acceptodds
Under review as a conference paper at ICLR 2027

VPEvolve: A Self-Evolving Virtual Process Engineer for Computational Lithography

Abstract

Optical proximity correction (OPC) recipes grow as engineers add local rules to repair newly discovered lithography hotspots. Each correction can interact with existing rules, while lessons from commercial-tool trials remain scattered across code and logs. VPEVOLVE combines a Virtual Process Engineer (VPE) harness with a Skill Bank of measured engineering experience. The harness equips a frozen language model with process manuals, layout analysis, recipe editing, and commercial-tool evaluation. The actor proposes changes to the global parameters, local targeted rules, or diagnostic trials. After each evaluation, an LLM reflector and curator turn the measured response into evidence-linked judgments. The actor retrieves them before its next trial. Feasible improvements update the retained recipe; every measured trial informs the Skill Bank that guides the next edit. The model weights remain fixed. On a FreePDK45-derived benchmark with ten commercial-tool evaluations per case, VPEVOLVE reduces the mean per-case maximum edge placement error from 18.294 to 5.361 nm on Poly and from 22.052 to 15.692 nm on Metal1. Every final recipe satisfies the predefined quality constraints and improves the maximum error by at least 0.1 nm. Project website: https://vpevolve.github.io/vpevolve/.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
LLM agents, self-evolving agents, virtual process engineer, computational lithography

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.25 over 4 reviews, above 96% of the others.
Reviewer_NdsiOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper automates OPC recipe development. Its VPE harness, the Pi agent harness plus domain prompts and tools, lets an LLM change global parameters, add local correction rules, or run diagnostic trials, and Calibre scores each trial once. After every trial, two other agents follow the ACE idea to look back at the result and add what they learn to a skill bank. On a FreePDK45 benchmark the method does well against six general agent methods and Bayesian optimization.

Critical strengths
  1. Though I don't work on OPC myself, I think the application has its value. There isn't much work on agents for OPC yet, and here a commercial lithography tool is in the loop. The constraints the paper works under, such as mask rules, SRAF printing, and caps on mean EPE and PVB, might well be what an engineer actually cares about.
  2. The paper provides its own evaluation protocol for agentic OPC recipe search.
Critical weaknesses
  1. The comparison with the baselines in Table 1 does not show how much of the gain comes from the experience mechanism. In the supplementary code, the method presets for the memory baselines (Textual Memory, ReasoningBank, ACE) and ReAct turn off the engineering guidance that VPEvolve gets, and VPEvolve and its ablations also get a separate prompt on how to search and a longer instruction. The paper doesn't say whether these baselines get the guidance. Frozen Experience, which never updates its experience, already beats every baseline in Table 1, so much of the gap may come from the prompts rather than from what the agent learns. The continuation study in Section 4.4.1 does hold the guidance fixed, but its intervals include zero and each case is run only once.

  2. It is not clear why OPC needs its own harness. The introduction lists what recipe work needs (process knowledge, layout analysis, recipe editing and tool feedback), but these could be given to a general self-evolve harness such as Hermes Agent as tools and domain prompts. Also, since the method is named for evolving, I expected some discussion of methods that optimize the harness itself, such as Meta-Harness [1] or AutoHarness [2], and neither is discussed. AnalogAgent [3], which already carries ACE-style memory into EDA, isn't discussed either. The code also reuses ACE's prompt templates (its NOTICE file says so), which the paper doesn't mention.

References

  • [1] Lee et al. Meta-Harness: End-to-End Optimization of Model Harnesses. arXiv:2603.28052.
  • [2] Lou et al. AutoHarness: improving LLM agents by automatically synthesizing a code harness. arXiv:2603.03329.
  • [3] Bao et al. AnalogAgent: Self-Improving Analog Circuit Design Automation with LLM Agents. arXiv:2603.23910.
Questions for authors

Clarification requests

  1. What does the VPE harness give that a general self-evolving harness with the same domain prompts can't? To demonstrate the difference, I recommend running Hermes Agent with the same prompts and tools on a few cases.
  2. Did the agent baselines in Table 1 other than Static Harness get the engineering guidance that VPEvolve gets? The paper doesn't say. If they did not, rerun ACE, the method VPEvolve's experience mechanism adapts, with the same guidance on a few cases.
  3. Why is Bayesian optimization the only traditional method in the introduction and the only numerical search baseline, and what standing does it have in OPC practice?
  4. Figure 7 is meant to test whether VPEvolve depends on a particular LLM (Section 4.4.4), but every API model ends behind the default 27B model on both layers. Why is that, and does the gap hold if the context length is matched? Rerun one or two API models with the full context on a few cases, or explain where the gap comes from.
  5. What should an NLP reader from the LLM agent community take from this work beyond OPC? For example, how much of the gain comes from domain guidance versus accumulated experience, and what kind of feedback makes the experience useful? The ablations in Table 2 start on this, and I think they deserve a short paragraph, or even a section in the experiments, on what they teach the community about agents.

Suggestions further

  1. Section 4.4.1 says the gains "show" that curated experience can improve later decisions, and Section 4.4.2 presents local corrections as another way to cut the worst errors, but all four paired intervals in Table A4 include zero. I would say tone both statements down to what the intervals support.
  2. Rewrite the contribution list. The first bullet calls recipe development "a learning problem", the text under it says "sequential search", and Section 3.1 says "sequential decision problem" (I come back to this in the next suggestion). I would pick one term and say in a sentence what it adds over the RL formulation in Intelligent OPC Engineer Assistant [4], which the paper itself cites and which framed recipe search as reinforcement learning back in 2024. The third bullet is titled as an outcome, self-evolving experience, but describes a mechanism, reflection and curation, so it should say which one is claimed. The fourth bullet is titled "an open testbed", but the recipe task itself still needs Calibre, which is closed source software. The title seems to promise more than what is released.
  3. I think the introduction should be rewritten around one thread. Right now the paper moves between the harness, the memory mechanism and the "learning problem" of the first contribution, and the last one confuses me a lot. The formulation has no loss function, and I don't see how it can be framed as a learning problem. At best I think it is a memory accumulation problem. The contribution also says it is a search problem, but I can't find it formally framed as one in the paper, which is really confusing.

References (continued)

  • [4] Chen et al. Intelligent OPC Engineer Assistant for Semiconductor Manufacturing. arXiv:2408.12775.
Reviewer_VEpYOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents VPEVOLVE, a self-evolving language agent framework designed to automate optical proximity correction (OPC) recipe development in computational lithography. Addressing the issue where human engineering experience is fragmented across trial-and-error logs and local rule edits, the framework couples a Virtual Process Engineer (VPE) harness with a versioned "Skill Bank". The harness equips a frozen LLM with domain manuals, layout analysis tools, recipe compilation checks, and commercial lithography evaluation (Siemens Calibre). Through an actor-reflector-curator loop, the system executes global parameter adjustments and targeted local rule corrections, distilling experimental outcomes into evidence-linked engineering judgments that guide subsequent iterations.

The method is evaluated on a FreePDK45 benchmark across 20 layout clips (Poly and Metal1) under a fixed 10-trial commercial evaluation budget. Results demonstrate substantial reductions in maximum edge placement error (EPE) from 18.294 nm to 5.361 nm on Poly and 22.052 nm to 15.692 nm on Metal1, achieving a 100% success rate under strict manufacturability and process variation constraints. The authors provide thorough mechanism ablations, paired continuation experiments, 100-trial long-horizon runs, and cross-model evaluations across six LLM backbones.

Critical strengths
  1. Novel and grounded formulation of OPC recipe engineering as an agentic decision process: Instead of treating recipe tuning purely as numerical black-box parameter search (e.g., standard Bayesian Optimization) or unconstrained text generation, VPEvolve elegantly combines global parameter tuning with geometry-aware local rule synthesis. Grounding the LLM with deterministic syntax/geometry compilers and commercial-grade lithography feedback (Siemens Calibre) ensures that all proposed interventions remain strictly valid and physically actionable.

  2. Effective experience distillation and rigorous empirical validation: The domain-specific adaptation of the Skill Bank (incorporating spatial hotspot reflection, scoped revision, and falsifiable experiment hypotheses) effectively solves the problem of "rule side-effects" and moving bottlenecks in lithography. The experimental campaign is comprehensive, featuring meaningful ablations, paired continuation studies isolating experience curation value, long-horizon stability tests (100 trials), and consistent generalization across six major foundation model APIs.

Critical weaknesses
  1. Benchmark scale and technology node limitations: All evaluations are conducted on small 10x10 um clips derived from the open-source FreePDK45 node. In modern commercial semiconductor manufacturing, OPC recipe challenges predominantly arise in sub-7nm/sub-3nm advanced nodes involving complex multi-patterning, EUV, and curvilinear masks. On full-chip layouts, proximity interactions span wider neighborhoods, and simulation runtimes are orders of magnitude higher. It remains unverified whether the current prompt and retrieval mechanisms scale to full-chip contexts without suffering from context saturation or severe latency bottlenecks.

  2. Marginal statistical confidence in local-action continuation experiments: While the overall system gains are substantial, the paired continuation study evaluating local actions (Section 4.4.2 and Table A4) relies on only 5 cases per layer. On Metal1, the 95% bootstrap confidence interval for local action advantage spans [-1.465, 2.883] nm, crossing zero significantly and accompanied by increased mean EPE and PVB. This leaves the independent statistical contribution of local geometric actions less conclusive than that of global tuning.

Questions for authors
  1. Advanced Node and Full-Chip Feasibility: How do the authors envision scaling VPEvolve to modern advanced nodes (e.g., 7nm/3nm ArFi/EUV) or full-chip layouts? Specifically, if the layout window is expanded to handle long-range optical effects, how does the token cost and context management scale when multiple complex hotspots arise simultaneously?

  2. Variance in Local Actions on Metal1: In Table A4, the 95% bootstrap interval for the maximum EPE advantage of local actions on Metal1 spans [-1.465, 2.883] nm. Could the authors clarify why local actions exhibited such wide variance on Metal1 compared to Poly, and whether certain density strata systematically disfavor local rule insertion due to dense interconnect spacing?

  3. Cross-Layout Skill Bank Transferability: In the main experiments, the agent starts with an empty Skill Bank (E0 = empty) for each window. Did the authors investigate whether a curated Skill Bank developed on one window/stratum can be transferred zero-shot to accelerate optimization on unseen layout windows without positive/negative rule transfer conflicts?

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

VPEvolve proposes an LLM-based system for improving optical proximity correction (OPC) recipes used in chip manufacturing. It combines a Virtual Process Engineer harness with an evolving Skill Bank that records lessons from previous optimization trials. The authors report substantial improvements over several agent and memory baselines. Ablation studies show that the full pipeline achieves the best maximum-EPE results among the tested variants.

Critical strengths
  1. Relevant engineering problem. OPC recipe development can require costly tool evaluations and substantial expert interpretation.
  2. Integration with a commercial engineering tool. Candidate recipes are evaluated through Calibre, connecting the approach to an established OPC workflow.
  3. Adaptation without model retraining. The system improves recipes and updates external experience while keeping the underlying LLM fixed.
  4. Substantial improvement on Poly. The mean of the per-window maximum edge-placement errors falls from 18.294 to 5.361 nm, substantially below the evaluated baselines.
  5. Evaluation across multiple LLM families. Six additional models improve upon the initial recipe, suggesting that the workflow is not restricted to the main model.
Critical weaknesses
  1. The contribution of the proposed memory mechanism remains uncertain. Substantial improvement occurs on the first trial, before experience accumulates. The paired continuation experiments help isolate memory quality, but their confidence intervals include zero. The evidence therefore supports the complete workflow more strongly than the specific contribution of evolving memory.
  2. The contribution beyond close prior approaches needs stronger evidence. The actor–reflector–curator structure builds on ACE, while the closely related Intelligent OPC Engineer Assistant is cited but not experimentally compared. The methodological differences and their practical benefits could be established more clearly.
Questions for authors
  1. How much of the improvement can be attributed to the evolving Skill Bank beyond the initial harness, engineering guidance, and raw trial history, particularly given the substantial first-trial gain and uncertainty in the paired comparisons?
  2. Did the reported baselines receive the same engineering guidance as VPEvolve? How do the configurations used in the experiments relate to the released presets?
  3. What are the main methodological advances over ACE and Intelligent OPC Engineer Assistant, and what evidence demonstrates their benefits beyond domain adaptation and system integration?
Reviewer_pYprOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Paper summary. This paper addresses optical proximity correction through a Virtual Process Engineer (VPE) framework equipped with a skill bank of accumulated engineering experience. The framework connects recipe optimization with experience reuse: observations from each trial are distilled into skills that inform subsequent experiments. The goal is to accelerate process optimization and reduce reliance on manual intervention by human process engineers.

Critical strengths
  1. The paper includes well-crafted figures and an accompanying website that help communicate the proposed framework.

  2. The experimental comparisons include recent baselines.

Critical weaknesses
  1. The technical novelty and task-specific insights are insufficiently articulated. The contributions listed in lines 88–98 primarily describe the framework’s objectives and capabilities, rather than explaining the technical advances that enable them. For example, equipping a frozen LLM with process manuals, layout analysis, recipe editing, and commercial-tool evaluation describes the system’s components, but does not clarify the challenges involved in integrating them effectively. Similarly, using LLM reflection and curation to extract reusable skills leaves important methodological questions unanswered: How are useful skills identified, validated, scoped, and retrieved? How does the framework prevent unreliable or conflicting experience from misleading subsequent experiments? The paper would be stronger if it identified the main bottlenecks in automating optical proximity correction, distinguished routine steps from those requiring substantial engineering reasoning, and explained how the proposed design addresses these challenges. Ablations of the key design choices would further clarify their contribution.

  2. The evaluation scope is limited. The experiments rely solely on FreePDK45, leaving the method’s generalizability to larger and more complex settings unclear. Evaluation on additional datasets, more diverse layouts, or other process technologies would strengthen the evidence for the framework’s effectiveness and scalability. If such evaluations are infeasible, the paper should more clearly discuss the limitations of the current benchmark and the extent to which its findings are expected to generalize.

  3. The reliability of the results is difficult to assess. Line 297 states that each case is evaluated only once using a single seed. Because LLM-driven optimization is stochastic, individual runs may produce different decisions and outcomes. The authors should repeat the experiments with multiple random seeds and report aggregate performance together with variability, such as standard deviations or confidence intervals. This would help establish whether the reported improvements are consistent across runs.

Questions for authors

see above

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.