural attributes of the generating expression, and regression onto eight optimizer-behaviour summaries from three probe optimizers. All training-only components are removed after pretraining. Experiments on the 24 noiseless BBOB functions validate the performance of the learned landscape representations.

Critical strengths
  1. The idea is generally interesting and comprehensive. It puts a sample-based descriptor interface, one encoder and three kinds of offline supervision into a single route, and each objective is ablated, so every branch has a measured contribution.

  2. The experimental results are sound. Two downstream tasks, two selectors, three sampling designs and a selection metric that charges the representation cost are reported against cost-matched baselines, which is broad.

Critical weaknesses
  1. The privileged information may provide incorrect guidance. The encoder sees 25d points, while each target is a property of the function behind them, and two functions can produce the same points with different targets. The same input is then associated with different targets across functions, and nothing in that input identifies which function the points came from. The ablation also leaves the size of this effect open, since the deltas in algorithm selection are within the reported spread.

  2. The generalization test is rather limited. Pretraining and evaluation both stay inside synthetic continuous functions, so the results show fit within one family rather than transfer to another. The only cross-dimensional test is a single unseen dimension.

Questions for authors
  1. Do the three privileged objectives help because of what they encode about the function, or only because they add another training signal?

  2. The eight structural attributes describe how a function was constructed, not what its sampled landscape looks like. Can they be recovered from 25d observations? If two different trees produce almost the same observations but different attributes, what should the encoder output?

  3. Pretraining uses generated expression trees and both downstream tasks are labeled from BBOB. How far is the representation expected to transfer beyond that family?

  4. The ablation removes both alignment branches together, and that is where the largest drop appears. Which of the two, the sampling view or the instance view, accounts for the gain?

  5. How much of the reported gain would remain without pretraining? Please compare with randomly initialized encoder of the same architecture.

3 comments

We thank the reviewer for carefully reading our paper and providing detailed feedback. We respond to each question below.

Q1: We tested the value of the supervisory information while retaining the auxiliary heads and loss forms. Structural targets and complete optimizer-behavior target blocks were shuffled across functions of the same dimension. For distillation, the EMA teacher received the student's 25d25d Anchor observations in place of the nested 100d100d view. We evaluated these controls alongside the corresponding objective-removal variants and multi-view-only pretraining under the same protocol.

ObjectiveVariantHLP Overall ↑Multi. ↑Global ↑Funnel ↑RF relERT ↓kNN relERT ↓
ReferencePGFL0.97 ± 0.010.95 ± 0.010.95 ± 0.011.00 ± 0.0111.80 ± 1.3713.39 ± 2.36
Multi-view only0.91 ± 0.010.90 ± 0.000.85 ± 0.010.99 ± 0.0013.56 ± 1.4915.36 ± 2.13
Budget distillation25d25d Anchor teacher0.93 ± 0.000.90 ± 0.000.90 ± 0.000.99 ± 0.0013.82 ± 1.1314.83 ± 1.55
w/o budget distillation0.93 ± 0.010.90 ± 0.010.91 ± 0.011.00 ± 0.0012.57 ± 1.2413.97 ± 2.66
Structural guidanceShuffled structural targets0.92 ± 0.000.90 ± 0.000.87 ± 0.010.99 ± 0.0014.32 ± 0.7617.27 ± 1.56
w/o structural guidance0.94 ± 0.010.91 ± 0.010.91 ± 0.011.00 ± 0.0012.52 ± 1.5514.20 ± 1.76
Behavior guidanceShuffled behavioral targets0.92 ± 0.000.88 ± 0.000.90 ± 0.000.99 ± 0.0013.07 ± 1.2315.18 ± 1.03
w/o behavior guidance0.93 ± 0.010.89 ± 0.010.90 ± 0.011.00 ± 0.0012.61 ± 1.4314.35 ± 1.62

PGFL achieves higher HLP Overall and lower mean relERT with both selectors than multi-view-only pretraining. The shuffled-target and 25d25d Anchor teacher controls also yield higher mean relERT than removing the corresponding objective. Together, these comparisons support a contribution from correctly paired targets and richer teacher observations beyond simply adding auxiliary losses.

Q2: The eight attributes are computed from the known expression trees and serve as privileged supervision during pretraining. We measured how accurately they can be predicted from 25d25d observations using the frozen PGFL encoder and its pretrained structural head on 1,024 independently sampled synthetic test instances. The baseline predicts, for each attribute, the median training target among functions of the same dimension. Results are averaged across three pretraining seeds.

Structural attributePGFL MAE ↓Dimension-median MAE ↓
Complexity0.080.09
Operator entropy0.080.10
Smooth-nonlinear ratio0.110.12
Nonlinear nesting0.150.24
Periodic ratio0.060.07
Nonsmooth ratio0.070.05
Multiplicative ratio0.090.13
Coordinate-structure ratio0.160.22

PGFL achieves lower mean MAE on seven of the eight attributes. These results support partial, attribute-dependent predictability within the generated function distribution.

When different trees produce nearly identical observations but different attributes, the encoder extracts a representation from the observed evidence, and the structural head estimates the attributes using relationships learned during pretraining. Nearly indistinguishable observations may therefore receive similar estimates despite different target labels. The prediction experiment measures their accuracy, while the controls in Q1 assess the downstream contribution of this supervision.

Q3: We expect transfer beyond BBOB to depend on pretraining coverage. BBOB was chosen as a standard ELA benchmark with established landscape labels, standardized protocols, and extensive public optimizer records, enabling reproducible comparisons with existing methods.

To examine transfer beyond BBOB, we evaluated UAV path planning in 6, 12, and 30 dimensions, using 400 problem instances per dimension and a portfolio of eight candidate optimizers. The PGFL and Deep-ELA encoders remained frozen, and the selectors, including PCT, were fitted using only the training folds. We report normalized fixed-budget objective scores (×100; lower is better) as mean ± SD over ten sampling repeats.

DSBSELA 25d RFELA 25d kNNELA 50d RFELA 50d kNNDeep-ELA Large 25d RFDeep-ELA Large 25d kNNDeep-ELA Large 50d RFDeep-ELA Large 50d kNNPCT (100)PGFL RFPGFL kNN
67.42 ± 0.007.93 ± 0.359.58 ± 0.677.70 ± 0.479.65 ± 0.637.09 ± 0.329.98 ± 0.857.12 ± 0.319.99 ± 0.566.87 ± 0.486.69 ± 0.759.75 ± 0.32
129.04 ± 0.009.96 ± 0.3813.36 ± 0.819.71 ± 0.3712.67 ± 1.46----10.13 ± 0.359.42 ± 0.3612.61 ± 0.36
306.20 ± 0.006.20 ± 0.007.26 ± 0.236.20 ± 0.007.15 ± 0.30----6.50 ± 0.185.67 ± 0.157.13 ± 0.21

PGFL uses 25d25d observations, with results averaged over three pretrained encoder seeds within each sampling repeat. “-” denotes unsupported Deep-ELA dimensions.

PGFL remains competitive with the evaluated methods on this application benchmark, providing additional evidence of representation transfer beyond BBOB.

As acknowledged in Appendix E.1 and E.4, the evidence remains limited to the evaluated problem settings. Future work will broaden pretraining through diverse function generators and application-derived optimization traces, and evaluate higher-dimensional, noisy, constrained, and mixed-variable problems.

Q4: We evaluated two variants, each retaining one alignment branch and all three privileged objectives.

VariantHLP Overall ↑Multi. ↑Global ↑Funnel ↑RF relERT ↓kNN relERT ↓
PGFL0.97 ± 0.010.95 ± 0.010.95 ± 0.011.00 ± 0.0111.80 ± 1.3713.39 ± 2.36
w/o instance alignment0.92 ± 0.000.87 ± 0.000.89 ± 0.000.99 ± 0.0012.87 ± 0.1914.82 ± 1.27
w/o sampling alignment0.90 ± 0.000.85 ± 0.010.85 ± 0.010.99 ± 0.0011.86 ± 0.3814.83 ± 1.23
w/o multi-view0.86 ± 0.010.80 ± 0.020.78 ± 0.020.99 ± 0.0012.79 ± 1.2516.23 ± 4.04

Both single-branch variants scored higher on HLP than the w/o multi-view variant. Their relative strengths depend on the metric. The w/o instance alignment variant scored higher on HLP, and the w/o sampling alignment variant achieved lower RF relERT.

Sampling alignment encourages consistency across different samples of the same function instance. Instance alignment encourages the representation to capture structure shared by transformed instances of the same base function.

Q5: We compared the reported PGFL results with a randomly initialized encoder of the same architecture. In both cases, we froze the encoder and fitted the downstream predictors on its representations.

VariantHLP Overall ↑Multi. ↑Global ↑Funnel ↑RF relERT ↓kNN relERT ↓
PGFL0.97 ± 0.010.95 ± 0.010.95 ± 0.011.00 ± 0.0111.80 ± 1.3713.39 ± 2.36
Random frozen encoder0.85 ± 0.010.78 ± 0.010.77 ± 0.010.99 ± 0.0015.81 ± 2.2418.83 ± 3.54

The pretrained encoder performed better on both HLP and AAS, suggesting that PGFL pretraining improves the representation for both downstream tasks.

Thank you for the rebuttal and the additional experiments. I appreciate the authors’ efforts to address the review. I have no further comments and will maintain my original rating.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions and Responses

We thank the reviewers for their constructive feedback. Below we summarize the manuscript revisions and address the main concerns shared across the reviews. Blue text identifies the main additions and clarifications in the revised paper.

1. Manuscript Revisions

LocationChangesReviewers
Sec. 4.4, Table 4; Sec. 3.4Added random-frozen and multi-view-only baselines; clarified frozen-encoder evaluation.u7CN Q1/Q5; ZJVW Q1; y5Rf Q1
Sec. 3.2; App. D.2Clarified instance alignment and separately ablated the two alignment branches.u7CN Q4; ZJVW Q2
App. D.2Added shuffled-target and low-budget-teacher controls for privileged guidance.u7CN Q1
App. D.3Evaluated structural-attribute predictability against a dimension-conditioned median baseline.u7CN Q2
Sec. 4.3; Apps. C.2, D.4, E.4Added UAV transfer in 6, 12, and 30 dimensions and clarified the evaluation scope.u7CN Q3; y5Rf Weakness 2

2. Responses to Shared Concerns

Contribution of pretraining and privileged guidance.

We compared frozen encoders with the same descriptors and architecture. Multi-view only uses the same pretraining functions and retains both alignment objectives, while disabling all three privileged objectives. Downstream predictors follow the same evaluation protocol.

VariantHLP Overall ↑RF relERT ↓kNN relERT ↓
Random frozen encoder0.85 ± 0.0115.81 ± 2.2418.83 ± 3.54
Multi-view only0.91 ± 0.0113.56 ± 1.4915.36 ± 2.13
PGFL0.97 ± 0.0111.80 ± 1.3713.39 ± 2.36

These comparisons support the benefit of pretraining and the additional contribution of privileged guidance. Appendix D.2 further retains the auxiliary heads and loss forms while shuffling targets or reducing the teacher budget. These controls yield higher mean relERT than removing the corresponding objective, supporting a contribution from correctly paired targets and richer observations beyond simply adding auxiliary losses.

Roles of the two alignment branches.

Separate ablations give HLP Overall scores of 0.90 without sampling alignment and 0.92 without instance alignment, compared with 0.86 without either branch and 0.97 for PGFL (Appendix D.2). Related instances may differ in optimizer response: alignment acts in projection spaces, while behavioral guidance encourages the downstream representation to retain optimizer-response information. Section 3.2 clarifies these roles.

Representation transfer beyond BBOB.

We added UAV path planning in 6, 12, and 30 dimensions, with 400 instances per dimension and eight candidate optimizers. The PGFL encoder remains frozen; downstream selectors are fitted on UAV training folds. The RF-based comparison is shown below; PCT selects directly. Scores are normalized final objective values ×100 (lower is better), reported as mean ± SD.

MethodObservations6D12D30D
SBS—7.42 ± 0.009.04 ± 0.006.20 ± 0.00
ELA + RF25d25d7.93 ± 0.359.96 ± 0.386.20 ± 0.00
ELA + RF50d50d7.70 ± 0.479.71 ± 0.376.20 ± 0.00
Deep-ELA Large + RF25d25d7.09 ± 0.32——
Deep-ELA Large + RF50d50d7.12 ± 0.31——
PCT1006.87 ± 0.4810.13 ± 0.356.50 ± 0.18
PGFL + RF25d25d6.69 ± 0.759.42 ± 0.365.67 ± 0.15

PGFL shows advantages in several evaluated settings, supporting representation transfer to this application benchmark. These objective-quality scores exclude observation cost. Full RF/kNN results and seed aggregation are provided in Appendix D.4.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.