acceptodds
Withdrawn from ICLR 2027

SAAC: Sizing Agent for Analog Circuit via Turn-Level Reinforcement Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Analog circuit sizing (ACS) is a critical yet challenging iterative optimization task in electronic design automation (EDA), requiring domain expertise, multiple expensive simulations, and the balancing of conflicting performance objectives. Existing optimization approaches face distinct limitations: traditional black-box methods like Bayesian Optimization ignore the reasoning process and prior knowledge, relying instead on extensive sampling that scales poorly with design complexity. While large language models (LLMs) offer promising reasoning capabilities, current LLM-based ACS workflows either depend heavily on the base model's inherent abilities or require costly human-annotated data for training, and existing RL-based methods also fail to perform fine-grained, turn-level optimization at each iteration. To address these gaps, we propose SAAC (Sizing Agent for Analog Circuit), an agentic framework specifically tailored for analog circuit sizing. SAAC integrates a professional-grade simulation environment that provides standardized tool interactions and verifiable reward signals, coupled with an automated data synthesis pipeline that eliminates the need for expert-annotated training data. To optimize the agent's iterative refinement capability, we also introduce Turn-Level GRPO (TL-GRPO), a lightweight, tool-integrated reinforcement learning algorithm that performs turn-level group sampling for fine-grained policy updates. Evaluated on 12 professional analog circuit tasks, SAAC trained with TL-GRPO achieves the best overall performance among the evaluated methods, outperforming Bayesian Optimization and other RL baselines. Moreover, SAAC demonstrates generalization to unseen circuit specifications, providing a practical foundation for LLM-driven automation in electronic design.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM Agent, Analog Circuit Sizing, Reinforcement Learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_zur6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces SAAC, which is based on tuning a pre-trained Qwen model for sizing analog circuits via iterative simulator-based feedback. The core of SAAC is Turn-Level GRPO (TL-GRPO), which evaluates candidate actions sampled from the same interaction sequence based on their reward to approximate the overall objective of finding the best candidate encountered along a trajectory. Experimental results show an improvement in the max-over-turns score compared to other methods tested.

Critical strengths

S1. Simulator-based sizing.
Candidate designs receive simulation feedback. Moreover, the evaluation includes held-out tasks. This gives the work relevance beyond a conceptual optimization setup.

S2. The paper compares several RL training variants.
The comparisons among trajectory-level GRPO, single-turn GRPO, Kevin, and TL-GRPO provide useful evidence about the proposed training choice. TL-GRPO reports the highest average score among these variants on the reported in-domain and held-out sets (Tables 1 and 2).

Critical weaknesses

W1. The primary training signal is not demonstrated to maximize the objective function that the paper focuses on.
The objective is to determine the best design along a trajectory, but the TL-GRPO algorithm is trained on the rewards of actions taken from the same history. As mentioned by the authors (Section 3.1.3 "this surrogate does not capture the delayed informational value of an action" and Appendix I "Turn-level reward design" paragraph), this proxy can fail to capture the informative value of the action in the long term. It is not clear when the turn-level signal reliably improves the trajectory-level objective, thus limiting the support for its central claim.

W2. The evaluation fails to demonstrate that the improvements are robust to settings other than the chosen one.
Training queries are based on parameter distributions and offsets generated synthetically (Section 3.3). Using three evaluation trials and selecting the best checkpoint does not provide clear information about variations across the trials or the effects of selecting the best checkpoint. More evidence is required before making strong claims about the capability of sizing.

Questions for authors

Q1. Clarification. TL-GRPO vs. GRPO.
What is being held constant when comparing TL-GRPO to GRPO at the trajectory level? Provide separate numbers for model generation and simulation evaluation, and explain how the original trajectory action and reward are used in the turn-level groups. Also, do the baselines have an equivalent amount of budget? If not, then how much of the reported gain holds up with equivalent budgets?

Q2. Clarification. Complete-Trajectory GRPO advantage.
While the comparison provided in the studies involves mean max-over-turn values, the issue remains unresolved in whether the advantage of TL-GRPO is due to improved exploration capabilities or selection of actions that yield greater immediate rewards. Are there any comparisons of best-so-far rewards and candidate-space exploration compared to trajectory-level GRPO for equal budgets? If not, how should the claim be restricted?

Q3. Future Work.
Gains sensitivity. What is the sensitivity of the gains to the synthesized training-query distribution? Indicate the per-task parameter ranges and target-offset distributions. If there have not been any tests on other ranges or target distributions, then how is generalization assured for this particular query generation?

Reviewer_J12YOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces SAAC, an agentic framework tailored for reasoning-guided analog circuit sizing. By integrating a professional simulation environment, an automated data synthesis pipeline, and Turn-Level GRPO (TL-GRPO) as its core optimization engine, SAAC enables scalable, annotation-free RL training.

Critical strengths

The paper combines simulation feedback with RL-based LLM post-training to develop a method that requires no manual annotation and achieves better results than larger models on the tasks used for training.

Critical weaknesses

The experimental evaluation is insufficient to support the claims of PDK generalization, practical efficiency, and improved reasoning capability.

Questions for authors

The PDK configuration for the experiments is unclear. Appendix B does not provide a complete set of task-to-PDK pairs. Moreover, the TSMC 0.18 um RF appears in both the training task OTA and the held-out OPAMP task; this means the OOD evaluation does not exclusively involve unseen PDKs. Please provide a per-task table specifying the PDK and training & testing assignment. Also distinguish results on seen vs. unseen PDKs to support the claimed generalization.

Please report the all-specification success rate for every method on both in-domain and OOD tasks in Tables 1 and 2, extending the results currently provided in Appendix E. A successful case should require at least one design within the five-turn budget to satisfy all required specifications simultaneously. The average scores in Tables 1 and 2 alone do not establish the proportion of sizing tasks successfully completed.

The authors also limit BO to five turns per task. However, considering the time required for LLM inference, restricting both methods to five turns does not provide a fair comparison of runtime efficiency. Please compare TL-GRPO and BO under the same wall-clock time budget for each task.

Moreover, the authors do not clarify whether the Analyze component helps the LLM generate high-quality results. Ablation studies with and without the Analyze component, and with and without thinking mode, should be considered.

Reviewer_3oVhOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SAAC, an LLM-agent framework for analog circuit sizing. It has three parts: (i) a Cadence-based, distributed master-worker simulation environment that returns verifiable rewards; (ii) a data-synthesis pipeline that randomizes initial parameter values and perturbs target specs to produce 10K queries without human annotation; and (iii) Turn-Level GRPO (TL-GRPO), an RL algorithm for iterative-optimization tasks.

The authors model ACS as a single-state POMDP whose trajectory value is the maximum per-turn reward. They argue that trajectory-level GRPO assigns a uniform advantage to every turn and so gives poor credit assignment. TL-GRPO samples one full trajectory, splits it into history prefixes h_t, and samples G−1 extra single-turn actions per prefix. Advantages are normalized within each same-history group using a unified geometric-mean spec-satisfaction reward. The authors argue this costs the same G·T rollouts as vanilla GRPO.

Critical strengths

A well-motivated problem structure and a cheap, practical algorithm. Observing that analog sizing has a fixed evaluator state and a deterministic, action-only reward is a useful framing. It makes same-history, turn-level group comparison possible without extra rollouts beyond GT.

A realistic, large-scale evaluation setting. The work uses industrial PDKs, netlists with up to 38 tunable parameters and 12K-token prompts, and a real Spectre simulator. In-domain gains over the RL baselines are large (0.97 vs. 0.85/0.83/0.55), and the controlled RL comparison (same base model, G, T, batch size) is a good design.

Critical weaknesses
  1. The generalization claims are stronger than the evidence. The four held-out tasks (OPAMP1–4) have similar netlists and differ mainly by process node, so they test little topology diversity, yet the Conclusion claims generalization to “unseen circuit topologies.” The OOD advantage comes largely from OPAMP4 (0.58). Without it TL-GRPO leads only narrowly (about 0.28 vs 0.27), and it trails BO on OPAMP1. The benefit of history-conditioning is also not isolated, since there is no ablation and per-turn scores plateau after turn 1.
  2. The experimental reporting is incomplete. There are no training-seed variances, the checkpoint selection criterion is unspecified, and no baseline tuning is reported. Kevin’s training also appears unstable. BO is given only 5 simulations, which is a very low-budget regime, so the BO comparison says little about practical sample efficiency.
Questions for authors
  1. How was the “best checkpoint” selected for each method (which data, which metric)? Are the OOD tasks completely excluded from selection?
  2. Under an equal simulation budget (not only an equal rollout count), how does BO perform with 20–100 evaluations, relative to SAAC’s 5? This would clarify the practical sample-efficiency claim.
  3. In-domain, what is the all-spec success rate (not only the geometric-mean score)?

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

Related papers

Discussion (0)

Sign in to comment.