AgenticSizing: A Large Language Model-based Multi-Agent Framework for Analog Circuit Sizing
Yijia Hao, Pratibha Verma, Dongxu Guo, Cristian Sestito, Michael O'Boyle, Christos-Savvas Bouganis, Themis Prodromakis
cs.AI · cs.AR
Quick verdict
Claims an LLM multi-agent sizing framework succeeds on a 55-transistor LDO where classical optimizers fail; promising, but n=5 evidence and inconsistent numbers weaken the headline claims.
Summary
AgenticSizing is a multi-agent LLM framework for analog transistor sizing. An LLM decomposer partitions a netlist into functional roles and canonical substructures; a planner coordinates role-scoped sizer agents that propose bounded variable updates evaluated in closed loop with Cadence Spectre; a post-run knowledge extractor distills trade-offs, role-performance, parameter-performance facts, and reusable heuristics. Against differential evolution, Bayesian optimization, and multi-agent RL across eight circuits (four seeds, four evaluation), it reaches 60% success (3/5 runs) on a 55-transistor, 60-variable integrated LDO where all baselines fail within 6 hours. On the standalone BGR it succeeds in 4/5 runs using ~23 simulations on average versus 936 (DE) and 208 (BO); the two-stage FC gives 80% success while all baselines fail. Ablations credit topology tagging (40% alone), agent specialization, and knowledge injection (60% combined). Limitations: five runs per configuration with no statistical testing; the abstract reports 83 average iterations where the tables and introduction support 89.7; Table IX timing data is internally inconsistent; the GPT backbone version differs between methods and acknowledgments; no process design kit is specified, so reproduction requires Cadence Spectre and a commercial LLM.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 3
● low 1
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A multi-agent LLM framework combining topology decomposition, role-specialized sizing agents, and an extracted knowledge base achieves feasible designs on a 55-transistor, 60-variable integrated LDO in 3/5 runs within six hours, where differential evolution, Bayesian optimization, and multi-agent RL all fail.
- AgenticSizing reaches feasibility with one to two orders of magnitude fewer circuit simulations than classical baselines (e.g., BGR: avg 22.75 simulations vs 936 for DE and 208 for BO).
- Topology understanding, design-knowledge infusion, and agent specialization provide complementary benefits: in the Table X ablation, topology tagging alone yields 40% success, agent specialization without knowledge collapses to 0%, and the full configuration reaches 60%.
- The framework scales beyond prior LLM-based sizing methods, whose largest reported circuits contain at most 20 transistors (Table I).
Strengths
Genuine scalability advance over prior LLM-based sizing work
Table III/Table I: prior LLM-enabled frameworks report at most 20 transistors (EEsizer) and at most 20 variables (LEDRO); this paper evaluates up to 55 transistors and 60 variables on an integrated LDO (Section IV-A), a category (LDO with BGR + FC) no prior LLM framework reports.
Honest, non-trivial ablation study
Table X (integrated LDO only): multi-agent specialization WITHOUT knowledge drops success from 40% (S1) to 0% (S2) — an unflattering result the authors report and interpret as coupled-subblock coordination failure; token costs (3-4M per run) and Fig. 7 limit-cycle oscillations in failed runs are also candidly acknowledged.
Large sample-efficiency gains where they matter most
Strong reproducibility effort relative to peers
Weaknesses
All headline effects rest on n=5 with no statistical treatment
Internal reporting inconsistencies
Knowledge-seeded asymmetry versus cold-start baselines
Critical implementation details missing; full reproduction blocked
Core value proposition contested by own data
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
AgenticSizing (arXiv 2609.25873) is a 1+N planner-sizer multi-agent LLM framework for analog transistor sizing, built on LangGraph with GPT-5.x. Its four phases (Fig. 1): a two-step LLM decomposer (functional-role partition, then substructure tagging against a fixed ontology, Fig. 2); closed-loop iterative sizing where a planner dispatches one role-scoped sizer per violated metric and candidates are evaluated in Cadence Spectre (Figs. 3-4); post-run knowledge extraction into four views (Fig. 5); and knowledge reuse, seeded from four low-complexity textbook circuits (Table III). The evaluation spans four medium/high-complexity circuits, culminating in a self-constructed 55-transistor, 60-variable integrated LDO. Headline claims: 3/5 success where DE, BO, and MA-RL all fail within 6 hours (Table VIII); 1-2 orders of magnitude fewer simulations (Table IX); and complementary component effects (Table X).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is AgenticSizing?
It is a multi-agent framework built on large language models (GPT-5.x) that automates analog transistor sizing: an LLM decomposes the circuit netlist into functional blocks and substructures, a planner coordinates role-specialized sizing agents, and candidate designs are checked by SPICE simulation (Cadence Spectre) in a closed loop. A knowledge extractor turns finished runs into reusable design knowledge.
How well does it perform compared to classical optimizers?
On the paper's hardest benchmark — a 55-transistor, 60-variable integrated LDO — it found feasible designs in 3 of 5 runs (60%) within a six-hour budget, averaging about 90 simulations (the abstract says 83, but the tables support 89.7), while differential evolution, Bayesian optimization, and multi-agent RL found none. On simpler circuits like the bandgap reference it used ~23 simulations on average versus hundreds for the baselines, though Bayesian optimization matched its success rate on the standalone LDO.
Which components actually matter in the framework?
An ablation on the integrated LDO found topology tagging alone gives a 40% success rate, adding agent specialization without a knowledge base collapses to 0%, and the full system with knowledge reaches 60% — so the components are complementary, with knowledge injection only effective when combined with role-specialized agents.
Is the code and data available; can I reproduce it?
The agent code is open-sourced (github.com/aprilaihub/agentic-analog-sizing) and the prompts are published in the paper. However, the results require a commercial Cadence Spectre license, a process design kit that the paper does not name, and paid OpenAI API calls of roughly 3-4 million tokens per run, so full independent reproduction is costly.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures are raster-heavy in the extracted text; circuit schematics (Fig. 6) and optimization trajectories (Fig. 7) were assessed only via their captions and tabular data.
- The technology node/PDK is absent from the extracted text; it may appear in an appendix or supplementary material not included in the extraction.
- Table IX's column semantics could not be fully reconstructed from the text extraction, which contributes to (but does not fully explain) the timing inconsistencies flagged.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.