acceptodds
Withdrawn from ICLR 2027

Never Step Back: Monotonic LLM Optimization for Circuit Layout Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Recent advances in large language models (LLMs) suggest strong potential for automating analog layout design. Yet existing iterative approaches suffer from non-monotonic refinement, where unsuccessful constraint updates may degrade previously verified layouts and propagate errors across subsequent iterations. Moreover, limited reuse of verified design experience across tasks leads to repeated failures and ineffective design exploration. To address these challenges, we propose **MONA**, a training-free multi-agent framework for monotonic analog layout optimization. MONA integrates Adaptive Layout Search (ALS) with Verification-Aware Memory (VAM) and an open-source EDA flow. ALS explores complementary constraint strategies to enable effective design exploration, while VAM employs a Verification Gate to preserve the best deliverable layout and Long-term Memory to distill verified experience for cross-task transfer, accelerating convergence without compromising previously achieved design quality. Across 6 SKY130 analog layout tasks, MONA achieves a deliverable-layout success rate, a higher success rate than the strongest evaluated LLM-based method, while attaining the highest mean FoM across all tasks. Moreover, MONA reduces LLM tokens per deliverable layout by , demonstrating reliable and resource-efficient analog layout optimization.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Large Language Model, Agentic AI, Analog Layout Automation, Electronic Design Automation

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_Br2kOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces MONA, a multi‑agent framework for analog layout refinement. MONA employs three specialized agents responsible for topology analysis, failure analysis, and candidate generation using Exploit, Diversify, and Explore strategies. A Verification Gate systematically accepts layout candidates that improve upon the best retained deliverable design. Both successful and failed evaluation trajectories are fed into a memory distiller to distill reusable design experience. Evaluated on six SKY130 analog layout tasks using an open‑source EDA flow, MONA achieves substantially higher deliverable‑layout success rates and figure‑of‑merit (FoM) compared with ALIGN, BO, and ICPI, while significantly reducing LLM token consumption.

Critical strengths

This work adopts a greedy-flavored monotonic acceptance strategy for the high cost analog layout design problem, which yields competitive performance under a limited evaluation budget. As demonstrated by the reported experimental results, MONA achieves strong empirical performance, and it may provide useful insights for future EDA-automation research.

Critical weaknesses
  1. The paper exhibits limited conceptual novelty and largely resembles a solid system-level engineering assembly tailored for the EDA setting. Many building blocks, including complementary search strategies, multi-agent collaboration, and short-/long-term memory management, are well-known techniques from recent literature. While the monotonic optimization loop realized by the Verification Gate serves as its core innovation, substantial portions of the implementation amount to careful system-engineering adaptations for analog layout design. It remains hard to disentangle how much performance improvement comes from the newly-proposed monotonic paradigm versus existing agent-based components.

  2. The experimental evaluation has several notable drawbacks. (1) Only six small-scale analog-circuit benchmarks are tested, which is insufficient to convincingly demonstrate the generalizability of the proposed method across broader analog-circuit families. (2) The fairness of token cost comparison between MONA and ICPI is not well justified. By design, MONA invokes multiple agents and generates three candidates per iteration. Even though token consumption per deliverable layout is reported, it is unclear why MONA achieves lower overall token overhead despite more LLM invocations. (3) Ablation studies are only conducted on top-level modules ALS and VAM. Many fine-grained sub-components underneath these modules remain untested, making it impossible to quantify the individual contribution of each sub-module.

Questions for authors
  1. Could you clarify whether there exist realistic circuit scenarios where high‑quality final layouts can only be reached by iterating from intermediate low‑quality states? If such cases exist, how does MONA mitigate potential performance loss? This question concerns how the monotonic greedy‑style strategy escapes local optima.

  2. The benchmark only contains six circuits. Are there fundamental barriers preventing evaluation on more open source analog layout tasks? What key factors make these six benchmarks representative enough to validate your monotonic layout optimization paradigm?

Reviewer_w14XOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MONA, a multi-agent LLM loop for analog layout constraint refinement. MONA combines adaptive layout search, which generates candidate constraints through exploitation, diversification, and exploration, with verification-aware memory, which preserves the best verified layout and reuses design experience across tasks. Experiments on 6 SKY130 PDK analog layout tasks demonstrate improvements in deliverable-layout success rates and LLM token efficiency over baselines.

Critical strengths
  • The paper provides a relatively comprehensive experimental evaluation using an open-source EDA design flow (SKY130PDK, Align, etc...), with detailed implementation settings and experimental trajectories. This makes the results more transparent and potentially reproducible. The transparent appendix (e.g. Table 14) is genuinely useful.
  • MONA achieves a higher reported success rate than the baselines, while reducing LLM token consumption. The component ablations also provide useful insights into the contributions of the proposed framework.
Critical weaknesses
  • The biggest problem is that the title “Never Step Back” doesn’t distinguish between “reporting the all-time best” and “continuing the search from the all-time best.” In its experiments, the paper compares their method that performs both tasks to a baseline that does neither, because the baseline is scored based on the final state. Consequently, the huge discrepancy in results obscures the effects of the scoring method and search quality, making it impossible to attribute the performance differences. In particular, Appendix D.2 shows that 14/15 runs found a deliverable layout on Task 4, but only 1/15 ended with one. Furthermore, the initial ALIGN layouts are already deliverable on Tasks 1, 3, and 6, yet the ICPI baseline reports success rates below 100% on these tasks. Under best-so-far evaluation, ICPI's success rate should be much higher, reducing MONA's reported 2.02× advantage much lower. This discrepancy is concerning, as it makes the readers difficult to attribute the performance gains to better optimization rather than just preserving candidates that were successful.

  • The Verification Gate essentially implements best-so-far preservation, and this is a standard optimization strategy, not novel.

Questions for authors
  1. Could the authors report success rate and FoM for all methods (baselines) under best-so-far evaluation, both with and without counting the initial ALIGN layout? This would clarify how much of the reported improvement comes from better search rather than solution preservation. (W1)

  2. Table 3 don't show whether always searching from the historically best layout is the most effective strategy. Based on my experience with analog layout, better solutions may require temporarily accepting worse intermediate results (which holds in general). Could the authors compare against ICPI+Gate, Direct+Gate, and non-greedy alternatives under best-so-far scoring?

  3. Removing only the Verification Gate results in worse performance (47.8% SR) than removing the entire VAM component (61.1% SR). Could explain why? As some tasks share identical netlists and sizing, evaluation on more distinct circuit designs would also be helpful to strengthen the cross-task transferability claim.

  4. Could the authors clarify whether interconnect resistance and coupling capacitance are included in extraction? The tight UGBW margins on Task 2,4 may be sensitive to the parasitics. Also, reporting area and electrical scores separately would also be helpful to analyze the FoM improvements.

  5. MONA uses 3 evaluations per iteration, whereas BO uses only 1 according to the Appendix. Could the authors provide comparisons under matched evaluation budgets?

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

This paper presents MONA, a training-free multi-agent framework for monotonic analog circuit layout optimization under a limited EDA evaluation budget. The framework combines Adaptive Layout Search, which uses specialized agents to analyze circuit topology and failures and propose complementary constraint candidates, with Verification-Aware Memory, which preserves the best verified layout and transfers distilled design experience across tasks. Candidate layouts are evaluated using an open-source EDA flow covering physical verification, parasitic extraction, and post-layout simulation, and only verified improvements are retained. Experiments on six SKY130 analog layout tasks show that MONA improves deliverable-layout success rate and layout quality while reducing LLM token consumption compared with the evaluated baselines.

Critical strengths
  1. The paper addresses a practically relevant problem in analog layout automation.
  2. The paper presents a relatively comprehensive iterative refinement pipeline tailored to this specific application and demonstrates some empirical improvements.
Critical weaknesses
  1. I find the novelty of the framework somewhat limited. It largely applies standard agent-based techniques to an EDA task without introducing substantial task-specific methodological innovations; the main EDA-specific component appears to be the task-dependent scoring and verification mechanism.
  2. The proposed monotonic optimization strategy may discard useful information from unsuccessful intermediate designs. Although inferior candidates can reveal failure modes and provide informative directions for subsequent correction, the framework prevents such candidates from becoming the starting point for future iterations.
  3. Some methodological details are described too vaguely. For example, the paper does not provide a concrete account of how the short-term and long-term memory components are constructed, updated, maintained, and retrieved.
Questions for authors

See weaknesses.

Reviewer_9S1fOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

MONA combines multi-agent constraint generation, verification gating, and cross-task memory for iterative analog circuit layout optimization, ensuring a nondecreasing retained score by preserving the best feasible layout found so far. The system achieves high success rates on six SKY130 tasks, but its methodological novelty, the necessity of its agent architecture, and its advantages under limited verification conditions and unequal search budgets remain insufficiently established.

Critical strengths

Integrating DRC, LVS, post-layout simulation, and best-so-far layout retention into an open EDA flow grounds the generated results in actual tool-based verification and offers practical engineering value.

Critical weaknesses
  1. “Monotonic optimization” largely amounts to best-so-far retention, overstating the methodological contribution. Refusing to overwrite a better result makes the retained score nondecreasing by construction, but does not establish a more effective search strategy or provide convergence or global optimality guarantees.
  2. Neither the novelty nor the necessity of the agent framework is adequately established. Topology analysis, failure diagnosis, candidate planning, and memory summarization primarily decompose a familiar workflow into roles without introducing a new coordination mechanism; retaining an 88.9% success rate after replacing the entire ALS module with random legal actions further highlights the need to establish the independent value of this complexity.
  3. “Adaptive search” lacks an explicit adaptive allocation mechanism. The method generates a fixed set of Exploit, Diversify, and Explore candidates in every iteration without adjusting their budgets based on gains, failure rates, or evaluation costs, making it closer to a fixed mixture of sampling strategies.
  4. The analog circuit optimization objectives are overly simplified. OTA stability, noise, offset, and dynamic response are not systematically evaluated, while the comparator is tested only under fixed input and clock conditions; improvements in scored metrics may therefore accompany degradation in important unmeasured properties.
  5. The definition of “deliverable” falls well short of practical electrical reliability requirements. Verification covers only the typical process corner, a fixed supply voltage, and capacitance-only parasitic extraction, without adequately testing interconnect resistance, PVT variations, or mismatch; the high success rate demonstrates compliance with limited test conditions rather than robust electrical performance.
  6. The main experiments do not match actual search budgets, weakening the fairness of key comparisons. MONA evaluates fifteen candidates over five iterations, whereas BO evaluates only five, and BO’s parameterization does not cover MONA’s full constraint vocabulary; although Figure 5 includes ICPI with the same verification gate, the main table omits a comprehensive success-rate and cost comparison with this essential baseline.
Questions for authors
  1. With identical best-so-far retention rules, search spaces, and EDA call budgets, how much independent benefit does the multi-agent framework provide?
  2. Given that each iteration generates three fixed candidate types, what search mechanism actually adapts to feedback?
  3. How much of the reported 98.9% deliverable-layout success rate remains under full RC extraction, PVT and mismatch analysis, and evaluation of additional critical electrical metrics?

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

Related papers

Discussion (0)

Sign in to comment.