arxiv: 2604.18445 · v2 · submitted 2026-04-20 · 💻 cs.LG · cs.AR

Recognition: unknown

AutoPPA: Automated Circuit PPA Optimization via Contrastive Code-based Rule Library Learning

Chongxiao Li , Pengwei Jin , Di Huang , Guangrun Sun , Husheng Han , Jianan Mu , Xinyao Zheng , Jiaguo Zhu

show 9 more authors

Shuyi Xing Hanjun Wei Tianyun Ma Shuyao Cheng Rui Zhang Ying Wang Zidong Du Qi Guo Xing Hu

Authors on Pith no claims yet

Pith reviewed 2026-05-10 04:33 UTC · model grok-4.3

classification 💻 cs.LG cs.AR

keywords PPA optimizationRTL circuit designautomated rule abstractioncontrastive code learningadaptive multi-step searchLLM-based circuit optimizationExplore-Evaluate-Induce workflow

0 comments

The pith

AutoPPA automates PPA optimization in circuits by learning rules through contrasting generated code pairs rather than human summaries.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper develops AutoPPA to make performance, power, and area optimization of RTL circuit designs fully automated. It replaces reliance on human-summarized rules or blind search with a process that generates multiple code versions, evaluates them, and induces general rules by contrasting the pairs. The goal is to discover optimization patterns that improve circuit metrics more efficiently. An adaptive search then applies the most suitable rules to new designs. Experiments indicate this leads to better results than manual efforts or previous automated methods.

Core claim

The central discovery is that an Explore-Evaluate-Induce workflow can automatically abstract effective optimization rules from pairs of generated circuit code, and these rules, when applied via an adaptive multi-step search, yield superior PPA performance compared to manual optimization and prior methods.

What carries the argument

The Explore-Evaluate-Induce (E²I) workflow that contrasts generated code pairs to abstract rules, paired with an adaptive multi-step search framework for rule application.

Load-bearing premise

The rules derived automatically from contrasting code pairs will be more generalizable and effective than those summarized by humans, and the adaptive search framework will consistently identify and apply the optimal rules to any circuit.

What would settle it

Observing that for some input circuits, the PPA metrics after AutoPPA optimization are inferior to those from manual rule application or unoptimized designs.

Figures

Figures reproduced from arXiv: 2604.18445 by Chongxiao Li, Di Huang, Guangrun Sun, Hanjun Wei, Husheng Han, Jiaguo Zhu, Jianan Mu, Pengwei Jin, Qi Guo, Rui Zhang, Shuyao Cheng, Shuyi Xing, Tianyun Ma, Xing Hu, Xinyao Zheng, Ying Wang, Zidong Du.

**Figure 1.** Figure 1: Constructing the Rule Library for PPA optimization. Explore Cnon Copt1 1.Contrastive Code-based Rule Library Learning Copt2 ... Copt50 LLM Sampling EDA Synthesis Verification Tools Functional Equivalent Code Pairs module multiply_constant ( input[7:0] a, output[15:0] mul ); assign mul = {4'b0000,a,4'b0000}+ {8'h00,a}; endmodule Automatically Constructed Rule Library Copt1 equiv1 PPA1 Copt2 equiv2 PPA2 … … … view at source ↗

**Figure 2.** Figure 2: Overview of AutoPPA. AutoPPA includes the pipeline of the rule library learning workflow and the adaptive rule-based PPA optimization. multi-step search framework is a rule-enhanced beam search method that leverages the rule to better guide the LLM’s exploration of higher-quality Verilog code samples, increasing the probability of PPA-optimized implementations against Challenge 3. Experiments show that A… view at source ↗

**Figure 3.** Figure 3: Area and Delay improvement comparison with vanilla LLM sampling. AutoPPA yields consistently higher impr@k and better growth rate than DeepSeek-V3. search budget increases [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗

read the original abstract

Performance, power, and area (PPA) optimization is a fundamental task in RTL design, requiring a precise understanding of circuit functionality and the relationship between circuit structures and PPA metrics. Recent studies attempt to automate this process using LLMs, but neither feedback-based nor knowledge-based methods are efficient enough, as they either design without any prior knowledge or rely heavily on human-summarized optimization rules. In this paper, we propose AutoPPA, a fully automated PPA optimization framework. The key idea is to automatically generate optimization rules that enhance the search for optimal solutions. To do this, AutoPPA employs an Explore-Evaluate-Induce ($E^2I$) workflow that contrasts and abstracts rules from diverse generated code pairs rather than manually defined prior knowledge, yielding better optimization patterns. To make the abstracted rules more generalizable, AutoPPA employs an adaptive multi-step search framework that adopts the most effective rules for a given circuit. Experiments show that AutoPPA outperforms both the manual optimization and the state-of-the-art methods SymRTLO and RTLRewriter.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

AutoPPA's E²I contrastive rule induction from code pairs is a clear step past manual rules or blind LLM feedback for RTL PPA, but the abstract supplies no metrics or benchmarks to show the claimed outperformance or rule transfer actually happens.

read the letter

AutoPPA automates PPA optimization in RTL design by generating code pairs, contrasting them to induce rules, and then using an adaptive multi-step search to apply the best ones to a given circuit. The E²I workflow is the main new piece: it pulls patterns directly from the differences in generated versions instead of depending on human-summarized rules or starting from scratch with feedback loops. That addresses the inefficiency the abstract notes in both prior LLM styles. The adaptive search is a sensible addition to improve how well the rules fit different inputs. The paper does a clean job laying out why existing approaches fall short and sketching a concrete alternative that could cut down on manual rule writing. If the full experiments include ablations on the induction step and results on standard circuit suites, the idea could be useful for EDA tool builders. The soft spot is the evidence. The abstract says AutoPPA beats manual optimization plus SymRTLO and RTLRewriter, yet it gives no numbers, no list of benchmarks, no effect sizes, and no mention of cross-circuit transfer tests. Without those, the central assumption—that contrastively induced rules will be more general and effective than human ones—stays untested. The stress-test concern about rules embedding benchmark-specific biases or the search failing on arbitrary inputs still looks live based on what is shown. This is for people working on automated hardware design flows or LLM-driven code optimization. A reader who wants fresh angles on rule extraction from examples could pick up the E²I framing even if the results need more checking. It deserves a serious referee to examine the experimental section, the rule quality, and whether the outperformance holds with proper controls. I would send it to review rather than desk reject.

Referee Report

2 major / 1 minor

Summary. The paper proposes AutoPPA, a fully automated PPA optimization framework for RTL circuits. It introduces an Explore-Evaluate-Induce (E²I) workflow that automatically abstracts optimization rules by contrasting diverse generated code pairs, rather than relying on human-summarized rules, and combines this with an adaptive multi-step search framework to select and apply the most effective rules for a given circuit. The central claim is that this yields better optimization patterns and outperforms both manual optimization and state-of-the-art methods SymRTLO and RTLRewriter.

Significance. If the experimental claims hold with proper validation, AutoPPA could meaningfully advance automated electronic design automation by reducing dependence on manually curated rules and improving scalability of LLM-assisted circuit optimization. The contrastive rule induction and adaptive search ideas are potentially valuable contributions that address limitations in prior feedback-based and knowledge-based approaches.

major comments (2)

[Abstract] Abstract: The claim that 'Experiments show that AutoPPA outperforms both the manual optimization and the state-of-the-art methods SymRTLO and RTLRewriter' is presented without any quantitative metrics (e.g., PPA improvement percentages, number of benchmarks, statistical details, or comparison tables). This is load-bearing for the central claim and prevents assessment of whether the outperformance is real or generalizable.
[Method (E²I workflow and adaptive search)] E²I workflow description: No evidence or experiments are described demonstrating that rules abstracted via contrast from generated code pairs transfer to arbitrary unseen circuits (e.g., via cross-benchmark transfer tests, rule-coverage statistics, or handling of cases with no applicable rule). This directly impacts the weakest assumption that the induced rules are more generalizable than human-summarized ones and that the adaptive search reliably selects them.

minor comments (1)

[Abstract] The abstract introduces the E²I acronym before its expansion (Explore-Evaluate-Induce), which could be clarified on first use for readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their insightful comments, which have helped us improve the clarity and rigor of our presentation. We have revised the manuscript to address the concerns raised regarding the abstract and the validation of rule generalizability.

read point-by-point responses

Referee: [Abstract] Abstract: The claim that 'Experiments show that AutoPPA outperforms both the manual optimization and the state-of-the-art methods SymRTLO and RTLRewriter' is presented without any quantitative metrics (e.g., PPA improvement percentages, number of benchmarks, statistical details, or comparison tables). This is load-bearing for the central claim and prevents assessment of whether the outperformance is real or generalizable.

Authors: We acknowledge that the abstract, as currently written, does not include specific quantitative metrics, which makes it difficult to immediately assess the strength of the outperformance claim. The full paper contains comprehensive experimental results, including comparison tables and statistical details across a set of benchmarks. To address this, we will revise the abstract to incorporate key quantitative findings, such as the average PPA improvements and the number of benchmarks used, drawn from our experimental section. This will provide readers with a clearer view of the results without altering the manuscript's core content. revision: yes
Referee: [Method (E²I workflow and adaptive search)] E²I workflow description: No evidence or experiments are described demonstrating that rules abstracted via contrast from generated code pairs transfer to arbitrary unseen circuits (e.g., via cross-benchmark transfer tests, rule-coverage statistics, or handling of cases with no applicable rule). This directly impacts the weakest assumption that the induced rules are more generalizable than human-summarized ones and that the adaptive search reliably selects them.

Authors: The E²I workflow aims to induce general rules by contrasting diverse code pairs, and the adaptive search is intended to apply them to new circuits by selecting the most effective ones. However, we recognize that explicit demonstrations of transfer to completely unseen circuits, such as cross-benchmark tests or rule coverage statistics, are not detailed in the current manuscript. We will add experiments or analysis to the revised version, including rule applicability statistics across benchmarks and a description of the fallback mechanism in the adaptive search when no rule directly applies. This will strengthen the evidence for generalizability. revision: partial

Circularity Check

0 steps flagged

No circularity: rule induction from generated code pairs is independent of fitted inputs or self-citations

full rationale

The paper presents an algorithmic workflow (Explore-Evaluate-Induce) that generates code pairs, contrasts them to abstract rules, and applies the rules via adaptive search. No equations, fitted parameters, or predictions that reduce by construction to the inputs appear in the description. The central claim rests on experimental outperformance against external baselines (SymRTLO, RTLRewriter, manual optimization) rather than any self-referential derivation or load-bearing self-citation. The method is explicitly positioned as replacing human-summarized rules with automatically induced ones, with no renaming of known results or smuggling of ansatzes via prior author work.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the unstated premise that LLM-generated code pairs contain extractable, generalizable optimization patterns superior to human rules; no free parameters, invented entities, or additional axioms are specified in the abstract.

axioms (1)

domain assumption LLMs can reliably generate diverse, functionally equivalent code pairs that differ meaningfully in PPA characteristics
The Explore-Evaluate-Induce workflow depends on this to produce contrastive examples from which rules are abstracted.

pith-pipeline@v0.9.0 · 5545 in / 1091 out tokens · 46904 ms · 2026-05-10T04:33:29.008695+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

31 extracted references · 14 canonical work pages · 4 internal anchors

[1]

Information Processing & Management39(1), 45–65 (2003)

Aizawa, A.: An information-theoretic perspective of tf–idf measures. Information Processing & Management39(1), 45–65 (2003)

2003
[2]

In: Proceedings of Government Microcircuit Applica- tions and Critical Technology Conference (2019)

Ajayi, T., Blaauw, D.: Openroad: Toward a self-driving, open-source digital layout implementation tool chain. In: Proceedings of Government Microcircuit Applica- tions and Critical Technology Conference (2019)

2019
[3]

In: 2023 ACM/IEEE 5th Workshop on Machine Learning for CAD (MLCAD)

Blocklove, J., Garg, S., Karri, R., Pearce, H.: Chip-chat: Challenges and opportu- nities in conversational hardware design. In: 2023 ACM/IEEE 5th Workshop on Machine Learning for CAD (MLCAD). pp. 1–6. IEEE (2023)

2023
[4]

Chang, K., Wang, Y., Ren, H., Wang, M., Liang, S., Han, Y., Li, H., Li, X.: Chipgpt: How far are we from natural language hardware design (2025), https://arxiv.org/abs/2305.14019

work page arXiv 2025
[5]

Chen, M., Tworek, J., Jun, H., Yuan, Q., de Oliveira Pinto, H.P., Ka- plan, J., et al.: Evaluating large language models trained on code (2021), https://arxiv.org/abs/2107.03374

work page internal anchor Pith review Pith/arXiv arXiv 2021
[6]

In: Proceedings of the 43rd IEEE/ACM International Conference on Computer-Aided Design

Cui, F., Yin, C., Zhou, K., Xiao, Y., Sun, G., Xu, Q., Guo, Q., Liang, Y., Zhang, X., Song, D., et al.: Origen: Enhancing rtl code generation with code-to-code aug- mentation and self-reflection. In: Proceedings of the 43rd IEEE/ACM International Conference on Computer-Aided Design. pp. 1–9 (2024)

2024
[7]

DeepSeek-AI, Liu, A., Feng, B., Xue, B., Wang, B., Wu, B., et al.: Deepseek-v3 technical report (2025), https://arxiv.org/abs/2412.19437

work page internal anchor Pith review Pith/arXiv arXiv 2025
[8]

In: Proceedings of the AAAI Conference on Artificial Intelligence

Ho, C.T., Ren, H., Khailany, B.: Verilogcoder: Autonomous verilog coding agents with graph-based planning and abstract syntax tree (ast)-based waveform tracing tool. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 39, pp. 300–307 (2025)

2025
[9]

ACMTransactionsonDesignAutomationofElectronicSystems(TODAES)26(5), 1–46 (2021)

Huang, G., Hu, J., He, Y., Liu, J., Ma, M., Shen, Z., Wu, J., Xu, Y., Zhang, H., Zhong, K., et al.: Machine learning for electronic design automation: A survey. ACMTransactionsonDesignAutomationofElectronicSystems(TODAES)26(5), 1–46 (2021)

2021
[10]

Hui, B., Yang, J., Cui, Z., Yang, J., Liu, D., Zhang, L., Liu, T., Zhang, J., Yu, B., Lu, K., Dang, K., Fan, Y., Zhang, Y., Yang, A., Men, R., Huang, F., Zheng, B., Miao, Y., Quan, S., Feng, Y., Ren, X., Ren, X., Zhou, J., Lin, J.: Qwen2.5-coder technical report (2024), https://arxiv.org/abs/2409.12186

work page internal anchor Pith review arXiv 2024
[11]

In: Pro- ceedings of the ACM SIGPLAN 1994 Conference on Programming Language Design and Implementation

Knoop, J., Rüthing, O., Steffen, B.: Partial dead code elimination. In: Pro- ceedings of the ACM SIGPLAN 1994 Conference on Programming Language Design and Implementation. p. 147–158. PLDI ’94, Association for Computing Machinery, New York, NY, USA (1994). https://doi.org/10.1145/178243.178256, https://doi.org/10.1145/178243.178256

work page doi:10.1145/178243.178256 1994
[12]

Li, Z., Zhang, X., Zhang, Y., Long, D., Xie, P., Zhang, M.: To- wards general text embeddings with multi-stage contrastive learning (2023), https://arxiv.org/abs/2308.03281

work page internal anchor Pith review arXiv 2023
[13]

In: 2024 IEEE LLM Aided Design Workshop (LAD)

Liu, S., Fang, W., Lu, Y., Zhang, Q., Zhang, H., Xie, Z.: Rtlcoder: Outperforming gpt-3.5 in design rtl generation with our open-source dataset and lightweight solu- tion. In: 2024 IEEE LLM Aided Design Workshop (LAD). pp. 1–5. IEEE (2024)

2024
[14]

In: Proceed- ings of the 43rd IEEE/ACM International Conference on Computer- Aided Design

Liu, S., Lu, Y., Fang, W., Li, M., Xie, Z.: Openllm-rtl: Open dataset and benchmark for llm-aided design rtl generation. In: Proceed- ings of the 43rd IEEE/ACM International Conference on Computer- Aided Design. ICCAD ’24, Association for Computing Machinery, New York, NY, USA (2025). https://doi.org/10.1145/3676536.3697118, https://doi.org/10.1145/367653...

work page doi:10.1145/3676536.3697118 2025
[15]

In: Proceedings of the 59th ACM/IEEE Design Automation Con- ference

Olofsson, A., Ransohoff, W., Moroze, N.: A distributed approach to silicon compi- lation: Invited. In: Proceedings of the 59th ACM/IEEE Design Automation Con- ference. p. 1343–1346 (2022)

2022
[16]

ACM Trans

Pan, J., Zhou, G., Chang, C.C., Jacobson, I., Hu, J., Chen, Y.: A survey of re- search in large language models for electronic design automation. ACM Trans. Des. Autom. Electron. Syst.30(3) (Feb 2025). https://doi.org/10.1145/3715324, https://doi.org/10.1145/3715324

work page doi:10.1145/3715324 2025
[17]

IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems18(1), 58–68 (1999)

Pasko, R., Schaumont, P., Derudder, V., Vernalde, S., Durackova, D.: A new algorithm for elimination of common subexpressions. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems18(1), 58–68 (1999). https://doi.org/10.1109/43.739059

work page doi:10.1109/43.739059 1999
[18]

Betterv: Controlled verilog generation with discriminative guidance,

Pei, Z., Zhen, H.L., Yuan, M., Huang, Y., Yu, B.: Betterv: Controlled verilog generation with discriminative guidance. arXiv preprint arXiv:2402.03375 (2024)

work page arXiv 2024
[19]

ACM Trans

Pinckney, N., Batten, C., Liu, M., Ren, H., Khailany, B.: Revisiting verilo- geval: A year of improvements in large-language models for hardware code generation. ACM Trans. Des. Autom. Electron. Syst.30(6) (Oct 2025). https://doi.org/10.1145/3718088, https://doi.org/10.1145/3718088

work page doi:10.1145/3718088 2025
[20]

Advances in Neural Information Processing Systems36(2024)

Shinn, N., Cassano, F., Gopinath, A., Narasimhan, K., Yao, S.: Reflexion: Lan- guage agents with verbal reinforcement learning. Advances in Neural Information Processing Systems36(2024)

2024
[21]

In: 2007 IEEE international conference on Microelec- tronic Systems Education (MSE’07)

Stine, J.E., Castellanos, I., Wood, M., Henson, J., Love, F., Davis, W.R., Fran- zon, P.D., Bucher, M., Basavarajaiah, S., Oh, J., et al.: Freepdk: An open-source variation-aware design kit. In: 2007 IEEE international conference on Microelec- tronic Systems Education (MSE’07). pp. 173–174. IEEE (2007)

2007
[22]

Synopsys, I.: Coding guidelines for datapath synthesis. Tech. rep., Synopsys, Inc. (June 2012), https://solvnet.synopsys.com/retrieve/print/015771.html, doc Id: 015771, Product: Design Compiler, Last Modified: 06/01/2012

2012
[23]

ACM Transactions on Design Automation of Electronic Systems29(3), 1–31 (2024)

Thakur, S., Ahmad, B., Pearce, H., Tan, B., Dolan-Gavitt, B., Karri, R., Garg, S.: Verigen: A large language model for verilog code generation. ACM Transactions on Design Automation of Electronic Systems29(3), 1–31 (2024)

2024
[24]

Thorat, K., Zhao, J., Liu, Y., Peng, H., Xie, X., Lei, B., Zhang, J., Ding, C.: Advanced large language model (llm)-driven verilog development: En- hancing power, performance, and area optimization in code synthesis (2024), https://arxiv.org/abs/2312.01022

work page arXiv 2024
[25]

In: Proceedings of the 61st ACM/IEEE Design Automation Conference

Tsai, Y., Liu, M., Ren, H.: Rtlfixer: Automatically fixing rtl syntax errors with large language model. In: Proceedings of the 61st ACM/IEEE Design Automation Conference. pp. 1–6 (2024)

2024
[26]

In: The Thirty-ninth Annual Conference on Neural Information Processing Systems

Wang, Y., Ye, W., Guo, P., He, Y., Wang, Z., Tian, B., He, S., Sun, G., Shen, Z., Chen, S., et al.: Symrtlo: Enhancing rtl code optimization with llms and neuron- inspired symbolic reasoning. In: The Thirty-ninth Annual Conference on Neural Information Processing Systems
[27]

https://yosyshq.net/yosys/ (2013)

Wolf, C.: Yosys open synthesis suite. https://yosyshq.net/yosys/ (2013)

2013
[28]

Yang, Y., Teng, F., Liu, P., Qi, M., Lv, C., Li, J., Zhang, X., He, Z.: Haven: Hallucination-mitigated llm for verilog code generation aligned with hdl engineers (2025), https://arxiv.org/abs/2501.04908

work page arXiv 2025
[29]

In: Proceedings of the 43rd IEEE/ACM International Conference on Computer-Aided Design

Yao, X., Wang, Y., Li, X., Lian, Y., Chen, R., Chen, L., Yuan, M., Xu, H., Yu, B.: Rtlrewriter: Methodologies for large models aided rtl code optimization. In: Proceedings of the 43rd IEEE/ACM International Conference on Computer-Aided Design. pp. 1–7 (2024) 16 C. Li et al

2024
[30]

IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems pp

Zhao, Y., Huang, D., Li, C., Jin, P., Song, M., Xu, Y., Nan, Z., Gao, M., Ma, T., Qi, L., Pan, Y., Zhang, Z., Zhang, R., Zhang, X., Du, Z., Guo, Q., Hu, X.: Codev: Empowering llms with hdl generation through multi-level summarization. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems pp. 1–1 (2025). https://doi.org/10.1109/TCAD...

work page doi:10.1109/tcad.2025.3604320 2025
[31]

Zhu, Y., Huang, D., Lyu, H., Zhang, X., Li, C., Shi, W., Wu, Y., Mu, J., Wang, J., zhao, Y., Jin, P., Cheng, S., shengwen Liang, Zhang, X., Zhang, R., Du, Z., Guo, Q.,Hu,X.,Chen,Y.:Qimeng-codev-r1:Reasoning-enhancedveriloggeneration.In: The Thirty-ninth Annual Conference on Neural Information Processing Systems (2025), https://openreview.net/forum?id=ly5DnRIgCZ

2025