acceptodds
Withdrawn from ICLR 2027

GLASS: A Glass-Box, Loop-Verified Method for Analog Circuit Sizing with Symbolic Priors

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Agentic analog circuit sizing is now practical, but each candidate design requires an expensive SPICE simulation, and unguided search carries little explicit design knowledge. In controlled comparisons, the failures we observe stem from what we term a cold-start trap: the agent exhausts its simulation budget in electrically degenerate regimes. We propose GLASS, a glass-box, loop-verified method that resolves this trap at initialization and preserves search momentum through stall recovery and numerical warm-starting. GLASS comprises: 1) SIZE-90K, a dataset of 90,858 sizing simulations from six amplifier topologies across four technology platforms; 2) Symbolic-Op, a calibrated behavioral symbolic layer that converts solver-derived candidates into injectable priors and audits the initial prior by re-simulation in the target testbench; and 3) Multi-Task Surrogate, a compact neural network that predicts six specification metrics and screens proposals by predicted cost. On 12 sizing tasks, GLASS reaches 60.0% strict Pass@1 and 81.7% Pass@3, surpassing the strongest evaluated baseline by 36.7 percentage points with same-topology calibration and surrogate training. Under matched proposal budgets on a single topology, the priors raise the share of search chains that meet specification from 26.3% to 56.0% and reduce the median evaluations required from 258 to 148. Across ablations, degenerate designs fall from 29/60 without symbolic guidance to 0/60 under the full operator set. We release our code at https://anonymous.4open.science/r/GLASS-analog-sizing-86CF

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
neurosymbolic AI, agentic AI, large language models, symbolic priors, surrogate models, interpretability, analog circuit sizing, electronic design automation

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 4 reviews, above 27% of the others.
Reviewer_hjTJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The article proposes GLASS, a neuro-symbolic method for analog circuit sizing that combines calibrated symbolic design priors with an LLM-guided search loop and a surrogate model that screens some optimizer proposals before SPICE simulation. Its contributions include SIZE-90K, a dataset of 90,858 simulations across six amplifier topologies and four technology platforms, and symbolic operators for initialization, stall recovery, and optimizer warm-starting. On 12 sizing tasks, GLASS achieves 60.0% Pass@1 and 81.7% Pass@3, outperforming the evaluated agentic baselines.

Critical strengths
  1. The flow is compared against against two agentic baselines and ablation study shows a substantial improvement over both seed-only and no-symbolic variants, giving evidence that the closed-loop interventions contribute beyond initialization.

  2. The loop verifies accepted designs with transistor-level SPICE and makes prior failures visible. A failure case is provided and the flow identifies a failed mapping instead of presenting the analytical prediction as a valid result.

Critical weaknesses
  1. The reported search budget excludes offline dataset generation; the paper estimates that its cost breaks even only after 800–1,200 chains. The evaluation is much smaller, and the per-topology effort to develop symbolic models is not quantified, making the pratical cost unclear.

  2. The benchmark doesn’t test the same topology across multiple PDKs in a controlled matrix. So it’s hard to tell whether success comes from the method’s portability or from a model and calibration tailored to each task. Moreover, six amplifier topologies on four different PDKs are useful coverage, but they don’t show the approach works for other analog design problems.

Questions for authors
  1. Have you tried the same circuit across the 4 PDKs?

  2. Have you tried the flow on other circuit families, like LDOs, Comparators, DAC/ADCs with their dedicated metrics?

Reviewer_KX4yOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

GLASS augments agentic analog circuit sizing with calibrated symbolic priors at initialization, stall recovery, and the transition to numerical optimization. A neural surrogate trained on 90,858 simulations screens some proposals, while SPICE determines specification satisfaction. On 12 tasks spanning six topologies and four technology platforms, the method reports 36/60 successful runs, or 60.0% Pass@1, under an explicitly disclosed 5% tolerance in each metric's native scale. Its most persuasive contribution is the controlled evidence that continued prior intervention improves on initialization alone, together with informative analyses of unsuccessful cases.

Critical strengths
  1. The controlled experiment supports a meaningful benefit beyond a good initial seed. Table 4 holds the scaffold, language model, prompts, and proposal budget fixed: success rises from 10.0% without symbolic operators to 31.7% with initialization and 60.0% with the full set. The seed-only comparison already shares symbolic initialization resources, so this result should receive credit as evidence for the combined downstream interventions. Reported degenerate final outcomes decrease from 29/60 to 7/60 to 0/60. Using domain priors to maintain productive operating regions, including interventions triggered by stagnation, is a useful design principle for simulator-driven agents, even without a new neural architecture.
  2. The paper exposes important boundaries and preserves simulator-based assessment. The PFC failures remain in the main aggregate; the seed-transfer audit is identified as diagnostic; session-grouped surrogate degradation is reported; and broader compression failures appear in Appendix G. SIZE-90K and the documented workflow also offer practical value for research on sparse, heterogeneous simulation feedback. These disclosures strengthen the contribution within its stated regime.
Critical weaknesses
  1. The screening component's net efficiency benefit remains unresolved. Table A5 reports 14/25 successes and 292 mean simulator calls for warm start, versus 2/12 and 180 calls for screening. These are small, unequal cohorts; they do not establish that screening causes harm, but fewer calls alone do not establish a better success–cost tradeoff either. It is also unclear whether the screening arm contains precisely the same preceding components. The 148-versus-258 median compares successful chains only, and the strongest compression evidence is confined to DFCFC. The disclosed 90,858 offline simulations and 800–1,200-chain amortization further restrict the efficiency interpretation to repeated use with prepared data.
  2. The transferable insight and evaluation scope need precise boundaries. The contribution concerns the integration and timing of priors, rather than the first use of equations or feedback: the cited White-Op (arXiv:2601.21321), self-calibrating framework (arXiv:2604.07387v1), and NEMESIS (arXiv:2607.05657) already employ related physics and simulation loops. Table 4 does not separately identify stall recovery and warm start, and its final measurement-failure counts do not directly diagnose every transistor's operating region. Moreover, the substantial session-grouped fidelity decline in Table A4 makes an explicit separation of final evaluation sessions from training and calibration important. That decline is not itself evidence of test leakage.
Questions for authors
  1. Please clarify the exact composition and unequal cohort sizes of Table A5. Can existing logs show success against cumulative SPICE calls with screening enabled and disabled, including unsuccessful chains?
  2. Were the final benchmark sessions excluded from both surrogate training and symbolic calibration? Please state the split unit, shared topology/specification information, and which resources are reused across tasks.
  3. Can the original 60 trials per method, including the external baselines, be re-scored without tolerance? Appendix C reports a different, larger cohort, so its 38.2% cannot replace the headline 60.0% directly.
Reviewer_LzdtOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes GLASS, a framework for improving analog circuit sizing with symbolic priors and SPICE verification. It addresses the cold-start problem, where poor initial designs lead to wasted simulations and inefficient searches. GLASS uses circuit knowledge to generate better initial designs and guide the optimization process through three strategies: initialization, stall recovery, and warm-start. It also trains a multi-task surrogate model on the SIZE-90K dataset to filter out unpromising designs and reduce simulation costs. Experiments on 12 sizing tasks show that GLASS achieves a 60.0% Pass@1 and 81.7% Pass@3, outperforming the evaluated baselines.

Critical strengths
  1. The paper addresses a practical problem in analog circuit sizing, where poor initialization can waste many SPICE simulations.
  2. The three symbolic injection strategies provide a reasonable way to guide the optimization process and help the search recover when it gets stuck.
  3. The experiments show clear improvements in success rate and a significant reduction in degenerate designs compared with the baselines.
Critical weaknesses
  1. The ablation study does not separately evaluate stall recovery and warm-start, so it is unclear how much each contributes to the improvement.
  2. The surrogate model performs much worse on unseen optimization sessions, which raises concerns about its generalization ability.
  3. The simulation efficiency improvements are mainly demonstrated on one circuit topology, and surrogate screening reduces the success rate, so its overall benefits remain unclear.
Questions for authors
  1. Could the authors provide separate ablation results for stall recovery and warm-start to clarify how much each contributes to the overall improvement?
  2. Could the authors further evaluate the surrogate model on unseen optimization sessions to better understand its generalization ability?
  3. Could the authors provide more evidence of simulation efficiency improvements across different circuit topologies, especially considering that surrogate screening reduces the success rate in the reported experiments?
Reviewer_ynUGOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

GLASS combines symbolic priors, LLM-based refinement, numerical optimization, and surrogate screening for analog circuit sizing. The results are promising, but the incremental novelty and robustness of the gains remain unclear.

Critical strengths

The evaluation covers 12 tasks with repeated trials. Final designs are SPICE-verified, and the ablations show improvements beyond initialization alone.

Critical weaknesses
  1. The novelty beyond combining established techniques is unclear. Stall recovery and numerical warm-starting are not individually isolated in the main ablation.

  2. All experiments use Gemma-4-31B-it. Whether the cold-start problem and the benefits of symbolic guidance persist with stronger LLMs remains untested.

  3. The comparisons do not establish whether existing optimizers given equivalent calibrated priors and offline data could achieve similar gains. Offline simulation costs are excluded from the online comparison.

Questions for authors
  1. Please repeat the GLASS ablations with a stronger frontier LLM under matched simulation budgets. Does symbolic guidance still improve success rates and simulation efficiency?

  2. What does stall recovery contribute separately from numerical warm-starting?

  3. How does GLASS compare with existing optimizers given the same priors and data, accounting for preparation costs?

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

Related papers

Discussion (0)

Sign in to comment.