acceptodds
Under review as a conference paper at ICLR 2027

In-Context Batch Black-Box Optimization via Flow Matching

Abstract

Batch black-box optimization is important for expensive systems where many candidates can be evaluated in parallel, yet only a few rounds of evaluation are affordable. In this regime, effective batch construction must identify multiple promising regions while avoiding redundant evaluations, and conventional Bayesian optimization can require increasingly difficult batch-selection procedures as the batch size grows. We introduce Flow Matching for Thompson-Sampling Refinement (FMTR), an amortized approach that learns offline a continuous, history-conditioned proposal distribution from which variable-size batches can be sampled. FMTR is trained offline by conditional flow matching on Thompson-sampling argmaxes refined toward high-value regions via a horizon-dependent Langevin pushforward map, adapting the exploration-exploitation tradeoff over the optimization horizon. At test time, the model parameters remain fixed and batch generation requires neither task-specific surrogate refitting nor numerical acquisition optimization. FMTR achieves strong performance against batch BO and in-context baselines on synthetic and real-world benchmarks, while substantially reducing the computational cost of batch selection.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Batch Black-Box Optimization, Bayesian Optimization, Flow Matching, In-Context 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_JaW2OpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes FMTR, a batch black-box optimization method that learns a flow model for proposal distribution, conditioned on previous observations and remaining budget. The flow model is pretrained with data that covers many generated synthetic objectives and trajectories, on a value-weighted target distribution derived from Thompson sampling, and is reusable at inference time, conditioned in-context on observations. FMTR is shown to generally outperform many batch black-box optimization methods across synthetic and real-world functions. Ablations on various design decisions are also included.

Critical strengths
  • The work is novel, pretraining a single flow model for an in-context conditional proposal distribution (across multiple dimensions and batch sizes), compared to online fitting which existing (generative-based) works do. Many design decisions key to the pretraining, such as the construction of synthetic objectives and tilted target, are also detailed and supported by ablation studies.
  • Good empirical results for FMTR. Best final average ranks and faster wall-clock time than existing batch BBO methods.
Critical weaknesses
  • The reason for FMTR (without Langevin and tilt) outperforming qTS is still unclear, given that TS targets were used to train the flow model. I wonder if an analysis of the different priors used by each method could be informative. Especially since there seems to be a center bias for synthetic objectives generated for pretraining (Section B.2, center for ftrendf_\text{trend} sampled from U(−0.375,0.375)\mathcal{U}(-0.375,0.375) with p=0.75p = 0.75).
Questions for authors

Clarifications:

  • For Eq. 10, what is TT, in TminT_\text{min}, TmaxT_\text{max}, and T(h,H)T(h,H)? Temperature? Can they be introduced briefly in words, when first used?
  • Have the authors tried a constant TT (and constant Kh,HK_{h,H})? For comparison with the scheduled TT to see how much the round-awareness matters.
  • FMTR implicitly represents the prior in the flow model to approximate the posterior distribution (if I'm understanding correctly). Is it possible to observe this prior to compare with qTS's prior, through the flow model without conditions or via pretraining data? Also, what is the prior used for qTS?
4 comments
AuthorsOpenReview ↗

Rebuttal 1/2

We thank the reviewer for their thoughtful feedback on our methodology, which helped us clarify key design choices and strengthen the manuscript.

Q1: FMTR implicitly represents the prior in the flow model to approximate the posterior distribution. Is it possible to observe this prior to compare with qTS's prior, through the flow model without conditions or via pretraining data? Also, what is the prior used for qTS?

As derived in Appendix D.3 of the revised manuscript, for a fixed pair (D,α)(\mathcal{D}, \alpha), FMTR is trained to approximate ptarget(x∣D,α)=Ep(τ,c∣D,α)[p~1(x∣τ,D,c)],p_{\mathrm{target}}(x \mid \mathcal{D}, \alpha) = \mathbb{E}_{p(\tau, c \mid \mathcal{D}, \alpha)}\big[\tilde p_1(x \mid \tau, \mathcal{D}, c)\big], where τ\tau denotes the synthetic task, c=(h,H)c = (h, H), and p~1\tilde p_1 is the tilted target in Equation 13. The pretraining prior enters through p(τ,c∣D,α)p(\tau, c \mid \mathcal{D}, \alpha). FMTR therefore implicitly represents its prior, but its target is a task-averaged refined proposal rather than a posterior over the maximizer.

Regarding qTS's prior, qTS uses the default BoTorch SingleTaskGP: an RBF kernel with lengthscale hyperprior ℓi∼LN(2+12log⁡d, 3)\ell_i \sim \mathcal{LN}\big(\sqrt{2} + \tfrac{1}{2}\log d,\ \sqrt{3}\big), a noise-variance hyperprior LN(−4,1)\mathcal{LN}(-4, 1), and a constant mean on standardized outputs.

Finally, FMTR cannot be conditioned on an empty history, because its decoder cross-attends to the context embedding (Figure S8). We therefore compare the priors through the distribution of maximizer locations they induce. For each prior, we draw 1000 independent 2D functions and record each maximizer; for qTS, lengthscales are sampled from the hyperprior above. Figure S15 (Appendix A.11) shows that both distributions spread over the whole search space. Thus FMTR's pretraining prior does not show a pronounced center bias. Trend centers are drawn from U(−0.375,0.375)\mathcal{U}(-0.375, 0.375) only with probability 0.750.75 and from U(−0.8,0.8)\mathcal{U}(-0.8, 0.8) otherwise, and the GP component further displaces the maximizer.

Q2: There seems to be a center bias for synthetic objectives generated for pretraining (Section B.2, center for ftrendf_{trend} sampled from U(−0.375,0.375)\mathcal{U}(-0.375, 0.375) with p=0.75p=0.75).

To address this follow up question, we evaluate on shifted versions of Ackley, Levy and Rosenbrock. We keep each objective unchanged and shrink the search box so that the global minimizer x⋆\mathbf{x}^\star is no longer near the center: Ackleyshift uses [−32.768,2]d[-32.768,2]^d (x⋆=0\mathbf{x}^\star=\mathbf{0}), Levyshift uses [−10,2]d[-10,2]^d (x⋆=1\mathbf{x}^\star=\mathbf{1}), and Rosenbrockshift uses [−5,2]d[-5,2]^d (x⋆=1\mathbf{x}^\star=\mathbf{1}). In the normalized [−1,1]d[-1,1]^d box that FMTR sees, the optimum lies at about +0.88+0.88, +0.83+0.83 and +0.71+0.71 per coordinate, well outside the central region favored by the prior. The result is shown in Figure S6. We can clearly see that FMTR still achieves good performance in this scenario.

Q3: For Eq. 10, what is TT, in TminT_{min}, TmaxT_{max}, and T(h,H)T(h,H)? Temperature? Can they be introduced briefly in words, when first used?

T(h,H)T(h,H) is a temperature controlling how strongly the target favors exploitation: it sets both the tilt strength in Equation 13 (lower TT concentrates mass on high-value regions) and the number of refinement steps Kh,HK_{h,H} in Equation 10. It decreases linearly from Tmax⁡=1.0T_{\max}=1.0 in the first round to Tmin⁡=0.4T_{\min}=0.4 in the last. We added a brief introduction of these parameters in the revised manuscript.

AuthorsOpenReview ↗

Rebuttal 2/2

Q4: The reason for FMTR (without Langevin and tilt) outperforming qTS is still unclear, given that TS targets were used to train the flow model. I wonder if an analysis of the different priors used by each method could be informative.

We agree that the difference lies in the priors. Without refinement and tilt, FMTR amortizes Thompson sampling under its pretraining prior, a mixture of kernel families and randomized hyperparameters with a global trend component. During pretraining, the TS targets are constructed using the known generating kernel and hyperparameters of each synthetic task. Conversely, qTS performs Thompson sampling under a stationary GP whose hyperparameters are estimated online from few observations. The two are therefore TS under different priors, and their relative performance should depend on which prior better matches the test functions.

To test this, we evaluated both methods on functions drawn from qTS’s own GP prior (RBF kernel, lengthscale ∼U(0.5,0.8)\sim \mathcal{U}(0.5, 0.8), no trend). Since the exact value of the sampled lengthscale leading to a function draw to be optimized can be retrieved, we gave it to the GP surrogate powering qTS, leaving it frozen during the trial. In this case, qTS is perfectly well-specified. On the contrary, such function draws occupy only a small part of FMTR’s broader pretraining prior (FMTR's lengthscale range is [0.01,0.5][0.01, 0.5], and includes trends). In this setting, qTS outperforms the TS-only FMTR variant (See Figure S5, Appendix A.5).

These results suggest that FMTR's advantage on our original benchmarks stems primarily from differences in prior specification, rather than from amortization itself. FMTR is pretrained on a diverse mixture of kernels, hyperparameters, and global trends, whereas qTS relies on a stationary GP with hyperparameters estimated from limited observations. When the GP prior is correctly specified, qTS outperforms the TS-only FMTR variant, as expected. Needless to say that this is seldom the case, and in the limited observations regime considered in this submission, learning hyperparameters seems wasteful. We have added this experiment and discussion in Appendix A.5.

Q5: Have the authors tried a constant TT (and constant Kh,HK_{h,H})? For comparison with the scheduled to see how much the round-awareness matters.

We thank the reviewer for this excellent suggestion. We trained a variant (scheduler D) with constant temperature T=0.7T = 0.7, and hence a constant refinement budget Kh,H=⌈K Tmax⁡/T⌉=10K_{h,H} = \lceil K\,T_{\max}/T \rceil = 10 in every round. Since its target no longer depends on the remaining horizon, we also removed α\alpha from the model input. We chose T=0.7T = 0.7 because it equals the mean temperature of scheduler A over the horizon, so this variant isolates the effect of the schedule's shape at a matched exploitation level. The result is shown in Figure S11 (Appendix A.8).

Averaged over dimensions, scheduler D outperforms schedulers B and C but remains behind scheduler A. The gap between A and D grows with dimension: D performs on par with A for d∈3,5d \in \\{3, 5\\}, whereas A clearly outperforms it for d∈10,15d \in \\{10, 15\\}. We draw two conclusions.

First, even without explicit horizon conditioning, FMTR's target adapts over the rounds: as observations accumulate, the posterior contracts and the target concentrates on high-value regions. This explains why a constant temperature is already competitive, especially in low dimensions, where this contraction is fast.

Second, explicit round-awareness provides an additional gain when the schedule is appropriate (scheduler A), particularly in higher dimensions, where the history alone constrains the posterior less. At the same time, the other schedules (B, C) do not outperform the constant temperature, showing that performance depends on how exploitation is distributed across the rounds, not only on whether the target depends on the horizon. The main gains come from refinement and value weighting (Fig. 4b); the temperature schedule is a secondary but beneficial component whose effect grows with dimension.

Thank you for the detailed response. I appreciate the additional experiments on shifted synthetic functions and constant temperature, the further analysis of FMTR without refinement and tilt, and the prior visualizations. I will maintain my positive score.

One question on the prior visualization (Figure S15): how were the argmax points computed? Were they from a discrete candidate set? Visually, the two subplots seem to have several points at matching spots.

AuthorsOpenReview ↗

Thank you for the response

Dear Reviewer JaW2,

Thank you for your response. We sincerely appreciate your time and consideration.

Regarding your question about the prior visualization, we computed the argmax locations by evaluating each sampled function at 3,000 points and selecting the points with the highest value. Although this approach uses a discrete set of points, the sampling density is sufficient for the two-dimensional domain [−1,1]2[-1,1]^2. We will clarify this procedure in the revised manuscript.

Best regards,

The authors

Reviewer_8tUgOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2(+1)
Presentation
2
Soundness
2
Read review
Paper summary

The paper does black-box optimization using conditional flow matching to learn distributions of promising samples derived from Thompson-sampling solutions, which are then further refined toward high-value regions. So its focus is on distilling an optimization heuristic over (synthetic) task families, not quite a new optimization technique. The training targets use privileged information such as gradients, so they may not be good for blackbox problems in general beyond synthetic problems (or at least not sample efficient if the gradient has to be estimated for offline training). The evaluations are on standard analytic functions and four very low-dimensional data-derived benchmarks.

Critical strengths
  • Test-time generation does not need task-specific surrogate refitting
  • The use of closed-loop pretraining is reasonable for adapting exploration versus exploitation over rounds.
Critical weaknesses
  • It mainly learns to reproduce refined Thompson-sampling proposals using flow matching, while related in-context and generative BO methods already exist, so novelty is limited.
  • Training uses privileged information that may be hard to assume in general BBO problems.
  • Evaluation is too limited to standard synthetic functions and problem with very low dimensions.
Questions for authors
  1. Do you have much higher-dimensional benchmarks with analysis on the results that show advantage of the proposed techniques specifically?
  2. How important is the gradient information used to construct the training targets?
  3. Would modeling interactions among points within each batch improve optimization performance the current independent sampling scheme?
5 comments
AuthorsOpenReview ↗

Rebuttal 1/2

We thank the reviewer for their thoughtful assessment and constructive criticism. We provide responses to their comments below.

Q1: It mainly learns to reproduce refined Thompson-sampling proposals using flow matching, while related in-context and generative BO methods already exist, so novelty is limited.

While refined Thompson-sampling proposals form the basis of FMTR's training targets, we believe that our contribution extends beyond, spanning horizon-dependent construction, value weighting, and more importantly, amortization for continuous batch optimization. We provide more details below.

FMTR learns, in context, a proposal distribution over continuous query locations, conditioned on the observed history and remaining budget. At deployment, batches of any size are then drawn from this proposal with fixed trained parameters. As shown in Figure 1, existing methods don't fully tackle this problem. In-context optimizers (NAP, PABBO, TAMO, FICBO) learn single-query policies over finite candidate sets. In-context surrogate/acquisition models (PFNs4BO, GIT-BO, α\alpha-PFN) amortize the surrogate modeling but still require acquisition optimization or candidate searching at each round. Whereas, generative BO methods (DiBO, GenBO) still refit their generator on each task as observations accumulate. None of them provides a pretrained history-conditioned sampler from which continuous batches can be drawn directly.

Extending reward-based training to this setting is non-trivial, as each action consists of a set of continuous query locations whose joint contribution must be appropriately credited, requiring careful reward design. Instead, we directly specify a well-defined training target using privileged information available only during pretraining. This provides an effective alternative to reward-based training while preserving the explore-early, exploit-late behavior. Empirically, FMTR outperforms existing baselines while requiring substantially less inference time, demonstrating the effectiveness of our approach.

Q2: Training uses privileged information that may be hard to assume in general BBO problems.

We want to clarify that FMTR does not assume privileged information about the target black-box problem. Privileged information is used only during offline pretraining, on synthetic tasks sampled from our prior, for which it is available by construction: ∇f\nabla f (for Langevin refinement) and the noise-free objective values f(x)f(x) (for reweighting loss function). At deployment, FMTR conditions only on the observed history and the remaining horizon fraction. It requires no gradients, model refitting, or other information about the target function. Regarding the effectiveness of pretraining on synthetic tasks, this follows a well-established paradigm. Prior-Fitted Networks [1,2] pretrained on synthetic tasks are outperform strong baselines on tabular benchmarks. Likewise, in-context or amortized methods for Black-box optimization [3,4,5] pretrain on synthetic tasks (usually GP draws) and show competitive results on a wide range of optimization problems. FMTR's additional step is to exploit information that is available in simulation but not at test time to construct better training procedure. This idea is known as Learning Using Prvileged Information [6] and is effectively used in robotics [7]. Finally, it's also worth to note that gradient-based refinement is just one instantiation of this idea. As discussed in Appendix C.2, it can be replaced by a gradient-free refinement scheme when pretraining on task families whose gradients are unavailable. This shows the flexibility of our method.

[1] Hollmann et al, Accurate predictions on small data with a tabular foundation model, Nature 2025

[2] Muller et al, Position: The future of bayesian prediction is prior-fitted, ICML 2025

[3] Maraval et al, End-to-end meta-Bayesian optimisation with transformer neural processes, NeurIPS 2023

[4] Zhang et al, In-context multi-objective optimization, ICLR 2026

[5] Zhang et al, PABBO: Preferential amortized black-box optimization, ICLR 2025

[6] Vapnik et al, Learning using privileged information: Similarity control and knowledge transfer, JMLR 2015

[7] Pinto et al, Asymmetric actor critic for image-based robot learning, arXiv preprint arXiv:1710.06542, 2017

AuthorsOpenReview ↗

Rebuttal 2/2

Q3: Evaluation is too limited to standard synthetic functions and problem with very low dimensions. Do you have much higher-dimensional benchmarks with analysis on the results that show advantage of the proposed techniques specifically?

Our work targets the large-batch, few-round regime, where the primary challenge is generating informative batches within a limited number of rounds, rather than scaling to extremely high-dimensional spaces. In this setting, d=15d=15 is already challenging. For example, selecting a batch of qq candidates requires GP-based methods to optimize an acquisition function over qdqd variables, which becomes increasingly challenging as the batch size qq grows. Or in few round regime, the total budget of evaluations can only provide sparse coverage of the search space. In contrast, very high-dimensional BO typically requires additional techniques, such as low-dimensional embeddings or trust regions, which are beyond the scope of this work.

Nevertheless, we investigated scalability to higher dimensions, retraining FMTR with a maximum dimension of D=30D=30. The model size is unchanged, but the training budget was doubled. We then evaluated the resulting model on three real-world benchmarks with d∈25,26,30d\in\\{25,26,30\\}. The result is shown in Figure S3 in Appendix A.3. In general, FMTR does not maintain its competitive performance in this setting. We acknowledge that scaling FMTR to higher dimensions remains an important direction for future work and have explicitly discussed this limitation in the manuscript.

Q4: How important is the gradient information used to construct the training targets?

An ablation study tackling this question was already present in the original manuscript. We isolated the contribution of gradient information through the no Langevin variant in Section 5.2 (Effect of Langevin Refinement and Distributional Tilt, Fig.4(b)). This variant removes gradient-based Langevin refinement while retaining value weighting, which relies solely on objective values. Its performance clearly degrades compared with the full FMTR, demonstrating that gradient information substantially improves the quality of the training targets. Nevertheless, the no Langevin variant still outperforms both the variant without refinement and distributional tilt and qTS, indicating that FMTR's improvements do not rely solely on gradient information.

Q5: Would modeling interactions among points within each batch improve optimization performance the current independent sampling scheme?

FMTR samples batch members independently given the history and remaining budget, without explicitly modeling within-batch redundancy. To investigate whether modeling such interactions helps, we introduce FMTR+DPP, which applies a determinantal point process (DPP) during training and inference. DPP is a standard approach for modeling repulsion in batch BO [1,2].

To incorporate DPP into sampler (p1~\tilde{p_1} or FMTR), we first draw a pool of mqm q proposals from the sampler and then select qq candidates using greedy MAP inference [3] with the kernel L=S⊙(vvT)L = S \odot (\mathbf{v} \mathbf{v}^T); e.g Lij=viSijvj,L_{ij} = v_i S_{ij}v_j,where Sij=exp⁡(−∣xi−xj∣2/2ℓ2)S_{ij}=\exp(-|x_i-x_j|^2/2\ell^2) measures the RBF similarity between proposals, encouraging diversity by discouraging the simultaneous selection of similar candidates. The length scale ℓ\ell is adjusted according to the remaining budget, ℓ=ℓ0α\ell=\ell_0\alpha, where α=h/H\alpha=h/H, encouraging broader exploration early on and more concentrated sampling as the budget decreases. We also consider an optional quality term [4], vi=(Sˉi/max⁡jSˉj)βv_i=(\bar S_i/\max_j\bar S_j)^\beta, where Sˉi\bar S_i denotes the average similarity of proposal ii to the other proposals in the pool. It favors proposals in regions where the model placed many of its own draws. β=0\beta=0 means pure diversity, and larger β\beta keeps the batch away from outlying draws. Setting m=1m=1 recovers independent sampling without post-hoc selection.

The result are shown in Figure S4 (Appendix A.4). We set m=4m=4, ℓ0=0.3\ell_0=0.3, β=0.5\beta=0.5 for DPP. We report both Log simple Regret and Top 10 Log simple regret metric. FMTR+DPP selection gave no improvement over independent sampling . So post-hoc diversification for within-batch interaction using DPP is not sufficient. Consequently, a fuller treatment is needed, for example by generating batch members sequentially and conditioning each on the earlier ones. We leave this to future work and mentioned it our Discussion section.

[1] Kathuria et al, Batched gaussian process bandit optimization via determinantal point processes, NeurIPS 2016

[2] Nava et al, Diversified sampling for batched bayesian optimization with determinantal point processes, AISTATS 2022

[3] Chen et al, Fast greedy map inference for determinantal point process to improve recommendation diversity, NeurIPS 2018

[4] Morshed et al,. Diverseflow: Sample-efficient diverse mode coverage in flows, CVPR 2025

AuthorsOpenReview ↗

Official Comment by Authors

Dear Reviewer 8tUg,

With the discussion period ending in approximately two days, we would greatly appreciate any further feedback on our rebuttal. We believe we have carefully addressed your concerns and would welcome your assessment of our responses. If you find that we have satisfactorily resolved your main concerns, we would respectfully appreciate your consideration of an updated score.

Thank you again for your time and consideration.

Best regards,

The Authors

Reviewer_Ta9aOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes FMTR, an amortized batch black-box optimizer. Offline, the authors pretrain a history-conditioned conditional flow-matching model on synthetic episodes. Each episode uses a random-Fourier-feature GP sample plus a quadratic trend. Training targets are Thompson-sampling (TS) argmaxes refined by a horizon-dependent Langevin-style procedure and reweighted by a Gibbs-style value tilt. At deployment, the model parameters are fixed, and a batch is generated by sampling from the learned proposal given the history and the remaining-round fraction, with no surrogate refitting and no acquisition optimization. The method is evaluated on Ackley, Levy, and Rosenbrock (d ∈ {3, 5, 10, 15}, q = 64; Figure 2), four data-derived real-world tasks (oil-4D, hpobenchXGB-6D, Asteroid-12D, pm25-14D; Figure 3), and ablations on the pretraining prior, refinement and tilt, and temperature schedule (Figure 4). Baselines are TuRBO, BEEBO, DiBO, GIT-BO, qLogEI, qUCB, and qTS-Matheron. The paper reports the lowest runtime (Table S1, measured on Ackley only).

Critical strengths
  • The algorithmic framing is clear and well motivated. Amortizing batch construction into a sampler over continuous queries, rather than optimizing a joint acquisition over qd variables, is a sensible response to the scaling difficulty of batch BO.
  • The ablation studies are substantial and separate the contributions of the main technical components. Figure 4 isolates the effect of the trend term in the pretraining prior, the Langevin refinement and distributional tilt, and the temperature schedule, which makes it easier to see which design choices drive the gains.
Critical weaknesses
  • Related work is limited and does not position the contribution against existing flow-matching methods in BO. The related work section does not discuss prior work that also uses flow matching within the BO loop, such as Zhai, Jeong, and Rockova (AISTATS 2026) [1]. Positioning FMTR relative to that work would clarify what is new here, whether in the choice of target distribution, the amortization, or the batch-generation procedure.
  • Important baselines are discussed but not compared experimentally. NAP [2] is cited as a central end-to-end in-context optimizer and is shown in Figure 1 as a key reference point for methodological novelty, yet it is absent from all experiments. Since NAP acts on candidate sets, the authors should adapt it to continuous domains, for example by supplying a dense Sobol candidate pool at each round, which is a standard and inexpensive adaptation. Including it would show directly whether FMTR’s continuous sampler offers an advantage over a strong amortized baseline.

[1] Percy S. Zhai, Sowon Jeong, Veronika Rockova Proceedings of The 29th International Conference on Artificial Intelligence and Statistics, PMLR 300:2044-2052, 2026.

[2] Alexandre Maraval, Matthieu Zimmer, Antoine Grosnit, and Haitham Bou Ammar, End-to-End Meta-Bayesian Optimisation with Transformer Neural Processes, NeurIPS 2023

Questions for authors
  • Does the model’s use of observation order and remaining horizon in the dataset matter? The paper states that no positional encodings are used, so the history is treated as a set. Have the authors checked whether adding order information changes performance, or whether the set-based design was validated empirically?
  • Why is the temporal information limited to the remaining-round fraction α and the context size |D|? Have the authors considered whether the model could learn round-dependent behaviour without α, and would an ablation removing α show that it is needed?
3 comments
AuthorsOpenReview ↗

Rebuttal 1/2

We thank reviewer for positive assessment of our work and the points you raised. We address your points and questions below.

Q1: The related work section does not discuss prior work that also uses flow matching within the BO loop, such as [1]. Positioning FMTR relative to that work would clarify what is new here, whether in the choice of target distribution, the amortization, or the batch-generation procedure.

We thank the reviewer for pointing us to this work. Upon reading the paper, it appears that [1] use conditional flow matching for likelihood-free posterior inference over model parameters, in line with flow matching approaches to simulation-based inference [2]. While our approach, FMTR, shares this methodological mechanism of conditional flow matching trained on simulated pairs, its target is different. Indeed, FMTR learns ptarget(x∣D,α)=∫p~1(x∣D,c)p(τ,c∣D,α)dτdc,p_{\text{target}}(x \mid \mathcal{D}, \alpha) = \int \tilde{p}_1(x \mid \mathcal{D}, c) p(\tau,c \mid \mathcal{D}, \alpha)d\tau dc, where c=(h,H)c=(h,H) and p~1\tilde{p}_1 is defined in Equation 13 of the main text (a detailed derivation is provided in Appendix D.3 of the revised version). This target is neither a posterior over model parameters nor over the maximizer. It is a family of proposals indexed by the remaining horizon α\alpha, designed for optimization rather than inference, and its samples are used directly as query batches in a sequential loop, which is not a case for posterior inference. We have added [1] and related flow-based SBI methods in the revised paper to better clarify our contribution.

Q2: Important baselines are discussed but not compared experimentally. NAP [3] is cited as a central end-to-end in-context optimizer and is shown in Figure 1 as a key reference point for methodological novelty, yet it is absent from all experiments.

We thank the reviewer for this suggestion and agree that FMTR should be compared with an end-to-end in-context optimizer.

NAP was not included for two reasons. First, NAP is meta-trained on tasks from the target task family (e.g., HPO, antibody design), with a separate model for each search space. It therefore relies on related tasks being available for meta-training and is not designed to generalize across dimensions. Second, NAP is designed for sequential single-query optimization over a finite candidate set, not for continuous batch selection. A fair comparison would therefore require retraining NAP on our synthetic prior and adapting it to batch selection, which would effectively change the method and could confound the comparison.

Instead, we have added a comparison with TAMO [4], a pretrained end-to-end in-context optimizer like NAP. Unlike NAP, however, its architecture supports varying input dimensions and numbers of objectives, and its released checkpoint is pretrained on synthetic data across input dimensions 1--5 and supports a single objective. This allows us to evaluate it exactly as released.

Following the reviewer's suggestion, TAMO selects from a denser candidate pool than in its default evaluation (2142^{14} rather than 2112^{11} Sobol points), and batches are built with the kriging believer strategy described in the TAMO paper. Results are shown in Figure S1, with the setup in Appendix A.1. FMTR consistently outperforms TAMO on the Ackley, Levy, and Rosenbrock function (d=2−5d = 2-5), and the gap widens as the input dimension increases.

[1] Zhai et al, Conditional flow matching for bayesian posterior inference, AISTATS 2026

[2] Wildberger et al, Flow matching for scalable simulation-based inference, NeurIPS 2023

[3] Maraval et al, End-to-end meta-Bayesian optimisation with transformer neural processes, NeurIPS 2023

[4] Zhang et al, In-context multi-objective optimization, ICLR 2026

AuthorsOpenReview ↗

Rebuttal 2/2

Q3: Does the model’s use of observation order and remaining horizon in the dataset matter? Have the authors checked whether adding order information changes performance, or whether the set-based design was validated empirically?

We thank the reviewer for this question. Our set-based design follows from the structure of FMTR’s training target. For a given task, the target p~1(⋅∣τ,D,c)\tilde p_1(\cdot \mid \tau, \mathcal{D}, c) (Eq. 13) depends on the history only through the GP posterior pτ(f∣D)p_\tau(f \mid \mathcal{D}), from which the Thompson-sampling maximizers are drawn; refinement and value weighting depend on the objective, not on the history. Since observations are conditionally independent given the objective; e.g pτ(f∣D)∝pτ(f)∏i=1np(yi∣f,xi)p_{\tau}(f \mid \mathcal{D}) \propto p_{\tau}(f)\prod_{i=1}^np(y_i \mid f, x_i), this posterior, and hence the target, is invariant to permutations of the observed pairs. The only temporal information the target requires, the remaining budget, is provided explicitly to the model through α=h/H\alpha = h/H instead of being encoded through observation order.

This is consistent with the belief-state view of sequential decision-making, in which the Bayes-optimal policy depends only on the posterior and the remaining horizon (POLAR [1], Sec. 3), and with permutation-invariance results for amortized experimental design (DAD [2], Theorem 3). Set-based history encoders are also standard in in-context optimization (e.g., [NAP, PABBO, TAMO, FICBO])

Q4: Why is the temporal information limited to the remaining-round fraction α\alpha and the context size ∣D∣|D|? Have the authors considered whether the model could learn round-dependent behaviour without α\alpha, and would an ablation removing α\alpha show that it is needed?

First, we would like to point out that FMTR conditions on the full observation history D\mathcal{D}, not only its size, and on the remaining-round fraction α\alpha. Indeed, these inputs serve different purposes: the history informs the model about the objective, while α\alpha specifies the remaining evaluation budget. From finite-horizon Bayesian decision theory, the Bayes-optimal acquisition policy generally depends on both the posterior belief p(f∣D)p(f\mid\mathcal{D}) and the remaining horizon hh, since the same observations may warrant different exploration-exploitation trade-offs depending on how many rounds remain. A related argument is developed in [1, Section 3], with the explicit horizon dependence following from the finite-horizon Bellman recursion. Moreover, our pretraining varies the initial context size n0n_0, batch size qq, and total horizon HH, so ∣Dr∣=n0+rq|\mathcal{D}_r|=n_0+rq does not uniquely determine the remaining horizon.

This being said, we agree that the model can learn changing behaviour from the evolving history without explicitly receiving α\alpha. To assess the benefit of this additional input, we trained a variant without α\alpha in the model conditioning. Figure S2 in Appendix A.2 compares this variant with FMTR on Ackley, Rosenbrock, and Levy at d∈10,15d\in\\{10,15\\} and q=64q=64. Both variants improve as observations accumulate, but FMTR achieves lower mean final log simple regret in all six settings, with the differences mainly emerging in later rounds. This suggests a benefit for explicitly informing the model of the remaining budget.

[1] Huang et al, Efficient Adaptive Data Acquisition via Pretrained Belief Representations, NeurIPS 2026

[2] Foster et al, Deep Adaptive Design: Amortizing Sequential Bayesian Experimental Design, ICML 2021

AuthorsOpenReview ↗

Official Comment by Authors

Dear Reviewer Ta9a,

With the discussion period ending in approximately two days, we would greatly appreciate any further feedback on our rebuttal. We believe we have carefully addressed your concerns and would welcome your assessment of our responses. If you find that we have satisfactorily resolved your main concerns, we would respectfully appreciate your consideration of an updated score.

Thank you again for your time and consideration.

Best regards,

The Authors

Comments on the paper

2 comments
AuthorsOpenReview ↗

Global comment

We thank the reviewers for their careful assessment and constructive feedback, which have helped us substantially improve the manuscript.

In response to the questions raised, we have added the following experiments to the revised paper:

  • Additional baseline (Appendix A.1): Comparison with TAMO, a pretrained end-to-end in-context optimizer, using its Kriging Believer batch-selection strategy.
  • Horizon conditioning (Appendix A.2): An ablation removing the remaining-round fraction (α)(\alpha) from the model input while retaining the horizon-dependent training targets.
  • Higher-dimensional evaluation (Appendix A.3): Experiments on three real-world benchmarks with 25, 26, and 30 dimensions.
  • Within-batch interaction (Appendix A.4): A determinantal point process (DPP) selection step to assess the benefits of post-hoc within-batch diversification.
  • Prior specification (Appendix A.5): Comparison with qTS on functions drawn from its GP prior, providing qTS with the generating lengthscales and keeping them fixed during optimization.
  • Potential center bias (Figure S6 and Appendix A.12): Comparisons of the maximizer-location distributions induced by the FMTR and qTS priors, and evaluations on benchmarks whose optima lie near the search-domain boundaries.
  • Target scheduling (Appendix A.8): A variant with constant temperature and refinement budget, without explicit horizon conditioning.

Additionally, we expanded the explanation of FMTR's target distribution in Appendix D.3 and added relevant references to simulation-based inference (SBI).

We address each reviewer’s questions and concerns in the individual comments below.

Dear Reviewers,

As the discussion phase draws to a close, please acknowledge the author's response and let them know whether it has addressed some or all of your concerns.

Thank you for the time and effort you and the other reviewers have put into this discussion.

Best,

Your AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.