acceptodds
Under review as a conference paper at ICLR 2027

Towards Expert-Level Analog Layout Design: Large Language Model as Optimizer

Abstract

Analog physical design requires coordinated decisions about device placement, routing, and the parasitics that determine circuit performance. Yet much of its automation optimizes individual stages, leaving engineers to reconcile downstream electrical failures with earlier layout choices. We introduce LAMBDA, a large language model based analog physical design framework that combines global reasoning with fast automated execution. To connect reasoning with execution, a customized B Graph unifies analog constraints, placement, and routing in an executable search space, enabling a multimodal LLM to optimize coupled design decisions using circuit knowledge, layout observations, and accumulated experience. Multi-fidelity optimization couples low-cost physical and parasitic models with detailed electrical evaluation feedback, balancing exploration efficiency with post-layout performance. On nine commercial 28-nm circuits, LAMBDA improves geometric mean full layout area and wirelength and the composite analog electrical score over SOTA baselines.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Analog physical design, Graph Representation, LLM, Joint placement and routing, Multimodal reasoning, Multi-fidelity optimization

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_HCSBOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work presents LAMBDA, a large language model based analog physical design framework that uses B* graph connects circuit-derived constraints, placement and routing in same search space for joint Place-and-Routing. Also, multi-fidelity optimization couples low-cost physical and parasitic models with evaluation feedback to balance exploration efficiency with post-layout performance. It improves performance on experiments with commercial 28-nm circuits, evaluated by geometric mean full layout area and wirelength and the composite analog electrical score.

Critical strengths
  • It focuses on a practical and valuable problem, which pushes the boundary of LLMs/agents.
  • The proposed framework makes sense to me. Each module within the framework is introduced clearly with explanation for what it is and why it should be added to the framework.
  • The overall writing is easy to follow. I'm not an expert for circuit design. While it is still difficult for me to understand some circuit-related terms, I can easily follow how LLM is used to solve the problem.
Critical weaknesses
  • Is there any performance gap between different LLMs/MLLMs, especially open-sourced models vs. close-sourced models and different family of LLMs/MLLMs (e.g., GPT, Claude, Qwen, Kimi, etc.)?
  • How many test cases are used in the evaluation? Does nine commercial circuits mean nine test cases? If that is true, the evaluation results may not be that reliable since the number of test cases used is too small.
  • It would be better if there is a separate Related Work section to introduce how LLM is used in circuit design more comprehensively.
  • Actually, the proposed framework is a LLM harness for circuit design. Is it possible to rollout data under this harness and then internalize this capability into LLMs by SFT or RL? I know it is not the core idea of this work, but such discussion is valuable for judging the future impact and potential of this work.
Questions for authors

See weaknesses.

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

This paper presents LAMBDA, an LLM-based framework for analog physical design that uses a new graph representation to jointly optimize placement, routing, and design constraints. The framework integrates multi-fidelity evaluation, including post-layout simulation, to guide iterative optimization. Experiments on nine circuits demonstrate improvements in layout area, wirelength, and electrical performance over existing automated baselines.

Critical strengths
  1. The proposed framework thoughtfully integrates multiple complementary components, including domain-specific knowledge and tools, multimodal feedback through layout images, and multi-fidelity optimization with post-layout simulation feedback. It effectively leverages the LLM's capabilities to coordinate these components and iteratively refine designs based on both physical and electrical feedback.

  2. The experimental evaluation goes beyond geometric metrics by incorporating design rule checking (DRC), layout-versus-schematic verification (LVS), parasitic extraction (PEX), and post-layout circuit simulation. This is particularly important for analog physical design, where layout quality cannot be adequately assessed through area and wirelength alone, and strengthens the practical relevance of the reported results.

Critical weaknesses
  1. Although the proposed framework is comprehensive, it remains unclear how much of its performance comes from the LLM's capabilities versus the domain-specific knowledge and reusable skills supplied by the framework. While the ablations demonstrate the importance of LLM-guided optimization and the overall framework, they do not isolate the contribution of the supplied domain knowledge and skills while keeping the optimization framework fixed. This distinction is important for understanding whether the framework can generalize to new circuit topologies or related design domains without substantial additional knowledge engineering. It is also important for assessing whether similar performance could be achieved with smaller, more accessible models supplied with appropriate domain knowledge, rather than relying on expensive proprietary LLMs.

  2. The experimental evidence does not fully support the breadth of the paper's claims, particularly regarding generalizability and expert-level performance. The evaluation is limited to nine fixed circuit instances in a single technology process, leaving uncertainty about performance on different circuit configurations and design requirements. Moreover, although the expert and LAMBDA layouts are evaluated using the same testbenches, the expert comparisons lack essential methodological details, including who developed the reference layouts, what objectives they optimized, and how much design effort was involved. Without establishing comparable optimization objectives and conditions, favorable results relative to these reference layouts do not necessarily demonstrate expert-level design performance.

Questions for authors
  1. Could the authors provide a concrete example of an LLM optimization trajectory, including the actual prompts and inputs provided to the model (circuit information, domain knowledge, and layout images), the model's responses, the resulting B* Graph modifications, and the physical and electrical feedback used in subsequent iterations? Although the paper describes the overall workflow and provides pseudocode, a representative end-to-end example in the main paper or appendix would make the LLM's role in the optimization process much clearer.

  2. Could the authors clarify the specific domain knowledge and reusable skills supplied to the LLM and how they are incorporated into its decisions? Although the paper describes these resources and includes relevant ablations, the contribution of the supplied knowledge is not isolated while keeping the optimization framework fixed. How much additional knowledge engineering would be needed to adapt LAMBDA to new circuit topologies or related design domains? Could a smaller or open-weight model achieve comparable performance with the same framework and supplied knowledge?

  3. The evaluation covers nine fixed circuit instances in a single technology process. How were these instances selected, and were they held out throughout framework development, including prompt engineering and workflow tuning? The paper states that the circuits were unseen tasks and absent from the pre-existing knowledge repositories, but it is not clear whether they were also excluded from development and tuning. What evidence supports generalization to other circuit configurations or technology processes?

  4. Could the authors clarify the provenance of the expert-designed layouts in Appendix A, including the designers' expertise, the objectives they optimized, the design effort involved, and their stopping criteria? Although the expert and LAMBDA layouts are evaluated using the same testbenches, it is unclear whether they were developed under comparable optimization objectives and conditions. This information is important for assessing the paper's expert-level claims.

  5. As a minor presentation suggestion, the layout visualizations in Figures 5 and 6 are difficult to inspect at their current size. Larger or higher-resolution versions, potentially with zoomed-in views, would make the layouts easier to examine.

Reviewer_Y1cROpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper presents LAMBDA, an LLM-guided framework that jointly optimizes analog placement, routing, and layout constraints through a B* Graph representation. It combines physical evaluation and simulation feedback. Experiments on nine commercial 28-nm circuits report improvements over five automated baselines.

Critical strengths
  1. The paper presents an interesting idea of using the multi-fidelity loop and B* graph to encode placement and routing information.
  2. The evaluation includes complete routed layouts using physical verification and post-layout simulation.
Critical weaknesses
  1. The novelty of the work should be clarified against previous works, including the difference between enhanced B*-tree and the proof for admissible property
  2. The experimental settings need more details such as the SA configuration. Some metrics are missing in the evaluation.
Questions for authors
  1. The enhanced B*-tree proposed in Ou et al. (2013) already jointly model placement, routing spaces, routing topology, and symmetry. Please clarify what B* Graph adds beyond the enhanced B*-tree and isolate its contribution from the LLM optimizer.
  2. Define “admissible B* Graph” and explain how decoding and graph edits preserve nonoverlap, symmetry, and mandatory constraints. Is an ASF-B*-tree or an equivalent mechanism used? The free-space partition proof does not establish these placement guarantees.
  3. In the evaluation, IF-amp gain alone cannot establish overall improvement without bandwidth and other relevant tradeoffs. Please report frequency response and application-specific specifications.
  4. For the experiment settings, please provide the per-circuit human-provided constraints, objective weights, annealing parameters, and search budgets for both SA configurations. Human constraint choices can substantially affect results; explain their selection and assess sensitivity to alternative reasonable sets and tuning. Please also show baseline layouts alongside LAMBDA.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.