acceptodds
Under review as a conference paper at ICLR 2027

TRACE: LLM-Orchestrated Bayesian Optimization for Analog Circuit Sizing

Abstract

An analog circuit sizing task must reconcile competing specifications across testbenches and process, voltage, and temperature corners through costly circuit simulations. Bayesian optimization (BO) reuses these evaluations to guide sampling, but a stalled loss curve alone does not identify the circuit-level cause or the intervention it calls for. We introduce TRACE, an LLM-orchestrated framework that composes the sizing workflow online as simulation evidence accumulates. At each decision point, TRACE refreshes six diagnostic views of the committed results, reads them alongside device roles and circuit topology, and decides whether, when, and at what scope to act: continue BO, refine a local candidate, explore an unsampled region, or revise the sizing task. Each decision stays bound to the evidence available when it was made, so an episode can be replayed and its design checked against simulator records. On four circuits in a TSMC 28 nm process with Cadence Spectre, TRACE returns a fully verified design in every run and achieves the lowest mean simulation count among the compared methods. Relative to TRACE, tSS-BO requires – as many simulations and – as much wall time. Action ablations show that neither single intervention is uniformly best across the tested circuits.

Primary area:
applications to robotics, autonomy, planning
Keywords:
LLM Agents, Adaptive Workflow Orchestration, Analog Circuit Sizing, Bayesian Optimization, Evidence-Grounded Decision Making

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

This paper proposes TRACE, an LLM-orchestrated Bayesian Optimization framework for analog circuit transistor sizing. Unlike previous LLM-based BO approaches that generate additional candidates or refine search spaces with the LLM, TRACE dynamically determines when and how to intervene during optimization by selecting among 4 actions: Hold, Refine, Explore, and Revise, where the decision is made by an LLM using diagnostics from circuit simulation results. Experiments across four analog circuits (while the circuits are relatively small) demonstrate reduced simulation counts and wall-clock time, showing its sample-efficiency compared with conventional BO and LLM-BO baselines.

Critical strengths
  • Finer control (dynamic intervention) clearly gives improved sample efficiency
  • The paper provides fairly comprehensive comparisons, and the experimental results look convincing.
Critical weaknesses
  • From an optimization perspective, I don’t see much methodological novelty, and the paper's main contribution is orchestrating existing operations using LLM. While TRACE's improvements over the baselines are promising, it is hard to say that the gap will still maintain with better LLM in the future.
  • The benchmark circuits are relatively small.
Questions for authors
  1. Can you add results for larger circuits such as full differential amplifier with CMFB and biasing circuits, or a three-stage OTA?
  2. The authors should clarify the actual circuit topologies. The reported number of devices and independent design parameters alone does not necessarily reflect the true circuit complexity. For example, it’s unclear how a circuit with 9 devices ends up with 17 independent design parameters. Have they accounted for parameter sharing among symmetric or matched devices, which would typically have the same sizing parameters?
5 comments
AuthorsOpenReview ↗

Novelty and stronger LLMs

We thank the reviewer for the assessment and for recognizing that dynamic intervention improves sample efficiency and that the comparisons are comprehensive. Below we address the scope of the contribution, the role of future LLMs, circuit complexity, and parameter counting, including a new three-stage amplifier experiment.

1. Methodological novelty

We agree that, from the perspective of the optimizer itself, TRACE does not propose a new surrogate or acquisition function; its BO backend is the existing tSS-BO. This is also the setting of recent LLM-assisted sizing methods: ADO-LLM pairs a standard GP/qEI loop with LLM-authored candidates, LEDRO runs TuRBO within LLM-proposed ranges, and AutoSizer selects among existing optimizers while revising the search space. Their contributions, like ours, lie in how LLM reasoning is coupled with numerical search rather than in a new surrogate or acquisition function. Within this line of work, our contribution is deciding when, and at what scope, to intervene in a running optimization. As the reviewer kindly noted in the summary, existing LLM-based BO methods use the LLM in a fixed role and on a fixed schedule. ADO-LLM inserts one LLM candidate after every four GP candidates, and LEDRO alternates LLM range proposals with fixed TuRBO budgets (Appendix D.3). In TRACE, whether, when, and at which scope to intervene (a single candidate, an unsampled region, or the task configuration) are decided online.

This decision is not trivial. The most effective single intervention differs across circuits: Explore only is cheaper than Refine only on LDO, whereas the opposite holds on BGR (Table 2). A fixed schedule is therefore unlikely to suit every circuit, and TRACE improves on the better single intervention on both circuits (by 9.8% and 15.3% in simulations). Within the same evaluator, orchestration reduces simulations by 1.46–3.64× relative to Initial design + BO. Replacing the LLM selector with a warm-started, cost-aware contextual-bandit selector over the same views and actions increased the number of evaluated candidates from 36 to 53 on LDO and from 49 to 64 on BGR (see our response to Reviewer GwXy). We will revise the contribution statement in Section 1 to state this scope explicitly.

2. Will the gap persist with stronger LLMs?

The main comparison uses the same LLM and reasoning effort for all LLM-based methods, so the observed differences are not due to a stronger model for TRACE. We cannot predict future gaps and limit our conclusions to the evaluated models. Our measurements do, however, indicate why the problem TRACE addresses should remain relevant. First, TRACE already adapts how much it relies on the LLM: with the same model, it makes about four decisions per run on OTA and COMP and 12–14 on LDO and BGR, using Explore mainly on the two harder circuits (see our response to Reviewer bKer). A fixed schedule, such as ADO-LLM's one LLM candidate in every round of five, cannot make this adjustment. Second, LLM interventions are already worth more than their share of the cost: LLM-authored candidates consume 15–22% of simulations, yet they proposed the final verified design in 13 of the 19 runs that entered BO. A stronger model would raise the value of each intervention, but LLM calls and simulations would still have a cost, so deciding when and at what scope to spend them remains the orchestration problem. Because TRACE makes this decision online, a more capable model can be used more often or at a broader scope without redesigning the workflow. The extrapolation to future models remains a hypothesis rather than a result, and we will add it to the discussion of limitations.

AuthorsOpenReview ↗

A larger three-stage amplifier

3. Larger and more complex circuits

Two of the current tasks are already of substantial complexity. The BGR has 43 devices, including a PNP core (fixed 1:8:1 emitter ratio), several current-mirror groups, an amplifier, compensation, and a startup circuit; it has 67 sizing parameters before matching and 31 independent parameters after matching, with a search space of about 103810^{38}. The LDO contains an error amplifier, a buffer stage, a PMOS pass device (multiplier 300), and RC compensation, and must meet 10 specifications, with phase margin and PSRR constrained at three load currents, over a search space of about 104810^{48}. All four circuits and their testbenches were built by experienced circuit designers in our group and verified in the TSMC 28 nm PDK with Spectre across three PVT corners.

We agree that a larger topology provides further evidence. Following the reviewer's suggestion of a three-stage OTA, we added a three-stage OTA with active-feedback frequency compensation (AFFC) (Lee & Mok, 2003). To ensure that the topology itself is correct, we took it from the public AnalogGym suite (Li et al., 2024), where it is listed as HoiLee_AFFC_Pin_3 (netlist), and migrated it to the same TSMC 28 nm PDK with a 2.5 V supply, evaluated with the same Virtuoso/Spectre flow as the existing tasks. The migrated schematic and testbenches were reviewed by experienced analog design engineers in our group. The amplifier has 30 MOS devices and two compensation capacitors: a folded-cascode first stage with a PMOS input pair, a second gain stage, a third-stage output, a feedforward pull-up driven by the first-stage output, a Miller capacitor C0, and an active-feedback path through which C1 is fed back. To allow direct comparison with the AnalogGym netlist, the device names and sizing groups below are aligned with AnalogGym. The task has 32 independent variables, the width W, length L, and multiplier M of each of the ten groups in the table plus C0 and C1; the other 20 devices follow their group, giving 60 linked variables. The groups and the multiplier ratios within each group are the same as in AnalogGym. The only differences are that the bias current is kept fixed rather than optimized (32 instead of 33 variables) and that the ranges are set for the 28 nm PDK: L ∈ [270, 1200] nm (step 10 nm), W ∈ [0.3, 6] µm (step 20 nm), M ∈ {1, …, 20}, and C0, C1 ∈ [20, 1200] fF (step 20 fF), about 106110^{61} grid points in total. The task has the 10 specifications below, evaluated at the same three PVT corners as the other tasks (TT25, SS125, and FF−40) with a 10 fF load; four testbenches run seven simulation jobs per corner (21 simulations per full evaluation, counted as in Table 1). The noise target is the same as for our two-stage OTA.

SpecificationTargetSpecificationTarget
DC gain≥ 110 dBSlew rate (rise and fall)≥ 80 V/µs
GBW≥ 30 MHzSettling time≤ 150 ns
Phase margin≥ 45°CMRR≥ 70 dB
Input-referred noise at 100 kHz≤ 120 nV/√HzPSRR≥ 70 dB
Static power≤ 220 µWOutput swing≥ 1.8 V
AuthorsOpenReview ↗

A larger three-stage amplifier (continued)

3. Larger and more complex circuits (continued)

The ten groups of the three-stage amplifier are listed below.

GroupDevicesRoleLinked sizing
G0M0–M7PMOS bias, mirror load (M5, M6), tail (M4), pull-up bias (M7)shared W, L, M; M4 at 2×
G18M13–M15, M18–M21, M61, M64NMOS bias and current sourcesshared W, L; M15 at 1×, M20 and M21 at 8×, others at 4×
G9M9, M10PMOS input pairidentical
G17M16, M17folded cascodeidentical
G22M22, M23second-stage NMOS mirroridentical
G59M59, M62active-feedback PMOS mirroridentical
G60M60, M63active-feedback common-gate pair (C1 at the source of M63)identical
G11M11second-stage transconductor—
G12M12feedforward pull-up from the first-stage output—
G25M25third-stage NMOS output—
MethodVerifiedSˉ\bar{S}Tˉ\bar{T} (min)
tSS-BO3/37883193.2
ADO-LLM3/32100110.3
LEDRO3/3168098.4
TRACE3/3108570.5

Means over three runs. As in Table 2, TRACE and tSS-BO use phased evaluation, while ADO-LLM and LEDRO evaluate every candidate fully.

All methods reached verified designs in all three runs. TRACE required 1085 simulations and 70.5 minutes on average, 7.3 times fewer simulations than tSS-BO and 1.9 and 1.5 times fewer than ADO-LLM and LEDRO, with correspondingly shorter wall time. It also evaluated fewer candidates than both LLM baselines (53 vs. 100 and 80), so its advantage does not come from phased evaluation alone. The gains observed on the original four circuits therefore carry over to this larger three-stage amplifier.

Reviewer_GwXyOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents TRACE, an LLM-orchestrated Bayesian optimization framework for analog circuit sizing. It introduces six diagnostic views to summarize optimization progress and allows the LLM to dynamically select among four actions: Hold, Refine, Explore, and Revise. Experiments on four analog circuits using Cadence Spectre demonstrate improvements in simulation efficiency and wall-clock time over existing methods.

Critical strengths

The idea of using an LLM to dynamically decide when and how to intervene in Bayesian optimization is interesting and reasonably novel. The six diagnostic views and four actions are well-designed and interpretable, providing a structured interface between the LLM and the optimizer.

The experimental results are promising, showing substantial improvements over traditional BO and competitive LLM-based methods. The use of commercial circuit simulators and multiple PVT corners makes the evaluation convincing. The overall framework is also clearly presented, with thoughtful designs for asynchronous execution and decision tracking.

Critical weaknesses

My main concern is whether the LLM is really necessary for action selection. The diagnostic views and available actions are highly interpretable, and many decisions seem possible to implement using simple heuristic rules or lightweight models. Although the paper provides action ablations, it remains unclear how much benefit comes specifically from LLM-based decision-making rather than the available actions themselves.

I also have some concerns about task revision and asynchronous execution. When parameter ranges or matching constraints change, BO is restarted with the previous simulation history. It is unclear how observations from different configurations are handled. Moreover, since BO continues running during LLM inference, some decisions might already be outdated when executed. For complex circuits, optimization stagnation is common, and unnecessary interventions could potentially disrupt the search.

Finally, the baseline comparisons involve different initialization and evaluation strategies. In particular, ADO-LLM and LEDRO do not use the same phased evaluation strategy as TRACE. While comparing complete systems is reasonable, these differences make it difficult to isolate the improvements brought by the proposed orchestration mechanism.

Questions for authors
  1. Have the authors considered a rule-based or lightweight controller using the same diagnostic views and actions? What is the specific advantage of LLM-based action selection?

  2. How does BO reuse historical observations after task revisions? Have the authors observed cases where an intervention becomes outdated or counterproductive due to asynchronous execution?

  3. Could the authors further clarify how differences in initialization and evaluation protocols affect the baseline comparisons?

11 comments
AuthorsOpenReview ↗

Is the LLM necessary for action selection?

We thank the reviewer for the detailed review and for finding the idea interesting and reasonably novel, the views and actions interpretable, and the evaluation with commercial simulators and PVT corners convincing. We understand the main concerns to be (i) whether the LLM is needed for action selection, and (ii) whether the gains can be attributed to orchestration given the different evaluation protocols. We address both with new evidence below, together with the questions on history reuse and asynchronous execution.

Is the LLM necessary for action selection?

We agree that the single-action ablations do not separate the value of the available actions from the value of having the LLM choose among them. We first note that the LLM in TRACE does more than choose an action label (Section 3.5): it also specifies circuit-dependent arguments, such as the parent and size changes for Refine, the region and candidates for Explore, and the configuration change for Revise.

To isolate action selection, we followed the reviewer's suggestion and replaced the LLM selector with a lightweight learned controller using the same views and actions: LinTS, Thompson sampling for contextual bandits with linear payoffs (Agrawal & Goyal, 2013). At each decision, it takes 10 features from the frozen view snapshot (e.g., recent improvement and slope, repeated-sampling ratio, stall length, boundary pressure, local-direction support, and outcomes of recent actions) and selects Hold, Refine, or Explore by posterior sampling; Revise is triggered by the setup-health rule of Appendix B.5. To give LinTS a fair starting point, we warm-started it with 97 recorded TRACE decisions on OTA and COMP; no LDO or BGR data were used. Because the actions differ in cost, the reward is cost-aware: a successful action, one whose own candidates (or, for Hold, the BO results during the wait) improved on the best design, earns 1 for Hold, 0.5 for Refine, and 0.2 for Explore, reflecting relative costs of 1, 2, and 5. The arguments of Refine and Explore are still generated by the same LLM, and all other settings are identical.

CircuitSelectorRunsVerifiedCandidatesTˉ\bar{T} (min)Hold / Refine / Explore per run
LDOTRACE (LLM selector)33/33655.43.0 / 4.0 / 1.3
LDOLinTS33/35394.73.7 / 4.0 / 2.7
BGRTRACE (LLM selector)33/34963.23.0 / 4.0 / 2.7
BGRLinTS33/36479.34.0 / 3.0 / 3.7

Means over three runs. All runs use full evaluation of every candidate. The LDO TRACE row is the all-views condition of the view ablation; the BGR TRACE runs were run under the same protocol.

On LDO, even with prior data and a cost-aware reward, LinTS required 53 candidates and 94.7 minutes, 1.5 and 1.7 times those of TRACE. On BGR, LinTS required 64 candidates and 79.3 minutes, compared with 49 and 63.2 for TRACE. LinTS made more decisions (10.3 vs. 8.3 per run on LDO and 10.7 vs. 9.7 on BGR) and, in particular, chose Explore more often (2.7 vs. 1.3 and 3.7 vs. 2.7 per run), the action that pauses BO and consumes a batch of candidates. With the same views and actions, the learned selector thus spends more of the budget on costly interventions that the LLM selector judges unnecessary.

This also points to a practical difference that holds regardless of these numbers. A learned selector needs decision data from related circuits, which here were themselves generated by TRACE's LLM, and a hand-specified cost for each action. The LLM selector needs neither: it weighs the expected gain and cost of each action from the topology and diagnostic views, and can be applied to a new circuit directly. We will include the full setting in the appendix of the revised paper.

AuthorsOpenReview ↗

Evaluation protocols and history reuse

2. Attribution under different initialization and evaluation protocols

We agree that the comparison with ADO-LLM and LEDRO measures complete sizing systems rather than orchestration alone. Following their original designs, both evaluate every candidate on all testbenches and corners and use their own initialization (Appendix D.3), so we do not attribute TRACE's lower cost relative to them to orchestration alone (Appendix D.5).

The orchestration effect is instead isolated within a shared protocol. The BO baselines and all TRACE ablations use the same phased evaluator, and TRACE uses tSS-BO as its backend. Moving from tSS-BO to Initial design + BO changes only the initialization; moving from Initial design + BO to TRACE keeps the initial design, BO, and evaluator fixed and adds only online orchestration:

Sˉ\bar{S} (Tˉ\bar{T}, min)OTACOMPLDOBGR
tSS-BO832 (70.9)912 (71.7)1446 (133.5)1364 (106.2)
Initial design + BO674 (60.2)669 (53.1)1011 (101.8)894 (82.2)
TRACE386 (32.7)184 (18.1)692 (60.4)493 (65.7)

The LLM-authored initial design reduces simulations by 19–34%, and orchestration then reduces them by a further 32–72% (1.46–3.64×). This second step is the controlled measurement of orchestration that the reviewer asks for. We will move this distinction from Appendix D.5 into Sections 4.1 and 4.2.

To further examine the role of phased evaluation, we also ran TRACE on LDO and BGR with full evaluation of every candidate, matching the evaluation protocol of ADO-LLM and LEDRO (three runs per circuit with GPT-6 astra; these are the TRACE runs of the selector comparison above). All runs reached verified designs.

CircuitMethodCandidatesSimulationsTˉ\bar{T} (min)
LDOTRACE, full evaluation3675655.4
LDOADO-LLM (Table 2)4186172.5
LDOLEDRO (Table 2)55115599.4
BGRTRACE, full evaluation4958863.2
BGRADO-LLM (Table 2)4655481.8
BGRLEDRO (Table 2)6072074.9

Each candidate costs 21 simulations on LDO and 12 on BGR. Baseline values are from Table 2, measured in a separate batch; candidate counts are rounded means.

With every candidate fully evaluated, TRACE still used fewer simulations than both LLM baselines on LDO; on BGR it used fewer than LEDRO and about as many as ADO-LLM (588 vs. 554). Its wall time remained lower on both circuits, partly because TRACE overlaps LLM reasoning with ongoing BO simulations, whereas ADO-LLM waits for each batch of five simulations to return before reasoning, leaving the simulators idle during inference.

3. History reuse after task revision

After a revision, BO reuses earlier simulation results, not earlier loss values or optimizer state. Earlier observations are reused when the parameter set (identities, types, and order), testbenches, corners, and netlist are unchanged; changed ranges do not prevent reuse. Each earlier design is re-scored under the revised task from its stored results, and designs lacking a newly required result are kept in the history but not used as warm-start observations. Designs are encoded relative to the revised bounds without clipping, so a design outside the new range keeps a coordinate outside [0, 1] and still informs the GP, while new candidates stay within the revised range. If a revision changes the parameter set, e.g., a matching revision that adds or removes parameters, BO does not reuse the earlier observations; the LLM repeats the initial setup of Section 3.3 and authors a new warm-start batch, so the agent that defines the search space also initializes it. In both cases, the LLM keeps access to the complete history through the views and raw records. In all our experiments, revisions were range extensions (Appendix D.6), so BO reused all completed earlier observations. We have added this rule to Section 3.5 and Appendix A and revised Algorithm 1 accordingly.

AuthorsOpenReview ↗

Asynchronous interventions

4. Outdated or unnecessary interventions under asynchronous execution

Acting on slightly outdated information is inherent to any parallel optimizer. BO itself runs four concurrent trials (Appendix C), so each BO candidate is proposed without the results of other pending evaluations; techniques such as the constant liar (Ginsbourger et al., 2010) and fantasized observations (Snoek et al., 2012) exist because this delay is unavoidable. In TRACE, the LLM's delay is no larger. While it reasons, about 2 of the 4 concurrent simulations complete on LDO and BGR, and about 2.8 on OTA and COMP, no more than the up to three evaluations pending whenever BO proposes a candidate. LLM-authored candidates are also dispatched to the next free simulator slot ahead of queued BO proposals, so their delay is not extended by queueing; we have made this rule explicit in Appendix A.

We have observed interventions that did not help: in the main comparison, 29 of 90 Refine actions and 8 of 23 Explore actions did not improve on the best design available when their results returned (see our response to Reviewer bKer). We do not attribute these specifically to asynchrony, since a local step or a probed region can also fail with fully current information. Such an intervention is treated like an unsuccessful BO candidate: it cannot replace the best design or bypass final verification, and its cost is bounded (one candidate for Refine, at most two rounds of five for Explore). The other 76 of 113 interventions (67%) improved on the best design available when their results returned, which already included every BO result that arrived during reasoning and simulation, so the delay rarely made an intervention obsolete. Regarding unnecessary interventions during stagnation, the agent chose Hold in 30–47% of decisions, and with the same evaluator and initial design, TRACE required fewer simulations than Initial design + BO on all four circuits, so the net effect of its interventions was positive in our experiments.

We hope that the new selector comparison and the clarified scope of each comparison address the reviewer's main concerns, and we would be grateful if the reviewer would consider them in the final assessment. We are happy to answer any further questions.

References

  • S. Agrawal and N. Goyal. Thompson sampling for contextual bandits with linear payoffs. ICML, 2013.
  • D. Ginsbourger, R. Le Riche, and L. Carraro. Kriging is well-suited to parallelize optimization. In Computational Intelligence in Expensive Optimization Problems, Springer, 2010.
  • J. Snoek, H. Larochelle, and R. P. Adams. Practical Bayesian optimization of machine learning algorithms. NeurIPS, 2012.
Reviewer_bKerOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes TRACE, a large language model (LLM)-orchestrated Bayesian optimization (BO) framework for analog circuit sizing, aiming to address the limitations of traditional BO in diagnosing search stagnation and applying targeted interventions. The framework integrates historical simulation results, optimization progress, and search-space information through six structured diagnostic views. By combining these views with circuit topology and device roles, the LLM adaptively selects among four operations: continuing BO search (Hold), local candidate refinement (Refine), regional exploration (Explore), and task configuration revision (Revise), enabling dynamic decisions on when and how to intervene. In addition, TRACE employs an asynchronous coordination mechanism between the LLM and BO, while recording the evidence and outcomes associated with each decision to ensure the traceability of the optimization process. The main contribution of this work lies in using the LLM as a dynamic orchestrator of the optimization workflow, enabling adaptive combinations of search strategies based on simulation feedback.

Critical strengths
  1. The research problem is practically significant, and the proposed method demonstrates novelty. TRACE addresses the limitations of traditional BO in diagnosing search stagnation and selecting appropriate interventions. Compared with methods that focus on individual optimization stages, TRACE adaptively coordinates interventions of different scopes based on the evolving search state, enabling more flexible optimization.
  2. The structured diagnostic views are well designed and improve LLM decision-making efficiency. Six diagnostic views summarize simulation history, optimization progress, search-space coverage, and intervention feedback, providing structured evidence for decision-making while reducing reasoning overhead.
  3. The framework is comprehensive and offers good practicality and traceability. TRACE integrates task initialization, BO search, LLM-driven interventions, and final verification into a closed-loop workflow. Its asynchronous execution and decision records support efficient coordination and traceable optimization.
  4. The experimental evaluation is systematic and supports the main conclusions. Experiments on four analog circuits, comparisons with existing methods, and multiple ablation studies demonstrate TRACE's effectiveness in reducing simulation counts and optimization time.
Critical weaknesses
  1. The individual contributions of the six diagnostic views are insufficiently analyzed. Section 4.3 only compares full diagnostic views with raw simulation records, without examining each view's contribution or potential redundancy. More fine-grained ablation studies would strengthen the analysis.
  2. The quantitative analysis of LLM-driven interventions could be further strengthened. Although Section 4.2 and Appendix D.4 provide ablation results and illustrative cases, systematic aggregate statistics on intervention frequency, execution costs, and action distributions across runs are limited. Additional analysis would improve the interpretability of the framework.
  3. Sensitivity analysis of key hyperparameters is lacking. Sections 3.4–3.5 and Appendix C.2 describe several fixed settings, such as the diagnostic window size, Hold interval, and Explore batch size, but provide limited justification for these choices and no systematic sensitivity analysis. Further experiments would help assess the robustness of these settings.
Questions for authors
  1. What are the individual contributions of the six diagnostic views to optimization efficiency? Is there any redundancy among them? Could finer-grained ablation studies further validate their necessity?
  2. Could the authors provide statistics on intervention frequency, execution costs, and strategy distributions to better explain how the LLM dynamically selects interventions and the role of each operation?
3 comments
AuthorsOpenReview ↗

Contribution and possible redundancy of individual views

We thank the reviewer for the positive and constructive assessment, and for recognizing the practical significance of the problem, the design of the diagnostic views, and the systematic evaluation. The three suggestions concern finer-grained evidence on the views, the interventions, and the numerical settings; we have added an analysis for each.

1. Contribution and possible redundancy of individual views

We agree that a finer-grained analysis clarifies the role of each view. Section 3.4 presents the six views as summaries of existing simulation history and does not claim that the set is minimal, and Section 4.3 compares the complete interface with raw records. It therefore supports the interface as a whole but does not show that each view is individually necessary.

Following the roles described in Section 3.4, we group the views into state views (specification trade-offs and BO progress), which locate the current design; scope views (local opportunity, search-space coverage, and setup health), which correspond to the scopes at which an intervention can act; and the retrospective view (feedback from earlier actions). On LDO, we removed these groups cumulatively, keeping the initial design, seeds, observation intervals, and stopping rule of Section 4.3. To keep episode time from depending on where phased evaluation stops, every candidate in this experiment was fully evaluated on all testbenches and corners. Removed summaries were unavailable during later retrievals, and raw records remained accessible throughout.

Views providedCandidatesT̄ (min)Inference (s/decision)Record queries (/decision)
All six views3655.438.66.8
State + scope views4760.544.16.8
State views only5669.362.27.8
None (raw records only)6375.267.78.9

Means over three runs per condition with GPT-6 astra. All four conditions were run together under the same code and service conditions, with full evaluation of every candidate (21 simulations per LDO candidate). Per-decision metrics are averaged over all decisions after initialization.

Removing views increased the cost at every step. Relative to all six views, removing the retrospective view increased inference time per decision by 14% without changing record queries. Further removing the scope views produced the largest increase, to +61% in inference time and +15% in record queries, and removing the state views as well raised them to +76% and +31%. The number of evaluated candidates and the episode time followed the same order, from 36 to 63 candidates and from 55.4 to 75.2 minutes, suggesting that the views make not only each decision cheaper but also the search more sample-efficient. Each group therefore contributes in this task, the scope views most, and we found no group whose removal came without cost. The retrospective view mainly shortens reasoning rather than retrieval, suggesting that its summary of earlier outcomes spares the agent from reassessing them. The all-views and raw-records conditions reproduce the reductions reported in Section 4.3 (−43% in inference time and −24% in record queries, compared with −43% and −25% in Figure 4); their absolute episode times are not directly comparable with Figure 4, which used phased evaluation.

This ablation uses the unrestricted LDO task, in which no configuration defect arises, so it likely understates the role of setup health and boundary coverage, which mainly inform Revise.

AuthorsOpenReview ↗

Intervention frequency, costs, and action distributions

We agree that Figure 3 and Appendix D.4 illustrate individual trajectories rather than behavior across runs. We have therefore computed aggregate statistics over the TRACE runs of the main comparison, counting the decisions made after the initial design.

CircuitDecisions/runHoldRefineExploreReviseSims from LLM candidatesRefine → new bestExplore → new bestFinal design from BO / Refine / Explore
OTA†4.252.002.000.25022.3%6/81/12/2/0
COMP4.41.62.80015.2%13/14–0/5/0
LDO12.64.66.21.8017.6%22/316/92/3/0
BGR14.24.27.42.6022.4%20/378/132/2/1

Action columns are mean counts per run. "New best" columns give totals over runs: the number of actions whose own candidates improved on the best design available when their results returned, out of all such actions; later BO improvements are not attributed to them (Section 3.6). "Sims from LLM candidates" is the share of testbench–corner simulations spent on candidates submitted by Refine and Explore. †One OTA run met all specifications during the initial design and never entered BO; OTA statistics are over the remaining four runs.

The statistics show three patterns.

Intervention effort follows circuit difficulty. On OTA and COMP, TRACE makes about four decisions per run and relies almost entirely on Refine. On LDO and BGR, it makes 12–14 decisions and adds coordinated exploration (1.8–2.6 Explore actions per run). This is consistent with Section 5: a few local refinements suffice on the easier circuits, while the gap opens on LDO and BGR. On COMP, all five final designs came from Refine, in line with Table 2, where Refine only nearly matches full TRACE.

Interventions are frequently productive. Between 54% and 93% of Refine candidates, and 62–67% of Explore batches on LDO and BGR, directly improved on the best design available when their results returned.

Interventions cost little relative to what they deliver. LLM-authored candidates consume only 15–22% of simulations, yet Refine or Explore proposed the final verified design in 13 of the 19 runs that entered BO. The agent also chose Hold in 30–47% of decisions, leaving BO to continue unchanged.

Revise was never selected in the main comparison, where no configuration defect arose; it is evaluated separately in the restricted-configuration tasks of Section 4.4 (Table 3).

AuthorsOpenReview ↗

Sensitivity of key settings

These settings reflect the trade-offs described in Section 3.5 rather than values tuned for the reported results. The Hold interval balances decision frequency against new evidence, and the Explore batch size balances the need for more than one sample per region against the cost of each round. The 12-candidate window of the BO progress view only sets how many recent candidates the stall summary covers; it does not restrict the actions or alter the optimizer. Since the other two settings directly change the search trajectory, we analyzed them on LDO, changing one setting at a time under the same full-evaluation protocol as the view ablation above.

SettingRunsCandidates per runMean candidatesMean T̄ (min)
Default (k = 3, batch 5)343, 30, 353655.4
Hold k = 1421, 78, 45, 575067.9
Hold k = 6261, 686582.3
Explore batch 2248, 545171.6
Explore batch 8260, 565875.0

Each candidate is fully evaluated with 21 simulations. The default row is the all-views condition of the view ablation above. We ran Hold k = 1 four times because its first two runs differed widely.

The default setting required the fewest candidates and the least time among the tested values, and moving either setting in either direction increased the cost, in line with the trade-offs in Section 3.5. Every default run (30–43 candidates) needed fewer candidates than any run with k = 6 or an Explore batch of 2 or 8 (48–68). With k = 6, interventions came later, and both runs needed more candidates (61 and 68). With k = 1, the agent decided after nearly every new BO result, and the outcome became highly variable, from 21 to 78 candidates, consistent with repeated decisions on nearly unchanged evidence making the trajectory less predictable. For Explore, a batch of 2 needed 48 and 54 candidates, consistent with too few samples to judge a region, whereas a batch of 8 needed 60 and 56, as each round spent more candidates. The defaults therefore lie at a favorable point of both trade-offs on LDO. These results cover one circuit with two to four runs per setting, so they indicate local behavior around the defaults rather than general optimality.

We will include all three analyses in the appendix of the revised paper and clarify in Section 4.3 that the original view ablation evaluates the interface as a whole. We thank the reviewer again for these suggestions, which have made the analysis of TRACE more complete.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

We sincerely thank all three reviewers for their careful reading and constructive feedback. We are encouraged that the reviewers found the problem practically significant and the idea of letting an LLM decide when and how to intervene in BO interesting (bKer, GwXy, XPfR), the diagnostic views and actions well designed and interpretable (bKer, GwXy), and the experiments systematic and convincing (bKer, GwXy, XPfR).

In response to the reviews, we have added the following analyses and revisions. Details are given in the individual responses.

New experiments and analyses

  1. LLM selector vs. a learned lightweight selector (GwXy, XPfR). We replaced the LLM's action selection with LinTS, a warm-started, cost-aware contextual-bandit controller using the same diagnostic views and actions, on LDO and BGR.
  2. Aggregate intervention statistics (bKer, GwXy, XPfR). Action distributions, intervention frequency, cost, and direct outcomes over all 20 TRACE runs of the main comparison.
  3. Grouped ablation of the diagnostic views (bKer). Retrospective, scope, and state views removed cumulatively on LDO.
  4. Sensitivity to the Hold interval and Explore batch size (bKer). One-factor-at-a-time analysis on LDO.
  5. An additional three-stage amplifier (XPfR). A three-stage AFFC amplifier taken from the public AnalogGym suite and migrated to the TSMC 28 nm PDK, evaluated with the same flow as the existing tasks.

Clarifications and revisions

  1. The rule for reusing observations after task revision, and the corresponding change to Algorithm 1 (GwXy).
  2. The scheduling priority of LLM-authored candidates in asynchronous execution (GwXy).
  3. The scope of each comparison: external baselines measure complete sizing systems, while the shared-evaluator ablations isolate orchestration; to be moved from Appendix D.5 into Sections 4.1–4.2. We also report TRACE on LDO and BGR with full evaluation of every candidate, the protocol used by ADO-LLM and LEDRO (GwXy).
  4. A summary of circuit topologies and parameter counts before and after matching (XPfR), and a sharper statement of the contribution's scope in Section 1 (XPfR).

We have uploaded a revised PDF with the changes to Section 3.5, Algorithm 1, and Appendix A (items 6 and 7); the other revisions listed above will be included in the next revision. We hope these additions address the reviewers' concerns, and we would be glad to provide further details during the discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.