acceptodds
Under review as a conference paper at ICLR 2027

H-OSIRIS: Hierarchical Layout-Aware Analog Circuit Sizing and Footprint Optimization using Deep Reinforcement Learning

Abstract

The automation of analog integrated-circuit layout remains challenging because circuit performance depends on coupled interaction among device sizing, parasitics, layout-dependent effects, and physical area. These effects are difficult to model accurately with rule-based or template-driven methods, particularly as circuits grow from individual functional blocks into multi-block systems. While machine learning has shown promising results for isolated stages of the analog design flow, an automated back-end optimization methodology that exploits the multi-block hierarchy of a layout while leveraging parasitic-aware feedback remains largely unexplored. This paper presents H-OSIRIS, a deep reinforcement learning methodology for post-layout optimization of multi-block analog circuits. H-OSIRIS operates at two optimization levels *(i)* it allocates area across circuit blocks globally and *(ii)* performs local ε-sizing of block devices. Guided by post-layout feedback, these mechanisms align post-layout functional metrics with their pre-layout references while reducing the overall layout footprint. Across four analog topologies and six metric sets, H-OSIRIS achieves 87% joint feasibility on held-out test samples while reducing the median layout area by 5.5%.

Primary area:
reinforcement learning
Keywords:
analog layout, reinforcement learning, optimization, hierarchy

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_1qkoOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

While machine learning has shown promising results for isolated stages of the analog design flow, an automated back-end optimization methodology that exploits a layout's multi-block hierarchy while leveraging parasitic-aware feedback remains largely unexplored. To address this, the authors present H-OSIRIS, a deep reinforcement learning methodology for post-layout optimization of multi-block analog circuits. Experimental results show joint feasibility rises from 0–9% to 72–99%, median area falls 3–8%.

Critical strengths

It uses genuine post-layout feedback (DRC/LVS-clean layouts, real extraction), not a parasitic proxy, which is very important. So the results are convincing.

Targeting an underexplored problem: back-end, parasitic-aware closure gets far less ML attention than front-end sizing or topology work.

Performance is strong, and the results are promising.

Critical weaknesses

The evaluation tasks are a little weak for the claims. Mostly target the Op-amp with one level of 2–3 blocks, yet “scalable” appears in the conclusions.

No variance is reported in the evaluation results.

The transfer and generalization capabilities are weak.

It is still a sizing work by considering physical feedback. The more challenging design automation part, i.e., placement and routing, is not addressed in this work.

Questions for authors

What is the relationship of this work with “OSIRIS: Bridging Analog Circuit Design and Machine Learning with Scalable Dataset Generation” ?

Can one trained policy handle a new target set, or a new initial sizing distribution?

Can you add more diverse designs?

What is the training cost (e.g., time) of this RL-based layout design method?

How is place-and-route addressed? Can you provide more details? If it is an open-source layout generator, no blind is needed.

2 comments
AuthorsOpenReview ↗

Response to Reviewer 1qko - Part 1

We thank the reviewer for recognizing the importance of genuine post-layout feedback, the relevance of the targeted problem, and the strength of the experimental results. We appreciate the questions concerning evaluation breadth, generalization, computational cost, and the role of place-and-route, which help us clarify the methodology’s scope and strengthen the supporting evidence.

W1. The benchmarks evaluate the targeted coordination problem: jointly optimizing block footprints and device sizing under coupled post-layout electrical constraints. They now cover five circuit compositions across seven metric sets. Amplifier-based benchmarks are also used in post-layout optimization studies such as OSIRIS [1] and Gao et al. [2]. The demonstrated result is improved optimization efficiency within these multi-block circuits: hierarchical training is 1.71.7-9.6×9.6\times faster than the evaluated flat formulation.

W2. We have added Section A.4.3 and Table 15, reporting 9595\\% Wilson intervals for terminal joint success and interquartile ranges for area ratios across the held-out starts. Terminal success ranges from 7272\\% to 9999\\% across the seven metric sets, and every set has a median area ratio below one. The new statistics show the spread behind the aggregate results and characterize variability for each trained policy.

W4. The proposed ϵ\epsilon-sizing mechanism is motivated by analog design workflows, where meeting post-layout performance targets requires close coordination between circuit designers and layout engineers. H-OSIRIS supports this process through an automated back-end optimization framework that connects device-sizing and block-footprint decisions with extracted post-layout performance feedback.

Q1. OSIRIS [1] and H-OSIRIS share the use of extracted post-layout feedback, but differ in their primary contribution, optimization variables, and coordination mechanism. OSIRIS presents a layout-dataset generation framework accompanied by an RL baseline that explores transistor fingers configurations and component displacements, using response discrepancy (pscore) and area as optimization objectives. H-OSIRIS, instead, targets joint electrical feasibility and footprint reduction in multi-block circuits. Its Manager adjusts block footprint boundaries, while block-specific Workers perform local ϵ\epsilon-sizing of device groups. This interface coordinates two complementary decisions: changing the geometric constraints of a block and adjusting its device parameters within those constraints. Their combined consequences are evaluated through full-circuit parasitic extraction and simulation. H-OSIRIS aligns two optimization levels with the circuit’s block structure and couples footprint control with local ϵ\epsilon-sizing, while requiring simultaneous satisfaction of explicit electrical-metric bounds and a global area constraint.

[1] Chiari, G., et al., “OSIRIS: Bridging Analog Circuit Design and Machine Learning with Scalable Dataset Generation”, ICLR, 2026.
[2] Gao, X., et al. "Post-layout simulation driven analog circuit sizing", Science China Information Sciences, 2024.

AuthorsOpenReview ↗

Response to Reviewer 1qko - Part 2

Q2, W3. To strengthen the experimental campaign, three additional batches of experiments have been conducted and added to Section 4 “Experimental Results”. First, Table 5 reports a greedy-Manager ablation that preserves the hierarchy, learned H-OSIRIS Workers, legal Manager action space, held-out starts, physical-design/simulation backend, and per-start action budget. Replacing the learned Manager with a fixed greedy rule reduces joint success from 7474\\% to 44\\% on 2-Stage Amp SA\mathcal{S}_A, from 9999\\% to 3434\\% on 3-OpAmp InAmp SB\mathcal{S}_B, and from 7272\\% to 2727\\% on Amp-Filter Comp SF\mathcal{S}_F. Together with the flat comparison in Table 4, this shows that the hierarchical decomposition alone is insufficient: learned Manager decisions provide additional value. Second, Table 6 compares H-OSIRIS with flat Bayesian optimization (BO), a non-RL baseline that directly searches legal circuit-wide sizings without Manager or Worker RL policies. BO uses the same 100100 held-out starts, layout-generation and post-layout simulation backend, and matched maximum per-start post-layout evaluation caps as H-OSIRIS: 253253 evaluations for Dual-OTA SE\mathcal{S}_E and 373373 for Amp-Filter Comp $\mathcal{S}_F$. These demanding tasks respectively contain the largest constraint set and a heterogeneous amplifier-filter hierarchy. Flat BO attains joint feasibility on 0/1000/100 SE\mathcal{S}_E starts and 5/1005/100 SF\mathcal{S}_F starts, compared with 92/10092/100 and 72/10072/100 for H-OSIRIS. This comparison is framed as a matched test-time evaluation-budget comparison. Third, Table 7 evaluates transfer across sizing ranges and metric assignments. Under sizing ranges 5050\\% wider than those used during training, the evaluated policies achieve 9494\\% success for 3-OpAmp InAmp SB\mathcal{S}_B and 7575\\% for Dual-OTA SC\mathcal{S}_C. For metric transfer within Dual-OTA, the source policies are fine-tuned on the target metric set before held-out testing. Transfer from SD\mathcal{S}_D to SC\mathcal{S}_C achieves 77%77\% success after 220220 fine-tuning episodes, while transfer from SE\mathcal{S}_E to SD\mathcal{S}_D achieves 9090\\% after 4040 episodes. The corresponding cold-start H-OSIRIS policies achieve 9393\\% and 9292\\%, respectively.

Q3. We added a new topology (Cascode), extending the evaluation beyond amplifier-based designs. It introduces two voltage targets, VBV_B and VOV_O, alongside the global area constraint. H-OSIRIS achieves joint feasibility on 93/10093/100 held-out starts and a 44\\% median area reduction. The topology, and its targets are detailed in Section 4.1, Table 2, and Appendix Figure 7. Table 3, Table 4, and Figure 10 report its results.

Q4. Table 4 reports the training time in hours for each topology-metric-set pair, for both H-OSIRIS and the matched flat baseline. H-OSIRIS requires 1010-6161 hours, compared with 5151-465465 hours for the flat baseline, corresponding to 1.7×1.7\times-9.6×9.6\times less training time. At the same time, H-OSIRIS achieves 7272-9999\\% joint terminal success, compared with 00-4040\\% for the flat baseline.

Q5. Place-and-route is handled by an existing open-source layout generator. Its reference was anonymized as a precaution to preserve double-blind review. The complete citation will be restored in the camera-ready version. The generator receives the updated device parameters and the block-footprint bounds selected by H-OSIRIS, and performs placement and routing under these constraints. The resulting layout undergoes DRC and LVS verification, parasitic extraction, and post-layout simulation, providing the electrical and area measurements used for the optimization feedback. H-OSIRIS does not propose a new placement or routing algorithm. Its contribution is the learned coordination of block-footprint constraints and local device sizing. To this end, in the revised paper, we expanded the state-of-the-art discussion in Section 2 “Related Work” and provided further evaluation in Section 4.2 “Experimental Results”.

Reviewer_uf56OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper presents H-OSIRIS, a hierarchical reinforcement learning framework for post-layout analog circuit sizing and area optimization. A graph-based Manager adjusts functional-block footprint constraints, while block-specific Workers optimize device parameters using post-layout simulation feedback. Experiments across multiple circuit topologies and performance objectives demonstrate improvements in design feasibility, layout area, and training efficiency compared with a flat RL baseline.

Critical strengths
  1. Unlike approaches that rely solely on pre-layout simulations, the proposed framework integrates automated layout generation, parasitic extraction, and post-layout simulation into the optimization loop, allowing it to account for layout-dependent effects.

  2. The experiments demonstrate substantial improvements in design feasibility and training efficiency over the flat RL baseline across multiple circuit topologies and performance objectives.

Critical weaknesses
  1. The methodological novelty of the proposed approach is limited and not convincingly established relative to prior work. The framework operates on fixed circuit topologies with predefined functional blocks, sizing ranges, and footprint constraints, while the underlying problem of optimizing device parameters and area under performance constraints has been extensively studied. Although the integration of hierarchical RL with post-layout feedback is interesting, the paper does not sufficiently explain what fundamentally distinguishes its optimization methodology from existing approaches or why RL is particularly beneficial for this task. Furthermore, the comparison with prior work in Table 1 raises concerns, as criteria such as block-level sizing are not clearly defined, and some classifications appear questionable. These issues make it difficult to assess the claimed novelty and significance of the contribution.

  2. The presentation makes the paper unnecessarily difficult to follow, even for readers familiar with analog circuit optimization. The extensive use of abbreviations, including unnecessary ones for commonly understood terms (e.g., “NL” for netlist), and the scattering of important explanations across different sections and appendices significantly reduce readability. Although Figures 1 and 2 illustrate the general analog design flow and the training procedure, respectively, a clear overview of the proposed end-to-end optimization workflow is still missing. A more informative overall pipeline figure, along with clearer organization and more selective use of abbreviations, would substantially improve the paper's accessibility.

Questions for authors
  1. Section 4.2 explains that the randomly sampled sizing configurations used in Figures 3 and 8 are distinct from the held-out initial states used in Table 3. Could the authors clarify how the held-out starting configurations were generated and whether they satisfy the nominal performance specifications at the schematic level? This distinction is important for assessing how closely the evaluation reflects practical post-layout optimization scenarios, particularly those starting from schematic-level designs that already meet their performance requirements.

  2. The results in Table 4 demonstrate substantial improvements over the flat RL baseline. However, what evidence supports the necessity or advantages of reinforcement learning compared with established circuit optimization methods? A comparison with non-RL optimization approaches would help establish whether the proposed method offers benefits beyond the hierarchical decomposition of the problem.

  3. To what extent are the improvements over flat RL attributable to the learned Manager policy, as opposed to the decomposition of the circuit into smaller optimization problems? An ablation retaining block-specific Workers but using a fixed or heuristic block-selection strategy would help clarify the contribution of the learned hierarchy.

  4. Some of the classifications in Table 1 appear questionable, and the criteria used to evaluate prior approaches are not sufficiently clear. Could the authors provide more precise definitions of the comparison criteria and justify how full, partial, or no support is assigned to each approach? Without clearer definitions and justification, these classifications may not accurately reflect the capabilities of prior work, making it difficult to assess the claimed novelty of the proposed approach.

2 comments
AuthorsOpenReview ↗

Response to Reviewer uf56 - Part 1

We thank the reviewer for recognizing the value of integrating post-layout feedback into optimization and the improvements in feasibility and training efficiency over the flat RL baseline. We appreciate the questions concerning methodological differentiation, evaluation, and comparison criteria, as well as the suggestions for improving readability, which help us clarify the contribution and strengthen its presentation.

W1. H-OSIRIS’s main contribution is the learned coordination of block-footprint constraints and local device sizing through full-circuit post-layout feedback. The Manager does more than select a block: it changes the geometric constraints under which the Worker optimizes. Subsequent Manager decisions account for the resulting electrical and area effects across the complete circuit. This couples layout compaction with electrical refinement while accounting for inter-block parasitic interactions. Fixed topologies and predefined bounds specify the refinement task. The policies learn how to navigate it. Across four circuits and six metric sets, H-OSIRIS achieves 7272-9999\\% held-out joint feasibility and median area reductions of 33-88\\%. Against the evaluated flat formulation, it improves feasibility from 00-4040\\% to 7272-9999\\% and reduces training time by factors of 1.71.7-9.69.6. Please see response Q4 (below) for clarification of Table 1.

W2. We have revised the presentation to make the workflow easier to follow. Unnecessary abbreviations, including NL, CN, and P&R, have been removed. Section 3 now opens with a brief introduction followed by Section 3.1 “Hierarchical Architecture” introducing the Manager-Worker organization before the detailed mechanisms. Figure 2 presents the complete optimization loop, connecting Manager decisions, Worker sizing, layout generation, post-layout evaluation, and policy updates. Its inputs labels have been revised to remove NL and CN. We have also reordered the appendix so that “Manager and Worker States” precedes the “Area-Boundary Action Example,” and expanded Section 3.4 “Reward Definition” with clearer links to its detailed definitions in Section A.1.3 Reward Definition Details. The paper now presents the concepts essential to understanding the methodology and interpreting its results in the main text, while reserving supporting technical derivations, implementation details, and experimental configurations for the appendix.

Q1. The held-out starting configurations are distinct from the random characterization samples shown in Figure 3 and Figure 8. For each test set, candidates are generated using a Latin hypercube over the sizing ranges, excluding configurations previously used for random characterization or as training-episode starting points. Section 4.2 explicitly clarifies this distinction to avoid confusion between the two sets of 100100 configurations. The random characterization samples are used to demonstrate the impact of layout-induced parasitics on circuit performance, whereas the held-out configurations are independently generated to evaluate the proposed methodology. Furthermore, we simulated the 100100 held-out starting schematics before layout and compared their electrical metrics with the target and done bands. Among starts with valid pre-layout simulations, all considered metrics jointly met their target bands for 0/1000/100 SA\mathcal{S}_A, 0/1000/100 SB\mathcal{S}_B, 11/10011/100 SC\mathcal{S}_C, 13/10013/100 SD\mathcal{S}_D, 4/1004/100 SE\mathcal{S}_E, 0/1000/100 SF\mathcal{S}_F, and 21/10021/100 SG\mathcal{S}_G starts. We added Figure 9 and expanded the discussion in Section A.4.2 in the revised paper. These results show that the held-out sizings often begin far from the desired operating region. Together with the post-layout outcomes in Figure 10, they further highlight the role of RL: the learned policy can recover desired post-layout performance even from poor starting states.

AuthorsOpenReview ↗

Response to Reviewer uf56 - Part 2

Q2, Q3. To strengthen the experimental campaign, three additional batches of experiments have been conducted and added to Section 4 “Experimental Results”. First, Table 5 reports a greedy-Manager ablation that preserves the hierarchy, learned H-OSIRIS Workers, legal Manager action space, held-out starts, physical-design/simulation backend, and per-start action budget. Replacing the learned Manager with a fixed greedy rule reduces joint success from 7474\\% to 44\\% on 2-Stage Amp SA\mathcal{S}_A, from 9999\\% to 3434\\% on 3-OpAmp InAmp SB\mathcal{S}_B, and from 7272\\% to 2727\\% on Amp-Filter Comp SF\mathcal{S}_F. Together with the flat comparison in Table 4, this shows that the hierarchical decomposition alone is insufficient: learned Manager decisions provide additional value. Second, Table 6 compares H-OSIRIS with flat Bayesian optimization (BO), a non-RL baseline that directly searches legal circuit-wide sizings without Manager or Worker RL policies. BO uses the same 100100 held-out starts, layout-generation and post-layout simulation backend, and matched maximum per-start post-layout evaluation caps as H-OSIRIS: 253253 evaluations for Dual-OTA SE\mathcal{S}_E and 373373 for Amp-Filter Comp $\mathcal{S}_F$. These demanding tasks respectively contain the largest constraint set and a heterogeneous amplifier-filter hierarchy. Flat BO attains joint feasibility on 0/1000/100 SE\mathcal{S}_E starts and 5/1005/100 SF\mathcal{S}_F starts, compared with 92/10092/100 and 72/10072/100 for H-OSIRIS. This comparison is framed as a matched test-time evaluation-budget comparison. Third, Table 7 evaluates transfer across sizing ranges and metric assignments. Under sizing ranges 5050\\% wider than those used during training, the evaluated policies achieve 9494\\% success for 3-OpAmp InAmp SB\mathcal{S}_B and 7575\\% for Dual-OTA SC\mathcal{S}_C. For metric transfer within Dual-OTA, the source policies are fine-tuned on the target metric set before held-out testing. Transfer from SD\mathcal{S}_D to SC\mathcal{S}_C achieves 77%77\% success after 220220 fine-tuning episodes, while transfer from SE\mathcal{S}_E to SD\mathcal{S}_D achieves 9090\\% after 4040 episodes. The corresponding cold-start H-OSIRIS policies achieve 9393\\% and 9292\\%, respectively.

Q4. We agree that Table 1 did not explain its comparison criteria clearly enough. We have clarified the headings and expanded the Related Work discussion to define block-local sizing, direct layout-area control, and post-layout feedback used during optimization. The revised text distinguishes these from circuit-wide sizing, analytical area estimates, and predicted performance; final post-layout validation alone is not counted as optimization feedback. We have also reviewed the full, partial, and no-support labels against these definitions.

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

The paper presents H-OSIRIS, a hierarchical reinforcement learning framework for analog circuit sizing and footprint optimization. A Manager adjusts block footprint bounds, while block-specific policies tune device parameters using full-circuit post-layout feedback. Across four circuit topologies and six metric sets, the method achieves 87% joint feasibility over 600 held-out initializations, with median area reductions of 3–8%.

Critical strengths
  • The separation between block footprint control and local device sizing is physically motivated and provides a reasonable structure for joint electrical and layout optimization.
  • Each sizing action is evaluated through full-circuit layout generation, parasitic extraction, and simulation, allowing the optimization to account for layout-dependent electrical effects.
Critical weaknesses
  • The methodological novelty appears limited beyond the footprint–sizing interface. Block-based decomposition and hierarchical RL have close precedents, and the paper does not clearly explain what additional coordination capability this interface provides.
  • The comparisons are limited to random and flat policies. Since the flat baseline also changes the geometry-control interface, the results do not isolate the benefit of learned hierarchical coordination or establish competitiveness against existing optimizers.
Questions for authors
  • Could the authors clarify the methodological differences from MA-RL and OSIRIS, particularly what the footprint–sizing interface adds beyond their existing decompositions?
  • How does the method compare with a simple heuristic Manager under the same layout backend, action space, and number of post-layout evaluations?
  • How does the Manager compensate when a Worker improves its assigned metrics but degrades another block or a global metric? A concrete example would help.
2 comments
AuthorsOpenReview ↗

Response to Reviewer Fxkt - Part 1

We thank the reviewer for recognizing the value in the separation between footprint control and local sizing, and the importance of full-circuit post-layout feedback. We appreciate the questions concerning methodological differentiation and the contribution of learned coordination, which help us clarify the scope and evidence of our contribution.

W1. H-OSIRIS' main contribution is the learned coordination of block-footprint constraints and local device sizing through full-circuit post-layout feedback. The Manager does more than select a block: it changes the geometric constraints under which the Worker optimizes. Subsequent Manager decisions account for the resulting electrical and area effects across the complete circuit. This couples layout compaction with electrical refinement while accounting for inter-block parasitic interactions. Fixed topologies and predefined bounds specify the refinement task. The policies learn how to navigate it. Across four circuits and six metric sets, H-OSIRIS achieves 7272-9999\\% held-out joint feasibility and median area reductions of 33-88\\%. Against the evaluated flat formulation, it improves feasibility from 00-4040\\% to 7272-9999\\% and reduces training time by factors of 1.71.7-9.69.6.

Q1. H-OSIRIS’s contribution is not block decomposition or hierarchical RL alone, but the coordination of physical footprint constraints and block-local device sizing through full-circuit post-layout feedback. MA-RL [1] decomposes circuit-parameter optimization among interacting block-specific agents. H-OSIRIS introduces a different coordination interface: the Manager selects a block and adjusts its allowable footprint, while the corresponding Worker adapts device sizing under that geometric constraint. Thus, the Manager does more than select the next subproblem. It changes the physical conditions under which local optimization proceeds. Layout regeneration and extracted simulation then reveal the combined electrical and geometric consequences. OSIRIS [2] provides a parasitic-aware layout-generation framework along an RL baseline separating finger permutations search from component movement optimization. H-OSIRIS instead organizes decisions around interconnected circuit blocks, coupling block-boundary control with local ϵ\epsilon-sizing. It also evaluates feasibility against explicit metrics, rather than relying on an aggregate response-discrepancy objective (pscore). The footprint-sizing interface therefore enables the policies to coordinate layout compaction with electrical adjustment: a tighter block boundary changes the physical implementation available to the Worker, while the resulting full-circuit performance informs subsequent Manager decisions.

[1] Bao, J., et al., “Multiagent based reinforcement learning (MA-RL): An automated designer for complex analog circuits”, IEEE TCAD, 2024.
[2] Chiari, G., et al., “OSIRIS: Bridging Analog Circuit Design and Machine Learning with Scalable Dataset Generation”, ICLR, 2026.

AuthorsOpenReview ↗

Response to Reviewer Fxkt - Part 2

Q2, W2. To strengthen the experimental campaign, three additional batches of experiments have been conducted and added to Section 4 “Experimental Results”. First, Table 5 reports a greedy-Manager ablation that preserves the hierarchy, learned H-OSIRIS Workers, legal Manager action space, held-out starts, physical-design/simulation backend, and per-start action budget. Replacing the learned Manager with a fixed greedy rule reduces joint success from 7474\\% to 44\\% on 2-Stage Amp SA\mathcal{S}_A, from 9999\\% to 3434\\% on 3-OpAmp InAmp SB\mathcal{S}_B, and from 7272\\% to 2727\\% on Amp-Filter Comp SF\mathcal{S}_F. Together with the flat comparison in Table 4, this shows that the hierarchical decomposition alone is insufficient: learned Manager decisions provide additional value. Second, Table 6 compares H-OSIRIS with flat Bayesian optimization (BO), a non-RL baseline that directly searches legal circuit-wide sizings without Manager or Worker RL policies. BO uses the same 100100 held-out starts, layout-generation and post-layout simulation backend, and matched maximum per-start post-layout evaluation caps as H-OSIRIS: 253253 evaluations for Dual-OTA SE\mathcal{S}_E and 373373 for Amp-Filter Comp $\mathcal{S}_F$. These demanding tasks respectively contain the largest constraint set and a heterogeneous amplifier-filter hierarchy. Flat BO attains joint feasibility on 0/1000/100 SE\mathcal{S}_E starts and 5/1005/100 SF\mathcal{S}_F starts, compared with 92/10092/100 and 72/10072/100 for H-OSIRIS. This comparison is framed as a matched test-time evaluation-budget comparison. Third, Table 7 evaluates transfer across sizing ranges and metric assignments. Under sizing ranges 5050\\% wider than those used during training, the evaluated policies achieve 9494\\% success for 3-OpAmp InAmp SB\mathcal{S}_B and 7575\\% for Dual-OTA SC\mathcal{S}_C. For metric transfer within Dual-OTA, the source policies are fine-tuned on the target metric set before held-out testing. Transfer from SD\mathcal{S}_D to SC\mathcal{S}_C achieves 77%77\% success after 220220 fine-tuning episodes, while transfer from SE\mathcal{S}_E to SD\mathcal{S}_D achieves 9090\\% after 4040 episodes. The corresponding cold-start H-OSIRIS policies achieve 9393\\% and 9292\\%, respectively.

Q3. We revised Figure 4 to make each quality metric’s trajectory easier to follow alongside the Worker actions and Manager decisions. Episode A illustrates a trade-off: after the Manager tightens the BUFFER block’s width allowance, the Worker reduces a transistor’s width twice. BUFFER bandwidth enters its done band and BUFFER gain moves closer to its target, while CORE bandwidth moves farther outside its done band. The Manager then reallocates to CORE. The ensuing CORE rollout overshoots the CORE bandwidth band and temporarily increases global area. Subsequent CORE and BUFFER boundary adjustments bring both bandwidths into their done bands and reduce global area below its done limit. The episode ends with all configured metrics within their done bands.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Overall Revision Comment

In response to the reviewers’ feedback, we revised the manuscript. The main revisions are:

  • Improved readability (uf56): reorganized the opening of Section 3 “The H-OSIRIS Methodology” with a dedicated “Hierarchical Architecture” subsection, removed unnecessary abbreviations, simplified Figure 2 labels, reordered appendix explanations, and clarified links between the main-text and appendix reward definitions.
  • Clarified comparison criteria (uf56): expanded the explanation of Table 1 in Section 2 “Related Work”.
  • Additional transfer-learning experiments (1qko): evaluated transfer to new quality-metric target sets and sizing ranges, reported in Table 7 and the “Transfer learning” paragraph of Section 4.2 “Experimental Results”.
  • Additional policy-ablation experiments (Fxkt, uf56): examined the contribution of learned policies, reported in Table 5 and Table 6 and the “Policy ablation” paragraph of Section 4.2 “Experimental Results”.
  • Additional held-out initialization analysis (uf56): evaluated the pre-layout performance of held-out sizing configurations. Reported in Figure 8.
  • Additional topology (1qko): added a new topology targeting new quality metrics. The results are reported throughout Section 4.2 “Experimental Results”.
  • Reported evaluation variability (1qko): added 9595\\% Wilson intervals for terminal success and interquartile ranges for area ratios across held-out starts in Section A.4.3 and Table 15.

Additions and edits are highlighted in blue in the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.