acceptodds
Under review as a conference paper at ICLR 2027

SizingAgent: Incentivizing Physics-Grounded Reasoning in LLM-Based Analog Circuit Sizing

Abstract

Large Language Models (LLMs) are increasingly adapted for analog circuit sizing, yet existing methods often treat them as direct parameter generators and defer all validation to SPICE. We observe that this leads to many risky parameter updates that sound plausible in language but contradict the local physical response of the circuit. In this paper, we introduce SizingAgent, a training-free LLM framework that empowers LLM agents with physics-grounded reasoning for sizing optimization in analog design automation. Instead of relying on the LLM to directly generate numerical design parameters, SizingAgent structures optimization as a verifiable sizing loop consisting of three modules: (1) Physics Initializer (PI), which extracts small-signal equations (S-Equations) from the netlist to provide circuit-specific physical priors; (2) Physics-Guided Reasoner (PGR), which analyzes specification gaps and derives targeted sizing updates through a physics-guided reasoning chain; and (3) Sizing Verifier (SV), which checks the physical consistency of the proposed updates using a sensitivity matrix. We further release, to the best of our knowledge, the largest open-source benchmark for LLM-based analog circuit sizing. Under a shared Gemini 2.5 Flash backbone, SizingAgent achieves 74.8% Pass@1, surpassing the strongest LLM baseline by 32.8 percentage points and dramatically outperforming classical optimizers, whose Pass@1 reaches only 16.6% for Genetic Algorithms and 26.2% for Bayesian Optimization. Replacing the backbone with GPT-5.5 further improves performance to 93.1% Pass@1. The framework also generalizes across LLM families, with commodity backbones achieving 69.6–76.1% Pass@1 and Qwen3-32B improving from 10.1% to 59.1% over zero-shot prompting. Code and benchmark are available at https://github.com/artifact-repro/sizingagent.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
large language models, analog circuit sizing, physics-guided reasoning

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_w8vmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SizingAgent is a training-free framework for analog sizing. It combines LLM-derived initialization with device lookup tables, a structured reasoning prompt supplied with finite-difference sensitivities, and a pre-simulation filter that checks proposed action directions. Multiple initializations and sensitivity-guided numerical exploration complete the workflow. The paper evaluates 27 SKY130 circuits and several proprietary and open-weight backbones. It reports 74.8% Pass@1 with Gemini 2.5 Flash, compared with 42.0% for the strongest evaluated same-backbone baseline.

Critical strengths
  1. The approach addresses a concrete failure mode with a deployable mechanism. Supplying measured sensitivities to an LLM and checking proposed directions before spending another simulation is sensible. The shared-backbone comparison, 27-circuit benchmark, and tests across multiple model families provide substantially more evidence than isolated examples. Table 7 is internally consistent: its counts sum to 303/405, yielding 74.8% Pass@1; all 27 circuits succeed at least once.

  2. The ablations already establish useful incremental effects. Within the separate ablation experiment, removing SV changes successes from 322 to 288, an 8.4-point reduction, and removing PGR yields 262, a 14.8-point reduction. Thus, there is evidence for these modules with the other components retained. The paper also explicitly acknowledges stale sensitivities, independently checked actions that are executed jointly, and equation/diagnosis fields that are not machine-verified. Those qualifications appropriately limit what “verification” means.

Critical weaknesses
  1. The baseline protocol and the largest ablation effect cannot currently be reconstructed unambiguously. Section 4.3, Table 2, and Appendix A specify 15 runs per cell. Yet several LLM-baseline entries, including 7.1/35.7 and 42.9/97.2 for Pass@1/Pass@5, match the estimator at 14 trials rather than 15. Separately, the 27 printed “w/o PI” counts in Table 6 sum to 171, not its reported 161. If the individual cells are correct, this implies 42.2% success and a 37.3-point decrease from 322/405, rather than the reported 39.8-point decrease. The other five ablation column totals are correct. These could be reporting errors, and the corrected PI effect would still be large; nevertheless, run denominators, excluded attempts, and the authoritative counts must be reconciled.

  2. The stronger attribution to LLM physics reasoning exceeds the controls. The no-PI ablation simultaneously removes LLM guidance and the gm/Id lookup mechanism, replacing the initialization with random sampling. It measures the value of that bundle, not an isolated LLM contribution as claimed in §4.5. This does not negate the separate SV/PGR gains. The novelty discussion should also distinguish the directional filter from earlier equation-based feedback systems, including White-Op (arXiv:2601.21321), Bujana and Karsilayan (arXiv:2604.07387v1), and NEMESIS (arXiv:2607.05657). Table 1's negative entry for AutoSizer's simulation-grounded loop is difficult to reconcile with that work's explicit feedback-to-LLM mechanism (§§3.6, 4.6, arXiv:2602.02849v2). A more precise positioning would strengthen the actual contribution without requiring a new model architecture.

Questions for authors
  1. What are the successful/attempted counts behind the affected Table 2 cells, and how are infrastructure errors and non-converging circuit simulations handled? Table 8 discloses one excluded GA run, but the corresponding LLM-baseline rules are unclear.

  2. Which “w/o PI” entries or aggregate are incorrect? Please reconcile them from existing records, and explain Appendix B's 20.45 simulations/run versus Table 7's 19.4. The main and ablation experiments are explicitly separate; their different aggregate rates are not, by themselves, contradictory.

  3. Is there existing evidence separating LLM initialization guidance from access to the same device tables and initialization procedure? For SV, can the recorded traces quantify useful proposals rejected or unhelpful proposals admitted as sensitivities become stale, beyond the illustrative checked steps?

5 comments
AuthorsOpenReview ↗

Response to Reviewer w8vm (Part 1/3)

We thank the Reviewer for the exceptionally careful audit of our records and the constructive comments. Below, we reconcile each flagged count with the archived records, add controls that isolate the contribution of the LLM, position SizingAgent precisely against the four works cited, and audit the recorded SV checks. All changes are highlighted in blue in the revised manuscript.

Q (W1). Why do some Table 2 cells appear to use 14 rather than 15 trials, and why do the 27 w/o PI entries in Table 6 sum to 171 rather than the reported 161?

Ans. We are grateful for this careful reconstruction. Both observations are correct, and we have reconciled them with the archived records. We have updated and re-checked every other cell of Tables 2 and 6 against the archived records and found no further discrepancy.

  1. Baseline denominators. The affected LLM-baseline cells were computed from 14 recorded outcomes, exactly as the Reviewer inferred from the estimator: an interrupted batch had left one outcome unrecorded in each of them. We reran each missing trial with its original seed, so every cell now has 15 completed runs. The revised Table 2 reports the updated values, and Table R8 (response to Q1) gives the successful/attempted counts.

  2. The w/o PI total: one cell was misprinted, not the total. The Reviewer's column sum is correct. The archived record for the PMOS-input OTA cell is 2/15, which the original Table 6 printed as 12/15. With 2/15, the 27 cells sum to the reported 161/405, and the decrease from the original full result (322/405) is 39.8 percentage points, as reported.

  3. The revised Table 6 reports the repeated batch. The revised Table 6 reports the repeated ablation with the main experiment's seeds (Table R3 in General Response 3). Without PI, initialization is random sampling, so this cell depends on the random initial designs that each seed draws, and the two batches use different seeds. We report these updated results in the updated manuscript so that each set of numbers matches its own records.

Overall, every baseline cell now rests on 15 runs, and the original w/o PI total and its 39.8-point decrease stand, with the PMOS-input OTA cell corrected to 2/15. The revised Table 2 reports 15 runs per cell, and the revised Table 6 reports the repeated ablation with the main experiment's seeds.


Q (Q2). Which w/o PI counts are correct, and why does Appendix B report 20.45 simulations per run while Table 7 reports 19.4?

Ans.

  1. w/o PI counts. With the corrected cell, the 27 cells sum to 161/405, and the decrease from the original full result (322/405) is 39.8 percentage points, as reported (see response to W1).

  2. Mean simulation cost. Appendix B's 20.45 simulations per run come from the ablation experiments, whereas Table 7's 19.4 comes from the main experiment. As the Reviewer notes, the two experiments are separate batches with the same benchmark and settings, so their means need not coincide. We have clarified this distinction in Appendix B of the revised manuscript.

  3. Consistency check. Repeating the ablation with the main experiment's seeds gives a full-pipeline success rate of 74.6%, compared with 74.8% in the main experiment (see Table R10 below).

Table R10. Repeated full-pipeline and component ablation using main-experiment seeds (27 circuits × 15 runs; 50 SPICE evaluations per run).

Removed moduleSuccessful runsPass@1Change vs. full (pp)
None (Full)302/40574.6%—
PI114/40528.1%−46.4
PGR271/40566.9%−7.7
SV278/40568.6%−5.9

Overall, the original w/o PI total and its 39.8-point decrease stand after the one-cell correction, and the two simulation-cost figures come from separate batches rather than inconsistent settings. The revised manuscript and Appendix B distinguish these experimental batches.


Our response is continued in Part 2/3 below.


AuthorsOpenReview ↗

Response to Reviewer w8vm (Part 2/3)

(Continued from Part 1/3 above.)


Q (W2). Does the w/o PI ablation isolate the LLM contribution, how does SizingAgent differ from prior equation-based feedback methods, and how should AutoSizer be classified in Table 1?

Ans. Thank you for these precise points. We agree on both: the w/o PI ablation measures the initialization module as a whole rather than the LLM alone, and SizingAgent should be positioned more precisely against the four works cited. We have added controls that isolate the LLM and revised the positioning as follows.

  1. Attribution of the w/o PI drop. Removing PI replaces both the targets of the LLM and the gm/Id lookup with random sampling, so its drop measures the module as a whole, and we have revised §4.5 accordingly. The controls in General Response 2 isolate the LLM: with the same gm/Id tables and downstream modules, LLM targets succeed in 55/85 runs, compared with 48/85 for fixed targets, and removing LLM guidance from both PI and PGR lowers success from 199/270 to 127/270. Our response to Q3 gives further detail.

  2. Positioning against earlier equation-based feedback systems. As shown in the table below, the four works differ from SizingAgent in what they produce and in what simulation feedback controls.

WorkOutputWhat simulation feedback controlsScreens each proposed update by measured sensitivity before simulation
White-Op [2]Behavioral op-amp design from a symbolic optimization problemRefines the design using theory–simulation discrepanciesNot described
Bujana and Karsilayan [3]Executable analytical sizing functionCalibrates device parameters with operating-point simulations; feeds prediction errors into later roundsNot described
NEMESIS [4]Performance equations (models) from netlists and schematicsRepairs the equation predictions with SPICENot applicable (builds models, no sizing loop)
AutoSizer [5]Circuit sizing with optimizers selected and configured in a two-loop frameworkSelects optimizers; adapts ranges and active variablesNot described
SizingAgentTransistor dimensions and bias parameters within a fixed simulation budgetIndividual parameter updates, proposed by PGR and screened by SVYes (SV)

SizingAgent's distinct element is the action-level filter: PGR proposes individual parameter updates, and SV screens the direction of each against the local response needed to close a specification gap, before the joint candidate is simulated.

  1. Table 1(d): AutoSizer. We agree that AutoSizer has a simulation-grounded loop with explicit feedback to the LLM (§§3.6 and 4.6 of [5]). We have changed its entry in Table 1(d) to Partial (marked in blue), with a note that its feedback adapts optimizer selection, ranges, and active variables rather than individual parameter updates.

  2. Evidence for the directional filter. In the repeated ablation, removing SV lowers success by 5.9 percentage points and removing PGR by 7.7 points. SV flags a directional conflict in 615 of 3,323 recorded checks (18.5%) before the candidate is simulated. In the telescopic OTA case (General Response 3), SV rejects an L5 update whose measured gain sensitivity (−0.200) contradicts the equation-based proposal. PGR revises it to W7 × 1.05, SPICE confirms that all four specifications are met, and a same-state offline replay confirms that the rejected update would have failed.

Overall, the PI ablation supports the initialization module as a whole, and General Response 2 isolates the contribution of the LLM. Relative to the four works, SizingAgent's distinct contribution is to screen individual sizing actions by their measured direction before simulation, within a fixed budget. The revised manuscript updates Table 1(d), discusses the four works in §2, and adds the telescopic OTA case to Appendix B.2 and the same-state comparison as Figure 4.


References

[2] Chen et al., White-Op, arXiv:2601.21321 (2026).

[3] Bujana & Karsilayan, Self-Calibrating LLM-Based Analog Sizing, arXiv:2604.07387 (2026).

[4] Ghosh et al., NEMESIS, arXiv:2607.05657 (2026).

[5] Yu et al., AutoSizer, arXiv:2602.02849 (2026).


Our response is continued in Part 3/3 below.


AuthorsOpenReview ↗

Response to Reviewer w8vm (Part 3/3)

(Continued from Part 2/3 above.)


Q (Q1). What are the successful/attempted counts behind Table 2, and how are infrastructure errors and non-converging simulations counted?

Ans. Thank you for this question. Every LLM-baseline cell now has 15 attempted runs (Table R8). A simulation that does not converge counts as a used, unsuccessful evaluation, and an infrastructure interruption is resumed, so each run yields exactly one outcome.

Table R8. LLM baselines with Gemini 2.5 Flash; successful runs out of 405 attempted (27 circuits × 15 runs), Pass@1, and Pass@5 averaged over circuits.

MethodSuccessful / attempted ↑Pass@1 ↑Pass@5 ↑
ADO-LLM131/40532.3%57.9%
AutoSizer87/40521.5%38.2%
EEsizer170/40542.0%70.0%
LEDRO94/40523.2%42.2%

These counts determine the Pass@1 entries of the revised Table 2. Pass@5 is computed for each circuit from its 15 runs and then averaged over the 27 circuits.

Infrastructure error handling.

  • Current protocol (stated in the supplement). A submitted candidate uses one SPICE evaluation even if ngspice does not converge or a required measurement is missing, and it cannot count as a success. Infrastructure interruptions are logged separately, and each resumed run contributes one completed outcome to the denominator.

  • Historical batch. The earlier runner treated a propagated exception as an unfinished run and a completed run without a feasible design as a failure; its logs cannot fully separate internally caught API errors from SPICE non-convergence. The GA run excluded in the earlier Table 8 came from this batch.

  • Refreshed GA and BO records. Each has 405 completed runs, with 70 and 95 successes, and the revised Table 8 lists 15 GA runs per circuit.

Overall, every baseline entry now traces to 15 attempted runs per circuit, and the supplement states how failed simulations and interruptions are counted.


Q (Q3). Can the effect of LLM-guided initialization be isolated from device-table access, and what do recorded SV traces reveal about false rejections, false acceptances, and stale sensitivities?

Ans. We thank the Reviewer for these valuable questions. Yes to both: a matched control separates the targets of the LLM from the device tables and the initialization procedure, and the recorded traces quantify the errors of SV in both directions.

  1. LLM initialization versus the same device tables. Holding the gm/Id tables and the downstream PGR and SV fixed, and changing only the source of the targets, LLM-generated targets succeed in 55/85 runs, compared with 48/85 for fixed targets (General Response 2). This isolated difference is modest. The combined control, which removes LLM guidance from both PI and PGR across all 27 circuits, is large and statistically clear: 199/270 (73.7%) versus 127/270 (47.0%), Fisher's exact p < 10−9. The gm/Id conversion is deterministic given the numerical targets: removing the equation text from 26 archived PI outputs while keeping the targets fixed yields identical initial dimensions. The archived lookup check is also reported in Appendix I.2.

  2. SV errors in both directions. Across 3,323 action-level checks, SV finds 2,253 directions consistent with the stored sensitivities (67.8%), flags 615 as conflicts (18.5%), and marks 455 as weak (13.7%). Because one LLM proposal can contain several actions, we also audited proposals against their SPICE outcomes: among 1,415 archived proposals, the directional rule would reject 440 (45 later retained by SPICE) and admit 975 (92 later rolled back by SPICE). The two errors have different costs. An admitted ineffective proposal costs one SPICE evaluation, which then exposes it; a rejected action returns to PGR for revision, as in the telescopic OTA case. SV is therefore a screen, and SPICE remains the final judge of every candidate. (Section 4.7, Figure 4; Appendix B.2)

  3. Stale sensitivities. These counts cover all recorded checks; we have not yet broken them down by the age of the sensitivities.

Overall, the matched control shows that the LLM's targets add value beyond the device tables, and the audit shows that SV errs in both directions while its net effect is positive: removing it lowers success by 5.9 percentage points.

We thank the Reviewer again for the careful audit, which has made our reported results more accurate and our contribution more precisely stated.

Reviewer_492AOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SizingAgent, a training-free LLM framework for analog circuit sizing with a fixed topology. It has three components:

  • Physics Initializer (PI): an LLM derives small-signal relations and gm/Id-style targets from the netlist, and a gm/Id lookup table built from SKY130 sweeps maps these to device dimensions. It uses up to 3 multi-start candidates.
  • Physics-Guided Reasoner (PGR): an LLM emits a typed five-field reasoning chain (observation, equation, constraint, diagnosis, action) with at most two parameter-scaling actions per round.
  • Sizing Verifier (SV): each proposed action’s direction is checked against a finite-difference sensitivity matrix, estimated once at a reference design with η = 0.2. Accepted actions are applied jointly and simulated once.
Critical strengths
  • A controlled, reproducible evaluation substrate. The authors fix the PDK, simulator, backbone, spec set and budget across methods, and release code, circuits, Sobol feasibility data, seeds and prompts. Prior LLM sizers report on incompatible PDKs, simulators, budgets and spec sets, so a shared testbed is useful to the community, to some extent.
  • The empirical gains are large and robust across backbones. Pass@1 is 69.6–76.1% across four commodity LLMs, and the method averages about 20 simulations per run. Even allowing for the concerns below, the end-to-end pipeline appears practically effective on the authors’ benchmark.
Critical weaknesses

1 Limited benchmark scale and scope:27 circuits, mostly textbook blocks. the four "hard" tier are not too hard to solve, where several baselines are able to solve them.

2 No held-out evaluation. The benchmark, spec targets (tuned via pilot simulations, F.4) and prompts were developed together, so the results may reflect fit to this suite rather than generalization.

3 The central claim that physics-grounded reasoning and verification drive the gains is not supported, and the “physics” framing is overstated. The equations are LLM generations from a hand-written design recipe, nothing in the pipeline checks whether the equation is correct for this topology.

Questions for authors

1 Why do the full-pipeline counts in Table 6 differ from the main-result counts in Table 7? for example current mirror 15 vs. 9, differential pair 12 vs. 15. and the system and benchmark are the same.

  1. How does the method behave as the parameter count grows (for example, 30–50 variables)
7 comments
AuthorsOpenReview ↗

Response to Reviewer 492A (Part 1/3)

We thank the reviewer for the careful assessment. Please find our point-by-point response below.


Q (W1). Is the 27-circuit benchmark too limited or easy, given that some baselines solve circuits in the hard tier?

Ans. Thank you for this suggestion. We agree that harder circuits make a stronger test, since several baselines solve our hard tier. We therefore evaluated the seven most complex amplifier sizing tasks in the public AnalogGym benchmark [1] (30–38 tunable parameters, about twice our largest circuit), where SizingAgent succeeds in 26/35 runs, compared with 18/35 for BO and 11/35 for ADO-LLM.

  1. The new circuits are substantially harder. Our benchmark has 27 SKY130 circuits with 4–17 tunable parameters and coupled specifications. The new circuits have 30–38, and the set is far from saturated: on Song DACFC and Qu2017 AZC tasks, no method succeeds in more than 2 of 5 runs.

  2. SizingAgent leads on every circuit. It outperforms ADO-LLM on all seven circuits and matches or exceeds BO on every circuit, with identical netlists, parameter bounds, specifications, testbenches, five seeds, and a 50-evaluation budget.

Table R1. Sizing performance on seven AnalogGym amplifier topologies (5 runs per circuit, 50 SPICE evaluations per run). Best result in each row is bold.

CircuitTunable parametersADO-LLMBOSizingAgent
Leung DFCFC1301/53/55/5
Peng IAC314/55/55/5
Song DACFC330/52/52/5
Peng ACBC333/51/54/5
HoiLee AFFC333/54/55/5
Yan AZ360/52/53/5
Qu2017 AZC380/51/52/5
Total—11/3518/3526/35

General Response 1 provides the experimental setup. Overall, on external circuits with about twice as many tunable parameters, where every method fails on some runs, SizingAgent keeps a clear lead over both baselines.


Q (W2). Since the benchmark, specifications, and prompts were developed together, how do we know the results generalize to unseen topologies?

Ans. Thank you for raising this point. We agree that a benchmark, specifications, and prompts developed together can overstate generalization. We now report two held-out evaluations, with SizingAgent's prompts and workflow frozen in both: held-out circuits and held-out LLM backbones.

  1. Held-out circuits. The seven AnalogGym amplifiers [1] were not part of our benchmark and played no role in developing SizingAgent. We selected all amplifier topologies with at least 30 tunable parameters, by size rather than by results. SizingAgent kept the same prompts, workflow, and hyperparameters, with no topology-specific tuning; we adapted only the circuit interfaces, specifications, and testbenches to our SKY130 flow, and fixed the specifications before running any method. SizingAgent succeeds in 26/35 runs, compared with 18/35 for BO and 11/35 for ADO-LLM (General Response 1).

  2. Held-out backbones. The same prompts, unchanged, also work with three LLMs not used in our main experiments. SizingAgent leads EEsizer and ADO-LLM with Gemini 3 Flash Preview (20/25 vs. 15/25 and 11/25) and Gemini 4 Argon (25/25 vs. 23/25 and 16/25), and it raises the open-weight Qwen3-32B from 10.1% to 59.0% Pass@1 on our 27 circuits (Table 3). In addition, a new backbone comparison is provided in Table R6.

Table R6. Newer-model comparison (5 circuits × 5 seeds; unsuccessful runs counted as 50 SPICE evaluations).

MethodGemini 3 Flash SuccessMean SPICE ↓Gemini 4 Argon SuccessMean SPICE ↓
ADO-LLM11/2535.0016/2527.60
EEsizer15/2523.2823/2513.04
SizingAgent20/2517.8425/256.00

Overall, SizingAgent transfers to unseen topologies and unseen backbones without any change to its prompts or workflow, which indicates generalization rather than fit to our suite. We have added both evaluations to the revised manuscript (Appendix I.1 (Table 11), Appendix I.4 (Table 14), and Table 3).


References

[1] CODA-Team, AnalogGym, https://github.com/CODA-Team/AnalogGym.


Our response is continued in Part 2/3 below.

AuthorsOpenReview ↗

Response to Reviewer 492A (Part 2/3)

(Continued from Part 1/3 above.)


Q (W3). Are the claims about physics-grounded reasoning overstated if the LLM-generated equations are not verified for the specific topology?

Ans. We thank the Reviewer for this important comment. We agree that SizingAgent does not verify its equations symbolically, and the revised manuscript says so. It does test each equation against this topology: SV checks the direction of every proposed update against SPICE-measured sensitivities of the actual circuit, and SPICE evaluates every resulting design.

Moreover, General Response 3 gives the full account; we summarize the main points here.

  1. The equations are hypotheses, and their predictions are tested. The LLM writes circuit relations that propose operating targets and update directions. The gm/Id tables map the targets to device dimensions using characterized device data, SV rejects any action whose direction conflicts with the measured sensitivity, and SPICE evaluates the joint candidate against all specifications. No step relies on an equation being correct.

  2. The check fires often. In 615 of 3,323 recorded actions (18.5%), the measured sensitivity contradicts the direction proposed by the LLM’s reasoning, and SV stops the action before the candidate is simulated.

  3. The check changes the outcome: the telescopic OTA case (Appendix B.2 and Figure 5). PGR uses Av ≈ gmRout to propose a longer L5, but the measured gain sensitivity to L5 is −0.200, so SV rejects the action. Because the gain sensitivity to W7 is positive (+0.399), PGR revises the update to W7 × 1.05, and SPICE confirms that all four specifications are met (gain 35.179 dB, GBW 15.136 MHz). Replaying the rejected length increase from the same state gives 34.546 dB and 14.791 MHz, both below target.

  4. The gains are attributed to the right components. In the repeated ablation (Table R3), removing PGR or SV lowers success by 7.7 and 5.9 percentage points, and removing the initialization module by 46.4 points. The last number measures the LLM targets and the gm/Id lookup together. Removing only the LLM guidance from PI and PGR, with the rest of the pipeline in place, lowers success from 73.7% to 47.0% (General Response 2).

Overall, the gains come from LLM-proposed actions that measured sensitivities filter and SPICE validates. We have revised the framing accordingly (Sections 3.5, 4.5, and 4.7), added the case study (Section 4.7, Figure 4, and Appendix B.2), and updated the ablation (Table 6).


Our response is continued in Part 3/3 below.


AuthorsOpenReview ↗

Response to Reviewer 492A (Part 3/3)

(Continued from Part 2/3 above.)


Q. Why do the Full Pipeline success counts differ between Tables 6 and 7, despite using the same system and benchmark (e.g., current mirror 15 vs. 9; differential pair 12 vs. 15)?

Ans. Thank you for the careful comparison of Tables 6 and 7. The two tables report separate batches of runs with different seeds. A circuit's count out of 15 varies between batches, but the totals agree within one run once the seeds are matched.

Both tables use the same system, benchmark, and protocol: 27 circuits, 15 runs per circuit, and 50 SPICE evaluations per run. The original ablation used different seeds from the main experiment. To check consistency, we reran the complete ablation with the main experiment's per-circuit seeds (Table R3 in General Response 3). Matched seeds do not make two runs identical, because the responses of the LLM vary from call to call.

Results. With matched seeds, the full SizingAgent reaches 302/405 (74.6%), within one run of the main result in Table 7 (303/405, 74.8%). Per-circuit counts vary more than the totals because each rests on only 15 runs. With two-sided Fisher's exact tests and Bonferroni correction for the 27 circuit-level comparisons, neither cited difference is statistically significant (adjusted p = 1.00 for the differential pair, 12 vs. 15; adjusted p = 0.46 for the current mirror, 15 vs. 9). Each module continues to contribute: removing the initialization module, PGR, or SV lowers success by 46.4, 7.7, and 5.9 percentage points.

Table R7. Repeated component ablation with the main experiment’s per-circuit seeds; successful runs out of 405 (27 circuits × 15 runs, 50 SPICE evaluations per run).

Removed module (role)Success ↑Pass@1 ↑Change vs. full (pp)
None (SizingAgent)302/40574.6%—
PI (sets initial device sizes)114/40528.1%−46.4
PGR (plans parameter updates)271/40566.9%−7.7
SV (checks update directions)278/40568.6%−5.9

Overall, the differences between Tables 6 and 7 reflect run-to-run variation between separate batches, not a different system or benchmark. With matched seeds, the totals agree within one run, and each module continues to improve success. The revised manuscript reports the updated ablation results in §4.5 (Table 4) and Appendix C (Table 6) and keeps the main result in Table 7.

Reviewer_8keyOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces SizingAgent, a training-free LLM framework that empowers LLM agents with physics-grounded knowledge for sizing optimization in analog design automation. It consists of three modules: physics initializer, which extracts small-signal equations from the netlist, physics-grounded reasoner, which analyzes specification gaps and derives targeted sizing updates, and sizing verifier, which checks the physical consistency of the proposed updates.

Critical strengths
  • The paper is well-written with a clear structure. It is easy-to-follow and get the main idea behind each modules.
  • The integration of physical-grounded knowledge is insightful and interesting.
  • The evaluation is conducted with different LLM families, providing comprehensive results.
Critical weaknesses
  • How about using general agent harness like codex, claude, pi, etc, directly for this task? Will they perform better than simply using the model itself? Will they even perform better than the proposed SizingAgent it they are allowed to use some design automation related tools?
  • How long it takes for the proposed SizingAgent to finish one task, e.g., several minutes or hours? any difference between different circuits?
  • For baselines presented in Table 2, do all of them use Gemini 2.5 Flash? If yes, since Gemini 2.5 Flash is relatively weak model, the comparison may not be that convincing.
Questions for authors

see weaknesses

5 comments
AuthorsOpenReview ↗

Response to Reviewer 8key (Part 1/3)

We sincerely thank you for the thoughtful and constructive feedback. Below, we provide detailed responses to each of your concerns.


Q. Can general-purpose agent harnesses such as Codex or Claude Code match SizingAgent when given access to the same circuit-design tools and SPICE feedback?

Ans. Thank you for this insightful suggestion. No: with the same model and the same SPICE evaluator, Codex and Claude Code succeed in 8/15 and 7/15 runs, compared with 14/15 for SizingAgent, which also needs about half as many simulations.

Setup. We ran two of the harnesses the Reviewer names, Codex and Claude Code, directly on the sizing task. All methods use Gemini 2.5 Flash, so only the framework around the model differs. We used five circuits from our benchmark that cover a basic input stage, multi-stage amplification with compensation, and cascode structures (differential pair, five-transistor OTA, two-stage Miller OTA, symmetric OTA, and folded-cascode OTA), with the same three seeds per circuit (15 runs per method). We tested two settings:

  • Without SPICE feedback: the direct LLM call or the harness returns one sizing, which SPICE then scores.

  • With SPICE feedback: Codex, Claude Code, and SizingAgent can each call the same SPICE evaluator up to 50 times. This gives the harnesses the design-automation tool access the Reviewer asks about.

A run succeeds only if every specification is met.

Table R4. General agent harnesses versus SizingAgent with Gemini 2.5 Flash (five original-benchmark circuits × three paired seeds; 50 SPICE evaluations maximum for feedback-enabled variants).

FrameworkSuccess without SPICE feedback ↑Success with SPICE feedback ↑Mean SPICE evaluations ↓
Direct LLM call0/15NANA
Codex1/158/1529.4
Claude Code0/157/1533.9
SizingAgentNA14/1516.6

NA : not applicable. The direct call receives no feedback, and SizingAgent always uses SPICE feedback.

Results. Without SPICE feedback, Codex and Claude Code succeed in 1/15 and 0/15 runs, close to the direct model call (0/15). SPICE access raises them to 8/15 and 7/15, whereas SizingAgent succeeds in 14/15 runs with 16.6 SPICE evaluations on average, compared with 29.4 and 33.9. Both differences in success are statistically significant (Fisher's exact test, p = 0.035 versus Codex and p = 0.014 versus Claude Code).

Overall, tool access alone does not close the gap. Given the same model and the same SPICE evaluator, SizingAgent succeeds in nearly twice as many runs as general-purpose harnesses and needs about half as many simulations, so its gain comes from its circuit-specific workflow. We have added this comparison to Appendix I.3 (Table 13) of the revised manuscript.


Our response is continued in Part 2/3 below.


AuthorsOpenReview ↗

Response to Reviewer 8key (Part 2/3)

(Continued from Part 1/3 above.)


Q. How long does SizingAgent take to complete a sizing task, and how does runtime vary across circuit types?

Ans. Thank you for this practical question. One SizingAgent run takes 121.2 s on average, and the time varies by circuit, from 15.7 s to 217.2 s. We recorded the wall-clock time of 270 SizingAgent runs, each under the 50-evaluation SPICE budget. Table R5 gives four representative circuits, from fast current-mirror and amplifier runs to slower multi-stage designs.

Table R5. Wall-clock time of SizingAgent per run; mean over the runs of each circuit and over all 270 runs.

CircuitMean wall-clock time per run (s)
Wilson current mirror15.7
Common-gate amplifier17.5
Inverter amplifier158.3
Three-stage OTA217.2
All 270 runs121.2

Runtime differs by about 14× across these circuits. Wall-clock time also depends on the hardware and the LLM service, so we also count SPICE evaluations, which do not. In the main experiment (405 runs, same 50-evaluation budget), SizingAgent needs 19.36 SPICE evaluations per run on average, compared with 41.99 for BO. By the 20th evaluation, it has succeeded in 260 of 405 runs, compared with 64 for BO (Figure 5).

Overall, SizingAgent completes a sizing task in about two minutes and needs fewer than half as many SPICE evaluations as BO. The simulation-efficiency analysis is in Appendix B (Figure 5) and Appendix I.5 (Table 15) of the revised manuscript.


Our response is continued in Part 3/3 below.


AuthorsOpenReview ↗

Response to Reviewer 8key (Part 3/3)

(Continued from Part 2/3 above.)


Q. Do all Table 2 LLM baselines use Gemini 2.5 Flash, and does SizingAgent remain effective with newer, stronger models?

Ans. Thank you for this question. Yes, all four LLM baselines in Table 2 use Gemini 2.5 Flash, so that differences come from the sizing workflow rather than the model. SizingAgent keeps its lead with two newer models, Gemini 3 Flash Preview and Gemini 4 Argon, and it also lifts the open-weight Qwen3-32B from 10.1% to 59.0% Pass@1. Please note that baselines GA and BO use no LLM.

Setup. We ran SizingAgent, EEsizer, and ADO-LLM with Gemini 3 Flash Preview and Gemini 4 Argon on the five circuits of W1, with the same five seeds per circuit (25 runs) and the same 50-evaluation SPICE budget. A run succeeds only if it meets every specification.

Table R6. SizingAgent and LLM baselines with two newer models; successful runs out of 25 (5 circuits × 5 seeds), and mean SPICE evaluations per run with unsuccessful runs counted as 50.

MethodGemini 3: SuccessGemini 3: Mean SPICE ↓Gemini 4: SuccessGemini 4: Mean SPICE ↓
ADO-LLM11/2535.0016/2527.60
EEsizer15/2523.2823/2513.04
SizingAgent20/2517.8425/256.00

Results. All three methods improve with the stronger models, and SizingAgent leads with both. With Gemini 4 Argon, the success gap to EEsizer narrows (25/25 versus 23/25), but SizingAgent needs less than half as many SPICE evaluations (6.00 versus 13.04), so its efficiency advantage grows as the model gets stronger. For reference, with Gemini 2.5 Flash on the same five circuits, SizingAgent succeeds in 67/75 runs, compared with 15/75 for EEsizer and 13/75 for ADO-LLM (15 seeds per circuit, a separate set of runs).

Beyond Gemini: an open-weight model. On the full 27-circuit benchmark, Qwen3-32B rises from 10.1% Pass@1 with direct parameter generation to 59.0% with SizingAgent, and it solves 23 of 27 circuits at least once. A locally run model also lowers inference cost and keeps proprietary circuit data in-house.

Overall, SizingAgent's advantage does not rely on a weak backbone: it leads the LLM baselines with newer models, its efficiency advantage grows with model strength, and it improves an open-weight model by 48.9 percentage points. The Qwen3-32B results are in Table 3 of the revised manuscript, and the Gemini 3 and Gemini 4 results are in Appendix I.4 (Table 14).

We thank the Reviewer again for these suggestions, which led to the harness, runtime, and backbone comparisons above.

Comments on the paper

6 comments
AuthorsOpenReview ↗

General Response: Summary of New Results

We thank all reviewers for their careful reading and constructive feedback. We appreciate their recognition of the clear framework design and integration of circuit knowledge (8key), the reproducible evaluation and performance gains across LLM backbones (492A), and the practical sensitivity-based action checks and informative module comparisons (w8vm).

In response, we conducted substantial additional experiments and analyses that we believe have significantly strengthened the paper. These include evaluations on more complex external circuits, controlled studies isolating LLM guidance, comparisons with general-purpose agent harnesses and newer LLM backbones, and a case study demonstrating how physics-grounded feedback guides actual sizing decisions. Together, these additions provide stronger evidence for SizingAgent's generalization, effectiveness, and practical value. Our key findings are:

  1. The gains transfer to more complex external circuits. On seven AnalogGym amplifiers with 30–38 tunable parameters, SizingAgent succeeds in 26/35 runs, compared with 18/35 for Bayesian optimization and 11/35 for ADO-LLM (Table R1).

  2. LLM guidance adds value beyond device tables and numerical search. With the same gm/Id tables, LLM-guided initialization improves success from 48/85 to 55/85. Starting from identical designs, LLM-guided update planning achieves 48/85 successes versus 46/85 without LLM guidance; the full workflow achieves 199/270 versus 127/270 for the no-LLM control (Table R2).

  3. The circuit-specific workflow outperforms general-purpose agent harnesses. Under the same Gemini 2.5 Flash backbone and SPICE evaluation budget, SizingAgent succeeds in 14/15 runs, compared with 8/15 for Codex and 7/15 for Claude Code (Table R4).

  4. The advantage extends to stronger LLM backbones. SizingAgent leads both EEsizer and ADO-LLM with Gemini 3 Flash Preview (20/25 vs. 15/25 and 11/25) and Gemini 4 Argon (25/25 vs. 23/25 and 16/25), while requiring less than half the SPICE evaluations of the strongest baseline with Gemini 4 (6.00 vs. 13.04; Table R6).

  5. Physics-grounded feedback improves sizing decisions. In the repeated ablation, removing PI, PGR, or SV reduces success by 46.4, 7.7, and 5.9 percentage points, respectively (Table R3). A recorded telescopic OTA case further shows how a sensitivity check redirects a proposed update toward a design meeting all specifications.

Below, we detail these results in the general responses and address each reviewer’s specific concerns.

AuthorsOpenReview ↗

General Response 1: External Topology Evaluation

Q. Can SizingAgent generalize to larger, external amplifier circuits beyond its original benchmark?

Ans. While our dataset is one of the largest benchmarks in the field, Reviewer 492A noted that our 27 benchmark circuits are mostly standard building blocks, that several baselines solve the hard tier, and that the benchmark, specifications, and prompts were developed together. We agree that harder circuits that played no role in developing SizingAgent give a stronger test. We therefore evaluated SizingAgent on seven external amplifiers with 30–38 tunable parameters, about twice the largest circuit in our benchmark, and it succeeds in 26/35 runs, compared with 18/35 for BO and 11/35 for ADO-LLM.

Task Selection. Industrial designs are rarely public [8], and proprietary PDKs complicate reproducible evaluation [7]. We therefore took the 7 most complex sizing tasks (30–38 parameters) we could identify in publicly available benchmarks from the AnalogGym [1], and adapted them to our interface and SKY130 evaluation flow. The circuits were selected by size alone, not by results, and none of them is part of our 27-circuit benchmark.

Setup and results. We compared SizingAgent, Bayesian optimization (BO), and ADO-LLM on these seven circuits with the same five seeds per circuit (35 runs per method), a 50-SPICE-evaluation budget per run, and identical netlists, parameter bounds, specifications, and testbenches. SizingAgent and ADO-LLM used the same Gemini 2.5 Flash backbone. As shown in Table R1, SizingAgent succeeds in 26/35 runs (74.3%), compared with 18/35 (51.4%) for BO and 11/35 (31.4%) for ADO-LLM. SizingAgent outperforms ADO-LLM on all seven circuits and matches or exceeds BO on every circuit. These results support transfer to external amplifier topologies with larger sizing spaces under the same simulation budget.

Table R1. Successful runs out of 5 per circuit on the seven AnalogGym amplifiers (50 SPICE evaluations per run). Higher is better; the best result in each row is in bold.

CircuitTunable parametersADO-LLMBOSizingAgent
Leung DFCFC1301/53/55/5
Peng IAC314/55/55/5
Song DACFC330/52/52/5
Peng ACBC333/51/54/5
HoiLee AFFC333/54/55/5
Yan AZ360/52/53/5
Qu2017 AZC380/51/52/5
Total11/35 (31.4%)18/35 (51.4%)26/35 (74.3%)

Overall, on unseen circuits with about twice as many tunable parameters as our benchmark, and without any change to its prompts or workflow, SizingAgent keeps a clear lead over both an LLM-based and a numerical baseline. We have added this experiment to Appendix I.1 (Table 11) of the revised manuscript.


References

[1] CODA-Team, AnalogGym, https://github.com/CODA-Team/AnalogGym.

[7] Leibl et al., LOADBench, arXiv:2610.21591 (2026).

[8] Li et al., AnalogFed, arXiv:2507.15104 (2025).

AuthorsOpenReview ↗

General Response 2: Contribution of LLM Guidance

Q. What is the independent contribution of LLM guidance beyond device tables and numerical optimization?

Ans. Reviewers 492A and w8vm raised an important question: How much does the LLM contribute beyond the device tables and numerical sensitivity search? While our ablation study (Table 4) measures the effect of removing the entire physical initialization module, we further added three controls that remove only the LLM and keep everything else fixed. Removing LLM guidance from both initialization and update planning lowers success from 73.7% to 47.0%.

  • PI, w/ vs. w/o LLM. LLM-generated operating targets versus fixed targets, with the same gm/Id lookup tables and the same downstream PGR and SV.

  • PGR, w/ vs. w/o LLM. Structured LLM reasoning versus numerical sensitivity search, from identical initial designs and with the same specifications, SPICE feedback, and sensitivities.

  • SizingAgent, w/ vs. w/o LLM. The full pipeline versus a control without LLM guidance in either module. The control initializes 17 circuits from fixed targets and device tables and the other 10 with a common parameter-bound rule, and then applies numerical sensitivity search to all 27 circuits.

Table R2. Contribution of LLM guidance; successful runs / total runs, 50 SPICE evaluations per run. For PI and PGR, “w/o LLM” removes the LLM only from the module tested; for SizingAgent, from both.

Comparisonw/ LLMw/o LLMChange (pp)
1. PI (initialization)55/85 (64.7%)48/85 (56.5%)+8.2
2. PGR (update planning)48/85 (56.5%)46/85 (54.1%)+2.4
3. SizingAgent (both modules)199/270 (73.7%)127/270 (47.0%)+26.7

Results. With the same device tables, LLM-generated targets raise success from 48/85 to 55/85 (+8.2 percentage points). From identical starting designs, LLM-guided update planning adds a smaller gain, 48/85 versus 46/85. Removing LLM guidance from both modules has the largest effect: 199/270 versus 127/270 (+26.7 points; Fisher's exact test, p < 10−9). Without the LLM, initialization must fall back on fixed targets or a generic parameter-bound rule, whereas PI derives circuit-specific operating targets for every circuit from its topology.

Overall, the LLM adds value beyond the device tables and the numerical search, and the effect is largest when it guides both initialization and update planning. We have revised §4.5 to attribute the w/o PI drop to the initialization module as a whole, and added these controls to Appendix I.2 (Table 12) of the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.