- 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?