{"id":"4db4265d-c26b-4e3a-a7cd-05a62656d5a5","arxiv_id":"2608.12687","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"TTARO retrains a shallow neural representation along with the Gaussian process at each Bayesian optimization step, improving sample efficiency on two analog topology benchmarks relative to frozen-embedding baselines.","lead":"This paper tries a simple idea for analog circuit topology search: keep updating the machine-learning representation used by the Bayesian optimizer after each new circuit is tested, instead of freezing it. On two public benchmarks, the approach cuts the average \"regret\" area under the curve by about 15% versus a fixed-representation optimizer, though gains vary by setting.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unreported initial sample size n0 and budget B control the DKL comparison; without these, the headline 20.7% regret-AUC reduction could partly reflect a data-starved DKL baseline rather than online adaptation.","rationale":"The reader's weakest assumption identifies exactly the same load-bearing concern: the missing n0 and B make the DKL baseline's strength ambiguous. This is the single most important issue because the strongest claim is framed as a relative reduction over DKL, and the magnitude of that reduction depends directly on how much initial data DKL is allowed for representation learning. The paper gives substantial evidence for the direction of the effect, with TTARO numerically better in 37 of 40 settings, but direction is not enough to validate the headline magnitude. A DKL model trained on a very small initial set is a weak baseline; TTARO's retraining advantage would then be expected even without any fundamental benefit to online adaptation per se. The proposed test varies n0 while holding everything else fixed, so it isolates whether the reported advantage survives when DKL is given a more generous initial sample. Other gaps, such as missing significance tests and absent code, are real but secondary; they would be fixable alongside the protocol reporting. I would keep the reader's CONDITIONAL verdict: the central idea is plausible and the experiments are broad, but the paper should not be accepted without reporting n0 and B and demonstrating robustness to their choice.","tokens_in":16571,"tokens_out":6113,"duration_ms":61598,"concrete_test":"Re-run the full 40-configuration comparison with B=100 and n0 in {5, 10, 20, 50, 100}, using the same 20 seeds and identical feature-map architectures for DKL and TTARO. Report aggregate TTARO-vs-DKL regret-AUC reduction and the number of configurations favoring TTARO at each n0. If the 20.7% headline falls below roughly 10% or the win count drops below a clear majority when n0 is large, the claimed benefit is not robust to the initial-sample protocol. If the advantage persists at n0=50, the concern is resolved. Also report paired per-seed significance, e.g., Wilcoxon signed-rank, for the n0 value used in the paper.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central comparison is TTARO, which retrains the deep-kernel representation after every evaluation, versus DKL, which freezes its feature map after training only on the initial n0 evaluated circuits. Algorithm 1 lists n0 and B as inputs, but Section V never reports either. Figure 5 and Table III imply B=100 through 'BO budget (%)' and 20% checkpoints, yet n0 is never given. If n0 is small, say 10 or 20, DKL's feature map is learned from very few FoM labels while TTARO eventually trains on all 100, so the reported 20.7% aggregate regret-AUC reduction relative to DKL could be partly a data-quantity effect rather than a specifically online-adaptation effect. This is not an internal inconsistency, but it is the load-bearing unknown for the headline claim: the protocol parameter that determines whether DKL is fairly trained or deliberately starved is absent from the manuscript, and the limitation section discusses only computational overhead.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces TTARO, an online deep-kernel Bayesian optimization procedure for finite-bank analog circuit topology search. Starting from pretrained fixed circuit embeddings, TTARO retrains a two-layer MLP feature map jointly with a GP surrogate at each BO iteration on all FoM labels observed so far, then applies a standard acquisition function to select the next candidate. The method is compared against a fixed-embedding GP, a DKL baseline whose feature map is trained only on the initial sample and then frozen, and a GP-oracle whose feature map is trained on all labels before optimization. Experiments cover two OCB benchmarks, four encoders per benchmark, two kernels, and three acquisition functions (40 settings at 20 seeds), and report regret-AUC reductions of 19.6% over GP on Ckt-Bench-101 and 12.2% on Ckt-Bench-301, which the abstract aggregates to 15.2% vs GP and 20.7% vs DKL.","tokens_in":16711,"tokens_out":9481,"duration_ms":96632,"significance":"If the results hold, the contribution is a simple, general, and practically useful wrapper around standard BO: retraining the deep-kernel representation as FoM labels accumulate improves sample efficiency on a real analog-topology benchmark. The strengths of the paper are the breadth of the evaluation (40 settings x 20 seeds on public benchmarks), the method's agnosticism to encoders, kernels, and acquisition functions, and the clear separation of GP, DKL, and oracle baselines. The main weaknesses are that the headline comparison against DKL depends on unreported protocol parameters (n0 and B), the per-setting gains are not supported by significance tests, and the oracle condition is not consistently an upper bound. These issues are fixable with additional reporting and analysis.","major_comments":[{"comment":"The initial sample size n0 and the total evaluation budget B are never reported, although Algorithm 1 lists both as inputs and Figure 5 and Table III implicitly use a budget expressed in percent. Because the DKL baseline trains its feature map only on the n0 initial FoM labels, an unreported small n0 would make the headline 20.7% regret-AUC reduction relative to DKL partly a data-quantity effect rather than an effect of online adaptation. Please state n0 and B for every configuration, and provide an ablation in which a retrained-DKL baseline receives additional training at the same checkpoints as TTARO (or otherwise controls for the number of training labels) to separate adaptation from increased training data.","section":"V.D and Algorithm 1"},{"comment":"The claim that TTARO 'outperforms prior art in most settings' is not supported by significance testing. In several settings the mean differences are smaller than the standard error, for example Ckt-Bench-101 D-VAE-GCN Linear/TS (29726.3±930.2 vs 31408.0±622.8) and Ckt-Bench-301 DAGNN Linear/EI (11183.7±1472.5 vs 12144.2±1253.1). Because the experiments use paired seeds, report paired tests (e.g., Wilcoxon signed-rank or paired t-test) across the 20 seeds, or at least report per-setting confidence intervals for the regret-AUC reduction, so that the 37-of-40 directional count can be interpreted as statistical evidence.","section":"VI, Tables I and II"},{"comment":"The GP-oracle condition is described as an 'upper-reference condition,' but it is not consistently better than the fixed GP. For example, Ckt-Bench-101 CktGNN RBF/UCB has oracle regret AUC 30246.8±2835.0 vs GP 24013.6±1432.6, and several other oracle rows are worse than the corresponding GP row. If training the feature map on all available FoM labels can degrade BO performance, the paper's motivating claim that objective-aligned geometry improves the surrogate needs qualification (for example, overfitting or optimization difficulty of the MLP), and the term 'upper-reference' should be either justified or revised.","section":"V.C and VI"}],"minor_comments":[{"comment":"The abstract contains LaTeX artifacts such as 'T est-Time' and 'O ptimization'; these should be fixed before publication.","section":"Abstract and Section I"},{"comment":"The statement 'As far as we are aware, no prior art has explored this framework for analog circuit topologies' should be narrowed: Section II.D itself cites representation-adaptive BO methods (SILBO, CoBO, LOCo), so the novelty claim should explicitly be limited to the analog-topology finite-bank setting rather than suggesting no prior art exists for adaptive representations in BO generally.","section":"Section I"},{"comment":"The dropout rate, optimizer, learning rate, number of training epochs, and stopping criterion for the MLP/GP fit are not reported. These details are needed for reproducibility, especially because the feature map is a roughly ten-thousand-parameter MLP trained on at most about one hundred labels.","section":"Section IV.A, Eq. (10)"},{"comment":"The Top-2% KNN@5 diagnostic is computed from the same FoM labels that trained the feature map, so it is not an independent measure of representation quality. The text should state this circularity explicitly rather than only noting that the quantity is 'used only as a diagnostic.'","section":"Section V.D and Figure 4"},{"comment":"Because the budget is reported in percent, the caption should state the absolute evaluation budget B and the initial sample size n0.","section":"Table III"},{"comment":"References [19] and [25] are technical reports or non-standard venues; if published versions exist, the authors should cite those instead.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"I am moderately confident that the method works as claimed, but the manuscript needs protocol transparency and statistical rigor before acceptance. The missing n0 and B is the single most important issue; if the authors supply these values and the paired significance tests are favorable, the paper would be a reasonable contribution to the analog-design and BO communities. The novelty over existing adaptive-embedding BO is incremental but the analog-circuit application and the large systematic evaluation support publication. The 'most comprehensive evaluation' claim is unverifiable as written and could be softened."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know up front. First, the algorithmic contribution is a re-implementation of deep kernel learning retrained after every BO evaluation — the paper says as much, citing Deep Kernel Bayesian Optimization [19]. Second, what is genuinely new is the scale and consistency of the empirical study: 8 encoder–dataset pairings, 5 kernel/acquisition settings, 20 seeds, 40 headline configurations, with TTARO beating fixed-embedding GP in 37 of 40 settings. That is a solid, useful result for anyone doing BO in finite-bank topology search.\n\nThe paper does several things well. It uses the public OCB benchmarks, compares against a reasonable fixed-embedding GP baseline, and includes an oracle upper bound that is properly described as not deployable. The writing is clear about the distinction between DKL (feature map trained once on the initial set) and TTARO (feature map retrained every step). The acknowledgment of the DKL lineage is honest.\n\nThe soft spots are real but fixable. The main one: n0 and B are never reported. Algorithm 1 takes them as inputs, and B can be inferred as 100 from Figure 5 and Table III, but n0 is absent. If n0 is small, the DKL baseline is trained on very few labels while TTARO accumulates labels across all 100 evaluations. That would make the 20.7% regret-AUC reduction relative to DKL partly a training-data effect, not purely an online-adaptation effect. The paper needs to report n0 and, ideally, run DKL with larger initial samples as a robustness check. Second, no significance tests are provided; several standard-error bars overlap, especially in the small-gain configurations. Third, the Fig. 4 Top-2% KNN@5 diagnostic is circular — the labels used to measure FoM concentration are the same labels that trained the feature map. The paper does call it a diagnostic for visualization, so it is not deceptive, but it should not be used as evidence for optimization quality. Fourth, the abstract says 'outperforming prior art,' but the comparison set is GP, DKL, and an oracle, not the SILBO/LOCo/CoBO family the related-work section discusses.\n\nOverall, the central claim — online representation adaptation improves sample efficiency in this domain — is plausible and supported by the direction of the results. The magnitude is less certain. If I were the editor, I would send this to peer review, and in the revision require n0, B, per-seed data or significance tests, and a corrected abstract. The paper is a useful contribution to EDA-oriented BO, but it needs a bit more care before the numbers are citable.","headline":"A competent, unusually broad empirical study of a known idea (retraining DKL at each BO step) applied to analog topology search; the headline gains are plausible but the unreported initial sample size makes the DKL comparison hard to trust.","tokens_in":17331,"tokens_out":3194,"would_cite":true,"duration_ms":30037,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"TTARO claims that online retraining of a neural feature map and Gaussian-process surrogate on every accumulated figure-of-merit label improves Bayesian optimization sample efficiency, cutting regret AUC by 15.2% on average and up to 46.7%.","keywords":["analog circuit design","Bayesian optimization","representation learning","topology synthesis","electronic design automation","online adaptation","deep kernel learning","test-time adaptation"],"falsifier":"Vary the initial sample size $n_0$ from very small to nearly the full budget while holding the evaluation budget fixed, and plot regret AUC for TTARO against the deep-kernel baseline that freezes its representation after the initial sample; if TTARO's advantage shrinks toward zero as $n_0$ grows, the reported gain is mostly extra training data, not online adaptation. Alternatively, give the frozen baseline access to the same total number of training labels that TTARO has at the end of the run and check whether the 20.7% average reduction survives.","tokens_in":16247,"feed_emoji":"⚡","tokens_out":9678,"duration_ms":84191,"temperature":0.7,"pith_summary":"Analog circuit topology search is expensive because each candidate must be simulated before its quality is known, so methods that find good circuits in fewer evaluations are practically valuable. This paper claims that Bayesian optimization over pretrained circuit embeddings leaves sample efficiency on the table, because the embedding geometry is organized by circuit structure rather than by the performance metric being optimized. It introduces TTARO, which after every circuit evaluation retrains a small neural feature map and a Gaussian-process surrogate on all figure-of-merit labels collected so far, and uses that updated geometry to select the next candidate. Across 40 encoder/kernel/acquisition settings on the Ckt-Bench-101 and Ckt-Bench-301 op-amp benchmarks, TTARO reduces regret area under the curve by 15.2% on average relative to fixed-embedding BO and by 20.7% relative to deep kernel learning, with individual reductions as large as 46.7%. If the claim holds, representation-based BO should not freeze its representation: re-aligning the search geometry to the objective is a general and inexpensive route to better acquisition decisions.","feed_headline":"Retraining circuit embeddings mid-search cuts regret up to 46.7%","feed_subtitle":"Refitting the kernel on every new label aligns circuit similarity with performance, so better topologies surface sooner.","key_machinery":"The load-bearing object is the learned feature map $\\phi_{\\theta_t}$, a two-layer multilayer perceptron that sends each normalized fixed embedding $\\bar{h}_i$ to a 16-dimensional adapted representation $z^{(t)}_i = \\phi_{\\theta_t}(\\bar{h}_i)$ at iteration $t$. It is fit together with the Gaussian-process hyperparameters by minimizing the negative log marginal likelihood on the accumulated labels $\\mathcal{D}_t$, and the GP then acts on the adapted coordinates through a scaled linear kernel or a scaled RBF kernel. The mechanism does its work through the covariance: because the feature map enters every off-diagonal entry of the training covariance $K_t$ and every cross-covariance vector $k_{t,j}$, retraining $\\theta_t$ rewires which circuits the surrogate treats as similar, and that rewired geometry is what changes the acquisition function's next pick.","core_discovery":"The paper's central discovery is that the surrogate's kernel geometry, not just the surrogate itself, should be updated during Bayesian optimization. Starting from any pretrained circuit embedding, TTARO normalizes the candidate bank and at each iteration fits a feature map -- a two-layer MLP mapping the normalized embedding into 16 adapted coordinates -- jointly with an exact Gaussian-process surrogate by minimizing the negative log marginal likelihood on the observed figure-of-merit labels. The adapted coordinates replace the original embeddings in the kernel, so every pairwise similarity, posterior mean, posterior variance, and acquisition score changes with each retraining. The comparison against deep kernel learning isolates the mechanism: DKL uses the same surrogate structure but trains its feature map only on the initial evaluated set and then freezes it, and on Ckt-Bench-301 DKL actually degrades relative to plain GP (regret AUC rising from 19,901 to 23,634 and final best FoM falling from 152.7 to 134.0), while TTARO improves both. The paper also reports that TTARO reaches 252.8 FoM by 60% of the budget on Ckt-Bench-101, already above GP's final value of 247.6, and matches GP's final 152.7 on Ckt-Bench-301 by 60% of the budget with 40% of evaluations still unused.","pith_inferences":["A natural extension, not tested in the paper, is that the same online-adaptation recipe transfers to any finite-bank search with pretrained embeddings and expensive labels, such as molecule selection, materials screening, or device sizing.","Because the kernel changes at every iteration, standard GP regret bounds do not directly apply to TTARO; a regret analysis for a sequence of adaptively learned kernels is an open question the paper leaves implicit.","The paper's own diagnostic suggests a testable mechanism: if Top-2% KNN@5 in the adapted space rises monotonically and predicts regret reduction across encoders, then objective-aligned geometry is the cause; if not, the gains may simply come from refitting the surrogate on more data.","A cheap ablation would retrain the feature map on each iteration but with the FoM labels randomly permuted; if the regret reduction persists, the benefit is from surrogate retraining rather than from representation alignment."],"forward_implications":["If TTARO is correct, latent-space BO pipelines for analog design should keep representation learning coupled to the search loop rather than treating the encoder as a frozen preprocessing step.","The gains are not confined to one encoder or acquisition function: TTARO improves regret AUC in 37 of 40 settings, spanning learned GNN embeddings, graph-autoencoder embeddings, high-dimensional WL features, linear and RBF kernels, and EI, UCB, and Thompson sampling.","Freezing a deep kernel after the initial sample can be actively harmful on larger, more heterogeneous search spaces: DKL degrades below plain GP on Ckt-Bench-301, whereas TTARO avoids the degradation.","TTARO's benefit appears early enough to matter under a fixed budget: at 60% of the budget it already exceeds GP's final best FoM on Ckt-Bench-101 and matches it on Ckt-Bench-301."],"supporting_citations":[{"why":"Supplies the two Open Circuit Benchmark search spaces (Ckt-Bench-101 and Ckt-Bench-301), the CktGNN encoder, and the precomputed FoM labels that all experiments use.","marker":"[3]"},{"why":"Defines deep kernel learning, the frozen-after-initial-sample baseline whose failure modes on Ckt-Bench-301 isolate the value of TTARO's online update.","marker":"[7]"},{"why":"Prior automatic op-amp topology synthesis in a learned latent space; establishes the problem setting of pretrained-embedding BO that TTARO targets.","marker":"[2]"},{"why":"Earlier graph-embedding BO over op-amp topologies, representing the fixed-representation approach TTARO is compared against.","marker":"[4]"},{"why":"Prior adaptive deep-kernel BO in protein and nanophotonics settings; TTARO adapts this idea to finite-bank analog topology search.","marker":"[19]"},{"why":"Prior contrastive embedding method for structured-space BO, part of the adaptive-representation literature TTARO positions itself within.","marker":"[21]"},{"why":"Prior work encouraging correlation between latent-space distance and objective-value difference, which motivates the objective-aligned geometry TTARO learns.","marker":"[23]"}],"fun_headline_variants":["Retrain circuit embeddings mid-search for up to 47% less regret","Online representation adaptation cuts BO regret by 46.7%","TTARO adapts embeddings during search to cut regret","Mid-search embedding retraining improves analog circuit BO","Don't freeze embeddings, retrain during BO for better circuits"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison assumes the initial sample size $n_0$ and evaluation budget $B$ are chosen so that the frozen-after-initial-sample deep-kernel baseline is fairly matched; the paper never reports $n_0$ or $B$, so if $n_0$ is small, TTARO's reported advantage over that baseline may reflect more training labels rather than online adaptation.","fun_headline_variants_meta":{"raw":{"variants":["Retrain circuit embeddings mid-search for up to 47% less regret","Online representation adaptation cuts BO regret by 46.7%","TTARO adapts embeddings during search to cut regret","Mid-search embedding retraining improves analog circuit BO","Don't freeze embeddings, retrain during BO for better circuits"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000799,"raw_usage":{"total_tokens":3590,"prompt_tokens":1094,"completion_tokens":2496,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":710,"completion_tokens_details":{"reasoning_tokens":2412}},"tokens_in":710,"tokens_out":2496,"duration_ms":19403,"temperature":1.0,"reasoning_tokens":2412,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:34:24.244608+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Vary the initial sample size $n_0$ from very small to nearly the full budget while holding the evaluation budget fixed, and plot regret AUC for TTARO against the deep-kernel baseline that freezes its representation after the initial sample; if TTARO's advantage shrinks toward zero as $n_0$ grows, the reported gain is mostly extra training data, not online adaptation. Alternatively, give the frozen baseline access to the same total number of training labels that TTARO has at the end of the run and check whether the 20.7% average reduction survives.","supporting_citations":[{"cited_title":"CktGNN: Circuit graph neural network for electronic design automation,","cited_arxiv_id":null,"evidence_quote":"Supplies the two Open Circuit Benchmark search spaces (Ckt-Bench-101 and Ckt-Bench-301), the CktGNN encoder, and the precomputed FoM labels that all experiments use."},{"cited_title":"Deep kernel learning,","cited_arxiv_id":null,"evidence_quote":"Defines deep kernel learning, the frozen-after-initial-sample baseline whose failure modes on Ckt-Bench-301 isolate the value of TTARO's online update."},{"cited_title":"ATOM: An automatic topology synthesis framework for operational amplifiers,","cited_arxiv_id":null,"evidence_quote":"Prior automatic op-amp topology synthesis in a learned latent space; establishes the problem setting of pretrained-embedding BO that TTARO targets."},{"cited_title":"Topology optimization of operational amplifier in continuous space via graph embedding,","cited_arxiv_id":null,"evidence_quote":"Earlier graph-embedding BO over op-amp topologies, representing the fixed-representation approach TTARO is compared against."},{"cited_title":"Contrastive embedding of structured space for bayesian optimization,","cited_arxiv_id":null,"evidence_quote":"Prior contrastive embedding method for structured-space BO, part of the adaptive-representation literature TTARO positions itself within."},{"cited_title":"Advancing bayesian optimization via learning correlated latent space,","cited_arxiv_id":null,"evidence_quote":"Prior work encouraging correlation between latent-space distance and objective-value difference, which motivates the objective-aligned geometry TTARO learns."}],"review_version":1}