acceptodds
Under review as a conference paper at ICLR 2027

AnaFlash: AI-driven Sizing of Analog Circuit using Fast Analytical Circuit Models

Abstract

Analog circuit sizing requires jointly satisfying coupled performance specifications whose response to device dimensions depends on the circuit topology, operating point, and technology process. Recent AI-based workflows either require substantial circuit-specific training data or ask large language models (LLMs) to directly propose numerical device updates from simulator feedback while the absense of an analytical solver of the equation of the circuit creates trial and error feedback loop approaches. This paper AnaFlash, a hybrid AI-analytical sizing framework in which a fast circuit model, rather than the LLM, drives numerical exploration. Starting from a physically valid seed, AnaFlash derives a SPICE-based quality-metric sensitivity model and computes sizing updates through constrained optimization. SPICE validates each candidate, while a lightweight LLM is invoked only when the analytical engine detects a locally unproductive sizing update. The LLM then revises metric priorities rather than transistor dimensions. Across six OTA topologies and five filter topologies, three technology processes, and thirty specification configurations per circuit–process pair, AnaFlash achieves training-free transfer with OTA success and filter success. AnaFlash requires only between 17 and 236 SPICE simulations on average for the circuit sizing to converge and achieves up to lower runtime than recent sizing workflows on shared benchmarks.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
analog circuit sizing, electronic design automation, large language models, analytical modeling

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_QE22OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

AnaFlash is a training-free, hybrid AI-analytical framework for automated analog circuit sizing that decouples numerical parameter exploration from high-level trade-off adaptation. Starting from a valid operating point initialized via gm/IDg_m/I_D lookup tables, the framework constructs a local logarithmic design Jacobian using SPICE small-signal parameters and adjoint Modified Nodal Analysis (MNA) sensitivities to compute numerical device updates via a trust-region Chebyshev linear program. A lightweight LLM (Qwen2.5-7B) is invoked only when the analytical solver stalls in a locally unproductive region, adjusting metric priority weights rather than guessing transistor dimensions. Validated across 990 sizing tasks spanning 6 OTA topologies, 5 GMCGMC filters, and 3 technology nodes, AnaFlash achieves 100% OTA and 94.7%–100% filter success rates in 17–67 average SPICE simulations, delivering up to a 56×56\times runtime speedup over recent RL and LLM-agent baselines.

Critical strengths
  1. Restricting the LLM to adjusting priority weights (ww) when stalled, while using an analytical sensitivity solver for numerical updates, avoids the poor mathematical reasoning of LLM-only sizing flows.

  2. The analytical engine alone solves 97% of OTA tasks, proving the physical model drives the 56×56\times speedup, while the 7B LLM resolves the remaining 3% of stalled cases and improves power/area.

Critical weaknesses
  1. Adjoint MNA (G+sC\mathbf{G} + s\mathbf{C}) only captures linear AC metrics; the paper omits how JijJ_{ij} is derived for large-signal Slew Rate or DC operating-point shifts without finite-difference sweeps. Also, "Infeasibility Proof" and "Verified Explainability" are claimed in Table 1 but never evaluated.

  2. The Ethics Statement calls the framework "ThunderGen" (line 502), Table 10 has corrupted column alignments.

Questions for authors
  1. How is the Jacobian JijJ_{ij} computed for large-signal metrics (Slew Rate) and DC operating-point shifts (∂OP/∂kj\partial \text{OP}/\partial k_j) using only linear AC adjoint MNA (G+sC\mathbf{G} + s\mathbf{C}) without perturbation sweeps?

  2. How many offline SPICE simulations are needed to build the Stage 1 gm/IDg_m/I_D LUTs, and what fraction of tasks already meet specifications at the initial seed x0x_0?

  3. Can you provide empirical validation for the "Infeasibility Proof" and "Verified Explainability" claims in Table 1?

2 comments

We thank the reviewer for the positive assessment, and for noting that restricting the LLM avoids the weak numerical reasoning of LLM-only flows. This is now the paper's central message.

Q1. Jacobian for large-signal metrics and DC shifts: For slew rate, the engine uses the operating-point estimate SR^=2max⁡d∈Dout∣ID,d∣/CL\widehat{\mathrm{SR}} = 2\max_{d \in \mathcal{D}_{\mathrm{out}}} \lvert I_{D,d}\rvert/C_L, where Dout\mathcal{D}_{\mathrm{out}} is the set of devices connected to the output and the currents ID,dI_{D,d} come from SPICE. For DC shifts, it linearises the DC Kirchhoff equations F(V,k)=0F(V, k) = 0 and solves GDC,∂V/∂ln⁡kj=−∂F/∂ln⁡kjG_{\mathrm{DC}}\\, \partial V/\partial \ln k_j = -\partial F/\partial \ln k_j for each knob kjk_j. It then combines these voltage shifts with local device derivatives and evaluates the metric on the perturbed states by centred differences in model space. The device derivatives come from perturbed isolated-device instances, batched into one SPICE operating-point run, so no full-circuit perturbation simulation per knob is needed. Against full-circuit SPICE central differences (5T and two-stage OTAs, PTM 45 nm), the geometry elasticities agree within 0.03 and 0.014 (absolute), and the bias elasticities within about 4%. This is now described in Section 3.2.

Q2. Offline LUT cost. The LUTs are built once per technology and reused across all circuits and specifications. Each polarity and channel length is one batched ngspice operating-point deck, with a 5 mV VGSV_{GS} grid and 9 VDSV_{DS} points:

TechnologyDecksDevice operating pointsBuild time
PTM 45 nm815,912~4.6 s
SkyWater 130 nm1238,988~81.7 s
GF 180 nm1271,388~47.8 s

These are single-device evaluations rather than circuit simulations. They are not included in the per-task counts, and Appendix A.2 (Table 9) now reports them.

Q3. Tasks met at the starting design. On the original benchmarks, the starting design already met 120 of 540 OTA tasks (22.2%), which is consistent with uXLa's point that those targets were modest. On the new benchmarks described in our response to uXLa (W3), it meets 50 of 320 hard OTA runs (15.6%) and none of the 450 PTM 45 nm filter runs, so the optimisation loop does the work. These fractions are now in Appendix A.2.

Q4. Infeasibility and explainability. These are now evaluated (general response, point 2). All 42 challenging cases were flagged with a quantitative reason and none had a satisfying design. The engine's sensitivity claims agree with SPICE finite differences in 569 of 573 cases. See Sections 4.3–4.4 and Appendices A.4–A.5.

Ethics Statement and Table 10. Both are fixed.

My concerns have been largely addressed. I'd like to maintain my score.

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

The paper targets analog circuit sizing and argues that both learned-surrogate flows (which need 0.4M–3.6M simulated examples) and LLM-agent flows (which ask a language model to propose numerical device updates from simulator feedback) are mismatched to the task — the former because of training cost and transfer, the latter because LLMs are unreliable at symbolic/analytical circuit reasoning (citing CircuitSense's sub-19% accuracy).

AnaFlash is a three-stage hybrid flow in which the analytical model, not the LLM, drives numerical exploration:

  1. Sizing initialization. The unsized netlist is parsed into a bipartite graph; functional transistor groups (differential pairs, current mirrors, cascodes, bias branches) are recognized; first-order relations map specifications to g_m and I_D requirements, propagated through mirror ratios; widths follow W_i = I_{D,i} / J_D(L_i, (g_m/I_D)_i) from a pre-characterized LUT, with symmetry constraints enforced and a SPICE operating-point check on saturation margins.
  2. Analytical model extraction. A SPICE operating-point analysis supplies g_m, g_mb, g_ds and parasitics, stamped into a sparse MNA system Y(s)z(s) = b(s), Y = G + sC. Adjoint sensitivities ∂y/∂p = −λᵀ(∂G/∂p + s ∂C/∂p)z are computed once per response and propagated by the chain rule to physical knobs, giving a log-log design Jacobian J_ij = d ln m_i / d ln k_j. Topology/PDK reachability bounds (intrinsic gain, transconductance efficiency) are checked up front.
  3. Hybrid sizing optimization. In log coordinates u = ln k, a trust-region weighted-Chebyshev LP min_Δu max_i (w_i ṽ_i) s.t. |Δu_j| ≤ r_j (plus linearized saturation-margin and PDK constraints) produces each update; SPICE validates it; a gain ratio ρ compares predicted and observed reduction in V_w(u) = max_i (w_i v_i), driving acceptance, trust-region update, or directional backtracking. Only when backtracking also fails is a Qwen2.5-7B LLM invoked, and only to return schema-validated revised priority weights w — it cannot touch device dimensions, PDK constraints or SPICE results. After convergence the LLM writes a human-readable explanation.

Evaluation. Six OTA topologies (5T, class-AB 5T, current-mirror, two-stage Miller, telescopic cascode, fully differential folded cascode) and five Gm–C filter topologies (LP/BP/HP biquads, 3rd- and 4th-order LP), each in PTM 45 nm, SkyWater 130 nm and GF 180 nm, with thirty fixed specification configurations per circuit–process pair: 30 × 11 × 3 = 990 sizing tasks, no training or fine-tuning. Reported: 540/540 OTA success, 94.7–100% filter success, 15–66.5 mean SPICE calls, 19 s–3.7 min wall clock. Table 5 compares against published AnaCraft and AnaFlow numbers plus an agent-only AnaFlow reimplementation. Tables 6–8 ablate priority adaptation (no LLM / random / Qwen2.5-7B), post-feasibility design quality, and LLM scale (7B/14B/32B).

Critical strengths

S1. The architectural choice is right, and the paper's own ablation is the cleanest evidence in the literature for it. Putting the numerical update in an adjoint-sensitivity/trust-region solver and restricting the LLM to a schema-validated weight interface — explicitly barred from touching device dimensions, PDK constraints or SPICE results — is the correct division of labour for this problem, and it is implemented with discipline (log-domain Jacobian so that relative sensitivities are scale-free; a gain-ratio acceptance test with directional backtracking; linearized saturation margins carried as constraints so candidates stay physical). Table 6's "No LLM adaptation: 524/540 = 97.0%" is, to my knowledge, the first published number quantifying how much of an LLM-assisted sizing flow's performance is actually attributable to the non-LLM machinery, and the authors report it rather than hiding it. I want this result in the literature.

S2. Simulation efficiency on the shared topologies is a real and large improvement, and the artifact choices are reproducible. 21 SPICE calls for the two-stage Miller OTA and 45 for the folded cascode, versus 205–360 for AnaFlow and 1,035–1,808 for AnaCraft, is roughly an order of magnitude on the metric that actually matters and that is platform-independent (the paper is right to privilege simulation count over wall clock). The mean 15–66.5 calls across 990 tasks is consistent with a seed that is already near-feasible plus a well-conditioned local model. Using open-weight Qwen2.5 models and open PDKs (SkyWater 130 nm, GF 180 nm) rather than a frontier API and a proprietary foundry node makes the flow independently reproducible, which is unusual and welcome in this subfield.

Critical weaknesses

W1. The paper's central contribution — LLM-based priority adaptation — is statistically indistinguishable from random weight perturbation once the paper's own Tables 6 and 8 are read together. This is decisive, because the LLM is the only machine-learning component and therefore the entire basis for an ICLR submission.

  • Table 6: analytical engine alone 524/540 (97.0%); random priority updates 1,597/1,620 (98.5%); Qwen2.5-7B 540/540 (100%). The LLM's entire measured contribution over random numbers is 1.5 percentage points.
  • Table 8: Qwen2.5-14B 534/540 (98.9%), Qwen2.5-32B 532/540 (98.5%). Larger models are strictly worse, and the 32B controller lands exactly on the random baseline.
  • Pooling the three LLM configurations, which is the right estimate of "LLM-based priority adaptation" as a policy class, gives 1,606/1,620 = 99.1% against random's 1,597/1,620 = 98.5%. A two-proportion test gives z ≈ 1.49, p ≈ 0.14 — not significant. The headline "100%" is obtained by reporting the one model size that happened to score perfectly; under a 99% per-run success process, P(540/540) ≈ 0.4%, and under 99.5% it is ≈ 7%, so a perfect run is not improbable enough to carry the claim on its own.
  • The comparison is also not budget-matched in the direction that matters. Random adaptation is reported as a per-run rate aggregated over three runs per task, while the LLM gets one run. If random adaptation has a 1.5% per-run failure rate, two independent restarts drive failure to ≈ 0.02% — i.e. random restarts match or beat the LLM at a cost of one or two extra sizing runs (~25–50 SPICE calls), versus the LLM's 4.7 calls and 3.13k tokens per run. The paper never reports "best of N random restarts," which is the baseline its own numbers demand.
  • Table 7 reinforces this. On post-feasibility design quality the 7B controller's win rate against random is 50% for SkyWater power (ratio exactly 1.000×) and 53% / 53% for PTM and SkyWater area — three coin flips out of six cells. The text reports only the favourable cells ("reduces mean power … in PTM and GF", "also reduces area in most GF tasks").
  • Nothing in the paper tests the hypothesis these numbers suggest: that the LLM functions as a stochastic weight jitter. The obvious checks are absent — does the revised w correlate with the sensitivity structure it is shown? Does a trivial rule (raise the weight of the largest unresolved violation; cycle priorities round-robin) match it? Does shuffling the structured summary before giving it to the LLM change anything? If LLM capability is irrelevant (7B ≥ 32B) and random is within noise, the parsimonious conclusion is that any perturbation escapes the stalled trust region, and the paper's framing — "lightweight LLMs can effectively support analog design" — is not established.

W2. Table 5's head-to-head comparison is not valid as constructed, so the efficiency claim that anchors the abstract is unsupported.

  • AnaCraft solves a harder problem. Its title, in this paper's own reference list, is "AnaCraft: Duel-play probabilistic-model-based reinforcement learning for sample-efficient PVT-robust analog circuit sizing optimization." AnaFlash sizes at nominal only — there is no corner analysis, no Monte-Carlo mismatch, no yield constraint anywhere in the paper. Comparing 21–45 nominal simulations against 1,035–1,808 PVT-robust simulations and reporting it as a 17–56× speedup compares different tasks. A PVT-robust sizer must evaluate each candidate across corners, which alone accounts for a large multiple.
  • Published numbers from different setups. Table 5's AnaCraft and AnaFlow rows are "reproduced from the respective papers," so the specification targets, simulator, netlists, convergence criteria and host hardware all differ. The paper states only that the topologies match ("the two-stage Miller and fully differential folded-cascode OTAs considered by such state-of-the-art works"), never that the specification targets are the same. Given W3, this is the crux: if AnaFlash's targets are looser, the simulation-count gap is not attributable to the method.
  • The wall-clock comparison is not meaningful. AnaFlow's times are for API-served GPT-o3 and Gemini-2.5-Flash, where network latency dominates; AnaFlash runs a local Qwen2.5-7B. A 56× wall-clock ratio between those two deployment modes says nothing about algorithmic efficiency. The simulation-count comparison (≈10×) is the defensible one and should be the headline.
  • Minor but confusing: "up to 56× lower runtime" matches only one of the six available Table 5 pairings (Gemini-2.5-Flash two-stage, 16 min vs 17 s); GPT-o3 on the same circuit (46 min vs 17 s) implies ≈162×, and AnaCraft on the folded cascode implies ≈17×. Please state which comparison produces 56×.

W3. The benchmark's specification targets are far below what these topologies deliver, and the "thirty configurations" span ranges too narrow to support the robustness claim — so 990 tasks are not 990 distinct problems. Contribution 2 is the "extensive and realistic experimental campaign," so this matters.

  • The OTA targets are undemanding. Table 2 gives DC gain targets of 20.09–34.15 dB across all thirty configurations. A two-stage Miller OTA routinely delivers 60–80 dB, and a telescopic cascode 70–90 dB, in all three of these nodes. Requiring ≤ 34 dB from a telescopic cascode or a two-stage amplifier is a target that a g_m/I_D-initialized seed should very nearly meet before optimization begins — which is consistent with the reported 15–31 mean SPICE calls and with 540/540 success. GBW 14.85–30.65 MHz, PM 45.02–57.52° and input-referred noise 75.9–166.8 µV_rms are similarly modest.
  • The filter targets barely vary. Across the thirty configurations, corner frequency spans 4.15–4.59 MHz (±5%), quality factor 0.641–0.685 (±3%), stopband rejection 15.39–17.05 dB (±5%), passband gain −4.61 to −3.63 dB, output noise 40.0–43.9 µV_rms, power 872–955 µW. These are perturbations around a single design point, not a sweep of a trade-off surface. A method that meets 30 near-identical specifications has demonstrated repeatability, not transfer across specification distributions — which is precisely the generalization axis the introduction criticizes prior work for failing (§1: "often limited when the circuit topology, PDK, or target-specification distribution changes").
  • The many-decimal target values (20.09, 4.59, 0.641, 955) suggest the configurations were sampled randomly in a narrow box rather than designed to span difficulty. The evidence that difficulty is the binding variable is in the paper: the only failures are on "the most constrained fourth-order low-pass filter in PTM."
  • No corner, mismatch or yield analysis. Stage 1 explicitly "enforces symmetry constraints for matched devices," which is only motivated by mismatch, yet mismatch is never evaluated. For a 2027 sizing paper claiming a "realistic" campaign — and comparing against a PVT-robust baseline — nominal-only results are a material gap. PTM 45 nm is additionally a predictive model rather than a silicon-calibrated PDK.

Additional concerns.

  • The classical analog-EDA literature is neither cited nor compared against, and it is where the working parts of the method come from. Stage 1 is g_m/I_D-based sizing with pre-characterized LUTs (the Jespers/Murmann methodology, standard for ~25 years) plus structural recognition of differential pairs, current mirrors, cascodes and bias branches (Graeb's sizing-rules / structural-analysis line). Stage 2 is adjoint sensitivity analysis of an MNA system (Director & Rohrer, 1969). Stage 3 is a trust-region minimax (weighted-Chebyshev) step on a log-linearized model, i.e. the DELIGHT.SPICE / OPASYN / ASTRX-OBLX family, with geometric-programming op-amp sizing (Hershenson–Boyd) as the convex relative. The reference list contains none of these; related work covers only 2024–2026 ML papers. Since Table 6 shows this classical core supplies 97 of the 100 points, the paper must position against it and must include at least one classical baseline — a plain SQP/trust-region sizer on finite-difference SPICE gradients, and a Bayesian-optimization sizer (the paper mentions AnaFlow's "Bayesian optimization safepath" but never runs BO itself). Without them, I cannot tell whether AnaFlash is faster than a 1990s sizer.
  • Two capabilities are claimed in Table 1 but never evaluated. (i) "Infeasibility Proof": §3.2 describes conservative reachability bounds ("conservative stage-gain ceilings"), which are a necessary condition, not a proof of infeasibility — and since 100%/94.7% of targets were met, the mechanism is never exercised in any reported experiment. (ii) "Explainability Validation": the LLM "synthesizes a human-readable explanation," but no explanation is assessed for faithfulness or usefulness — no expert rating, no check that the cited sensitivities match the Jacobian, no example in the paper.
  • No dispersion reported. Tables 3–4 give means over thirty configurations with no standard deviation or range; Table 5 is a single run per cell; Table 7's baseline is the median of three runs while the LLM arm appears to be a single run.
  • Venue fit. The ICLR-relevant contribution is a 7B LLM adjusting a weight vector roughly 4.7 times per run, with a measured benefit that does not separate from random perturbation and no benefit from scale. The substantive contributions — the sizing flow, the sensitivity engine, the simulation-count reduction — are circuit-design-automation contributions. With the comparison fixed, harder specifications, corner analysis and classical baselines, this would be a solid DAC / ICCAD / TCAD paper. As an ICLR paper, the learning content is not there.
Questions for authors
  1. Is the LLM better than noise? Please report (a) the three LLM sizes pooled against random adaptation with a significance test, (b) best-of-N random restarts at a simulation budget matched to the LLM arm (N = 2 and N = 3), (c) at least three independent runs of the Qwen2.5-7B arm so the 540/540 can be distinguished from a lucky draw, and (d) two trivial deterministic policies — "raise the weight of the largest unresolved violation" and round-robin priority cycling. If (b) or (d) reaches 100%, the LLM is not doing work and the paper's framing needs to change. This is the single question that decides my recommendation.
  2. Does the LLM use the information it is given? Please report whether the revised weights w correlate with the device-level sensitivity rankings in the prompt, and what happens when the structured summary is shuffled or replaced by a uniform placeholder. If performance is unchanged, say so — that result is publishable and more interesting than the current claim.
  3. Why is a 32B controller worse than a 7B one? Table 8's monotone degradation (540 → 534 → 532) is the opposite of what the "LLM contributes circuit reasoning" hypothesis predicts. What is your explanation, and what are the per-size variances?
  4. Specification difficulty. For each of the eleven topologies and three PDKs, please report the achievable range of each metric (e.g. the maximum DC gain the two-stage Miller and telescopic-cascode OTAs reach in each node) alongside the thirty targets, so the reader can see where the targets sit relative to the topology's capability. Then please re-run with targets placed near the Pareto boundary — say at 70%, 85% and 95% of achievable gain and GBW — and report success rate and simulation count. A 20–34 dB target on a telescopic cascode does not test a sizer.
  5. Specification spread. Why do the thirty filter configurations span only ±3–5% in Q and corner frequency? How were the thirty configurations generated, and what do you consider the effective number of distinct design problems in the 990?
  6. Table 5 fairness. Are the AnaCraft and AnaFlow specification targets in Table 5 identical to yours? Please tabulate them side by side. And given AnaCraft optimizes for PVT robustness, please either (a) report AnaFlash's simulation count when it must also satisfy PVT corners and Monte-Carlo mismatch, or (b) remove the AnaCraft comparison and state plainly that AnaFlash is a nominal-only sizer.
  7. Corners and mismatch. What is the yield of the 990 sized designs under the PDK's statistical models? Analog sizing that passes nominal and fails at 3σ is not usable, and symmetry constraints in Stage 1 suggest you already consider mismatch important.
  8. Classical baselines. Please add (a) SQP or trust-region minimax on finite-difference SPICE gradients from the same g_m/I_D seed — i.e. AnaFlash's Stage 3 without the adjoint model — and (b) a Bayesian-optimization sizer, both at matched simulation budget. (a) isolates whether the adjoint Jacobian is worth building; (b) is the standard black-box reference.
  9. Infeasibility and explanations. Please include experiments where targets are unreachable and report what the reachability check says and how often it is correct (false-infeasible and missed-infeasible rates). And please evaluate the generated explanations — e.g. expert ratings on a sample, plus an automatic check that the sensitivities cited in the text match the computed Jacobian.
  10. Corrections. §3.3's reference to "Equation (8)" should be Equation (7); and please remove "ThunderGen" from the Ethics Statement (line 498) — it is not this system's name and is a possible anonymity leak.
2 comments

We thank the reviewer for a very thorough review, and for noting (S1) that the division of labour is right and that the no-LLM ablation is a result worth having in the literature.

Q1/W1. Is the LLM better than noise? We ran the requested controls on the new benchmark that represented a more realistic and challenging range for the topologies (described under W3 below; 16 topology–process rows × 2 edges, 10 seeds for stochastic arms):

PolicyLow edgeHigh edgeMedian sims (met)
Analytical engine, no LLM16/1616/1624 / 25.5
Qwen2.5-7B (T=0.7)150/160153/16024 / 18
Qwen2.5-7B, shuffled prompt150/160157/16024 / 18
Random priority155/160150/16024 / 18
Round-robin14/1615/1620.5 / 18
Largest violation first15/1615/1624 / 18

This answers (c) repeated LLM runs and (d) deterministic policies. For (a), on the original set the three pooled LLM sizes meet 1,606/1,620 against random's 1,597/1,620 (p ≈ 0.14). On the hard set, both reach 95.3%. For (b), best-of-N random restarts reach 99.2%/96.1% (low/high edge) at N = 2 and 99.8%/97.1% at N = 3, at a mean of 69–125 SPICE calls. The deterministic engine reaches 32/32 at a median of 24–25.5. Shuffling the structured summary makes no detectable difference, which we report as the reviewer suggested. All policies are reported in Section 4.6 and Appendix A.6 (Table 14).

On filters, where the deterministic engine stalls, re-prioritisation helps (same tasks, one run each):

ProcessLLMNo LLMDiscordant (LLM only / no-LLM only)Sign test
SkyWater 130 nm135/150122/15016 / 3p = 0.004
PTM 45 nm137/150135/1504 / 2p = 0.69
GF 180 nm108/150108/1506 / 6p = 1.00
Pooled380/450365/45026 / 11p ≈ 0.02

The LLM is therefore useful in both of its roles: it breaks stalled trade-offs at no extra simulation cost, with a stated reason for each change, so that every decision of the flow remains interpretable for the designer, which a random policy cannot provide; and it explains the result (general response, point 2).

Table 7 and model scale. We recognise that the differences in both tables are within noise. Table 7 favours the 7B controller in some cells and matches random in others, including SkyWater 130 nm. In Table 8, the 540 → 534 → 532 differences (6 and 8 of 540 tasks) are smaller than the seed-to-seed variation of the 7B model itself: on the hard OTA set, its 10 seeds individually meet between 29 and 31 of 32 problem-edges. This is consistent with the paper's design. Numerical exploration stays in the analytical engine, so a larger model is not expected to bring a significant improvement, and it does not. Without such an improvement, the smaller model is preferable, as it requires less memory and compute per inference. We will frame both tables this way in the revision; they are now in Appendix A.8 (Tables 19–20).

W2 and corners. Table 5, PVT robustness and mismatch. We agree that AnaCraft solves PVT-robust sizing, whereas AnaFlash sizes at the nominal corner. Table 5 now marks AnaCraft as a different task and shows it as context rather than as a speed-up; Section 4.5 states that AnaFlash is a nominal-corner sizer, and Section 5 lists corner- and mismatch-aware sizing as future work. AnaFlash was run on AnaFlow's published targets (DC gain ≥ 54.4 dB, GBW ≥ 200 MHz, PM ≥ 60°, slew ≥ 30 V/µs, noise ≤ 30 nV/√Hz at 1 kHz), which are well above the main benchmark's and are now listed with Table 5. We also agree that wall-clock time mixes deployment modes, so the revised paper compares the workflows primarily in SPICE simulations, where AnaFlash needs 21–45 against 205–360 for AnaFlow (Section 4.5, Table 5).

W3. Targets and spread. The original targets were indeed modest and the filter configurations narrow. We therefore added two benchmarks that push AnaFlash to the edge of what each circuit can deliver and across its whole trade-off surface. They complement the original ones, whose results are all kept.

Original benchmarkNew benchmarks
OTA DC gain20.1–34.2 dB33.5–90 dB
OTA GBW14.9–30.7 MHz21.3–40 MHz
OTA phase margin45–57.5°60–65°
OTA slew rate13.1–26.4 V/µs22.3–40 V/µs
Filter corner frequency (PTM 45 nm)4.15–4.59 MHz1.9 kHz–19.8 MHz
Filter Q0.64–0.690.53–1.67

Hard OTA benchmark. For each topology and process, all five targets (gain, GBW, phase margin, slew rate and noise) are placed simultaneously in the top quartile of the topology's first-order operating range (the bottom quartile for noise). For example, the two-stage Miller must reach 81–85 dB and the cascoded topologies up to 90 dB in GF 180 nm, about 50 dB above the original targets. AnaFlash meets 303 of 320 runs (94.7%) and 29 of the 32 tasks in every one of ten runs, with a median of 24 SPICE simulations, the same budget as on the original, much easier benchmark. The initial seed meets only 15.6% of these runs, so the result is produced by the sizing loop. A Bayesian-optimisation sizer reaches only 76.0% of runs with 3–4× more simulations (see Classical baselines below).

Trade-off filter benchmark. Each of the 30 configurations per filter and process requests 85% of the performance of a different valid design of that filter, spread along the corner frequency. The targets therefore cover the achievable trade-off surface: corner frequencies span 1.1–4 decades per cell (e.g., 1.9 kHz–19.8 MHz for the PTM 45 nm high-pass biquad), Q spans 0.53–1.67, and no task is met at the initial seed. AnaFlash meets 91.3% (PTM 45), 90.0% (SkyWater 130) and 72.0% (GF 180) of these tasks, with three independent runs in PTM 45 nm reaching 91.8% (95% CI 88.9–94.0%).

Together, these results show that AnaFlash does not only repeat a single design point. The same engine, with no training and no per-topology or per-process change, sizes six OTA and five filter topologies in three processes from modest targets to the limits of each topology, at a budget of tens of simulations.

We included in Table 2 the definition of all four benchmarks with the original and new targets, Table 3 (Section 4.2) reports success on each of them, and the detailed results are in Appendix A.7 (Tables 15–18). The construction and per-cell targets of the new benchmarks are in Appendix A.3 (Tables 10–11).

Classical baselines. We now cite and position against gm/ID sizing, sizing rules, adjoint sensitivity, and other works that expand the state of the art. As a black-box reference, we added a GP-EI Bayesian-optimisation sizer with the same SPICE accounting. It meets 73 of 96 hard OTA runs (76.0%) at a median of 72–94 calls, against 303 of 320 runs (94.7%) at a median of 24 calls for AnaFlash. The BO plus (1+1)-ES search used to build the filter targets is a second black-box reference. A finite-difference Jacobian costs one simulation per knob per iteration, whereas AnaFlash obtains its Jacobian from the operating point plus batched device-level evaluations.

Infeasibility and explanations. Both are now evaluated (general response, point 2). Because ground-truth feasibility labels are unavailable in general, we report flagged cases and found designs (42/42 challenging cases flagged with no satisfying design found; 0/21 controls flagged) rather than false-infeasible and missed-infeasible rates, and call the output a diagnosis rather than a proof.

Dispersion. The hard OTA results use 10 seeds per stochastic arm, with per-row counts in Appendix A.7 (Table 17). On PTM 45 nm filters, three independent runs meet 413/450 (91.8%, 95% CI 88.9–94.0%), with a run-to-run variation of ±1.6 points per cell; per-cell medians and IQRs of SPICE simulations are in Table 18.

Venue fit. We believe the paper addresses a question central to ICLR: where an LLM should, and should not, sit in a scientific optimisation loop. Across two circuit families, three processes and three benchmarks, the evidence points one way. Numerical decisions belong to a verified engine; the LLM is most useful breaking stalled trade-offs with a stated reason and turning facts the engine has already checked into explanations. We report this as a design principle backed by controls, and believe it is useful for the growing body of LLM-for-EDA work at ML venues.

Corrections. The equation reference in 3.3 and the system name in the Ethics Statement are fixed.

Reviewer_LyFhOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

AnaFlash is a hybrid AI-analytical framework in which an explicit circuit model drives the exploration of the sizing space. AnaFlash combines an analytical circuit model with a lightweight LLM. It has three stages:

Initialization: the netlist is parsed into functional device groups (differential pairs, mirrors, cascodes). Currents come from first-order design relations, and widths come from gm/ID lookup tables, giving a physically valid starting point.

Model extraction: a SPICE operating point is turned into a small-signal MNA system. Adjoint sensitivities are pushed through the chain rule to a log-log design Jacobian. Topology and PDK reachability bounds are also computed.

Optimization: each step solves a weighted Chebyshev LP over log-domain knob updates inside a trust region. SPICE checks each step using a gain ratio, with backtracking. Only when a step is unproductive is a small LLM (Qwen2.5-7B) called to revise the metric priority weights. The LLM also writes a final explanation of the design.

Critical strengths

The experiments are broad, as they span two circuit families, each of them with at least five circuit topologies, three technology nodes, and thirty specification configurations.

Among their contributions, demonstrating training-free sizing and up to 56× lower runtime than prior recent SOTA workflows is notable.

Ablation shows that the analytical engine does most of the work, strengthening the paper's contribution and role of the analytical engine rather than the LLM in their hybrid framework.

Critical weaknesses

What is new compared with classical (non-AI) circuit optimization? The analytical engine is built from tools in classical circuit design automation (sensitivity analysis, gm/ID lookup tables, trust-region optimization), and the LLM adjusts priority weights. Table 1 compares AnaFlash only with ML-based methods. Which parts of the analytical engine are new, and how does AnaFlash compare with classical simulation-based sizing optimizers?

How much does the LLM actually help, and what is the justification for using an LLM in the loop? In Table 6, the analytical engine alone succeeds on 97.0% of OTA tasks, random reweighting on 98.5%, and Qwen2.5-7B on 100%. If this is the case, what is the reason the authors chose an LLM to improve the accuracy rather than use a more costly but more precise classical analytical engine? Cost-wise, is this hybrid framework more efficient?

Questions for authors

Please see weaknesses.

2 comments

We thank the reviewer for noting the breadth of the evaluation and the ablation showing that the engine does most of the work. We agree with this reading and have made it the paper's central message (general response, point 1).

W1. What is new compared with classical optimization, and how does AnaFlash compare with classical sizers?

The analytical techniques AnaFlash builds on were developed one to two decades ago. They were usually separate tools that needed expert set-up for each topology, and recent sizing work has largely moved to AI-centred methods instead. We do not claim novelty for these techniques. We now cite and position against gm/ID sizing, sizing rules and structure recognition, adjoint sensitivity and geometric programming, added in Section 2.

Our contribution is to combine them into an end-to-end, automatic and verified flow. From an unsized netlist, AnaFlash runs structure recognition, the gm/ID starting design, Jacobian extraction and the trust-region loop without per-topology equations. The same code sizes 6 OTA and 5 Gm-C filter topologies in 3 processes, and SPICE checks every step. On top of this flow, it adds a quantitative feasibility diagnosis and a design report whose facts are checked against SPICE (general response, point 2).

As a classical simulation-based baseline, we added Bayesian optimisation (GP-EI) with the same SPICE accounting, evaluated on a new hard OTA benchmark (details in our response to uXLa, W3). BO meets 73 of 96 runs (76.0%) at a median of 72–94 SPICE calls, while AnaFlash meets 303 of 320 runs (94.7%) at a median of 24 calls (Section 4.2).

W2. How much does the LLM help, why not a more precise classical engine, and is the hybrid more efficient?

The role of the LLM, and why the numerical search is left to the analytical engine, are addressed in point 1 of the general response. The benefit of LLM re-prioritisation is quantified in our response to uXLa (Q1/W1), and the explanations it produces are evaluated in point 2 of the general response.

Regarding cost, the hybrid flow is computationally inexpensive. The LLM is a 7B open-weight model run locally, so no frontier model or external API is required. On the nominal OTA benchmark it is invoked 4.7 times per run on average, consuming about 3.1k tokens per run (Appendix A.8, Table 20). It does not increase the simulation budget: the mean number of SPICE simulations per run is 25.1 with the LLM and 24.8 without it. For comparison, AnaFlow relies on frontier models (GPT-o3, Gemini-2.5-Flash) and requires 205–360 simulations on the shared benchmarks, against 21–45 for AnaFlash (Section 4.5, Table 5).

Thank you for your response and clarifications. These explanations resolve some of my concerns. While I acknowledge the authors' efforts to integrate existing analytical techniques into an automated framework, I still consider the contribution to be primarily system-level integration rather than methodological novelty. I therefore maintain my original score.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Role of the LLM

We thank the reviewers for detailed and constructive reviews. We summarise the main changes here. For the reviewers' convenience, the main changes and additions in the revised paper are highlighted in blue.

The submission reported that the analytical engine alone meets 97% of OTA tasks (Table 6), and that the LLM never sets device sizes. All three reviewers read this as the engine doing the numerical work, and two describe it as a strength of the architecture (uXLa S1, QE22). We agree, but there is more we consider about the role of the LLM:

The LLMs should not do the numerical sizing as proven by CircuitSense that reports below 19% accuracy for current models on symbolic circuit derivation. In 9 of AnaFlow's 11 LLM configurations, its external numerical optimiser uses most of the simulation budget (e.g., 250 of 282, 312 of 360). Only one configuration reaches a solution with the LLM alone. For this reason, in AnaFlash the engine makes every numerical decision. Each decision comes from a SPICE-grounded analytical engine, and SPICE checks every step.

Instead in our proposal, the LLM explains and re-prioritises when the search stalls. It turns the engine's quantitative evidence into guidance a designer can act on: which trade-off is active, which devices dominate each metric, why a step was accepted or rejected, and why a target set is unlikely to be reachable on a given topology and process. Additionally, It can change which trade-off the engine pursues next and records a reason for each change. It never touches device sizes or SPICE results.

We propose this as a white-box sizing flow: Every numerical fact comes from SPICE or the SPICE-extracted Jacobian and can be checked. The LLM translates those facts into the designer's language.

The controls requested by uXLa confirm this division of labour. Where the engine does not stall, all re-prioritisation policies perform alike; where it stalls between conflicting specifications, LLM re-prioritisation significantly improves success (380 vs. 365 of 450 tasks on a new filter benchmark, p ≈ 0.02) at no extra simulation cost, with a stated reason for each change. Details are in our response to uXLa.

AuthorsOpenReview ↗

Feasibility diagnosis and SPICE-checked explanations

All reviewers asked us to evaluate the "Infeasibility" and "Explainability" entries of Table 1, the capabilities a white-box flow is meant to provide. We now evaluate both.

Feasibility diagnosis: By feasibility we mean checking, before any optimisation budget is spent, whether the requested specifications can be reached with the given topology in the given process. AnaFlash bounds what the circuit can deliver from the PDK device characteristics, e.g., the transconductance efficiency available at the allowed current. When a target set exceeds these bounds, it reports which specifications conflict and the physical limit responsible. The designer can then relax the targets or change the topology instead of waiting for the optimiser to fail.

We tested 63 cases: 7 OTA topologies × 3 PDKs × 3 target sets. Two sets are challenging (bandwidth vs. power, slew vs. power), one is a feasible control. All 42 challenging cases were flagged before optimisation, each with a quantitative reason. For the PTM 45 nm 5T OTA, for example, the checker estimates a GBW ceiling of 18.8 MHz against the 150 MHz requested at 5 µW, and a minimum power of 39.8 µW for that GBW. No search (the agent, a 100-call random search per case, or archived designs) found a design meeting any flagged target set. None of the 21 controls was flagged, and a satisfying design was found for 17 of them. Since ground truth is not available in general, we present this as a model-based diagnosis, not a proof.

We have updated the paper accordingly to highlight these details, providing the complete description and results in Section 3.2, Section 4.3 and Appendix A.4

Explainability: By explainability we mean a readable account of both the sizing process and the resulting design. It describes which specifications were binding, which knobs and devices determine each metric, why each step was taken or rejected, and which trade-offs the final design sits on. In AnaFlash this account is derived from SPICE rather than from the LLM's own reasoning. Every number in the design report is computed by code from SPICE and the SPICE-extracted Jacobian, and the LLM turns these facts into prose for the designer.

To check that these facts are correct, we compared the engine's sensitivity claims with independent two-sided SPICE finite differences: they agree in sign in 569 of 573 conclusive cases (99.3%). An optional per-run check re-verifies them at about 10 extra simulations. By contrast, directional statements written freely by the LLM agree in 92 of 126 cases. This is why the facts are computed by the engine and the LLM's interpretation is labelled as such. An excerpt of the code-generated part of a report (PTM 45 nm 5T OTA):

result: ALL SPECIFICATIONS MET after 14 simulations (8 for the seed)
metric     start        final        target
A0_db      37.04 dB     38.65 dB     >= 30 dB
gbw        50.81 MHz    55.99 MHz    >= 30 MHz
slew       29.78 V/us   33.73 V/us   >= 30 V/us
power      33.43 uW     37.75 uW     <= 1 mW
WHAT MOVES EACH METRIC (elasticity d ln m / d ln k)
A0_db  increases with g2.L       +5.56
gbw    increases with vb[vbias]  +10
step 1: priority=[slew], accepted, rho=0.804
  merit improvement: predicted 0.155, measured 0.125
Jacobian claims checked by SPICE finite differences:
  15 confirmed, 0 wrong sign

We added the complete design report in Section 3.4, and the audit and this excerpt are in Section 4.4 (Figure 3). The complete report is in Appendix A.5.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.