Pith. sign in

REVIEW 4 major objections 5 minor 46 references

RF-Agent: A Practical Framework for Building Language Agents for RFIC Design

T0 review · 4 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that a multi-agent pipeline distilling seven RF textbooks into over 11,000 question–thinking–solution–answer samples produces the first RF-domain reasoning benchmark, and that both fine-tuning on this data and semantic retr

desk verdict Useful new RF QA dataset and benchmark, but the evaluation is not independent of the data-construction pipeline, so treat the accuracy numbers as upper bounds until leakage is addressed. read the letter →

arxiv 2607.18772 v1 pith:VNAMYTAT submitted 2026-07-21 cs.CL

Yueqi Xing , Houbo He , Jolie Wang , Erin Ni , Shikai Wang , Qiufeng Li , Weidong Cao , Taiyun Chi This is my paper · ORCID
classification cs.CL
keywords RFcircuitdesignlargelanguagemodelsreasoningdatasetmultiple-choicebenchmarktextbookdistillationsupervisedfine-tuningretrieval-augmentedgenerationsemanticretrieval
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

RF circuit design lacks open datasets and benchmarks for evaluating language models, so RF-Agent sets out to build one from scratch. The authors argue that textbook-driven knowledge distillation — a multi-agent Question–Thinking–Solution–Answer pipeline run over seven canonical RF textbooks — can generate a large, reusable reasoning dataset and a standardized multiple-choice benchmark. On that benchmark they show that domain-specific supervised fine-tuning consistently improves RF reasoning across model families, with the largest gains for small and base models, and that semantic retrieval beats keyword and hybrid retrieval for grounding state-of-the-art LLMs. If right, the framework provides the first open RF reasoning dataset and a route for adapting general-purpose models to a specialized engineering domain without expert-labeled data.

What carries the argument

The load-bearing mechanism is the multi-agent QTSA distillation pipeline: a Question Agent generates five-perspective multiple-choice (mcQTSA) and open-ended (ndQTSA) questions from each textbook subsection, an Answer Agent produces structured Thinking–Solution–Answer chain-of-thought traces, and a Process Agent validates and normalizes the outputs. This pipeline converts 1,108 subsections into the training data and the mcQTSA benchmark; the SFT and RAG experiments then measure what that data can teach.

What would settle it

Have two or more RF design engineers independently check the factual correctness of a random sample of roughly 100 mcQTSA questions and answers against the source textbooks, and separately run each benchmark question as a retrieval query against the same seven-textbook corpus: if a sizable fraction of answers are wrong or if the top-3 retrieved chunks contain the verbatim originating subsection for most questions, the reported SFT and RAG gains would be substantially inflated by benchmark leakage or error.

Watch

Extended reading notes

Core claim

RF-Agent establishes that a multi-agent QTSA distillation pipeline can convert subsection-level textbook content from seven canonical RF texts into a first-of-its-kind RF-domain reasoning dataset of over 11,000 samples, including a 1,000-question multiple-choice benchmark (mcQTSA). On this benchmark, the paper demonstrates two things: (1) supervised fine-tuning on the QTSA data consistently improves RF multiple-choice accuracy across Llama and Qwen model families (0.6B–4B), with dramatic gains on small base models (e.g., Llama3.2-3B-Base rising from 37.1% to 70.5%) and complementary benefits from mixing multiple-choice and open-response training formats; and (2) retrieval-augmented generatio

Load-bearing premise

The benchmark scores reflect genuine RF reasoning only if the LLM-generated questions and answers distilled from the seven textbooks are factually correct and cannot be answered by simply retrieving the exact textbook passage they were generated from—neither of which expert review confirmed.

Editorial extensions

If this is right

  • The released RF-domain QTSA dataset and mcQTSA benchmark give the field a standardized, reusable resource for evaluating and training LLMs on RF circuit reasoning, previously absent.
  • Domain-specific SFT on QTSA data yields consistent accuracy gains across model families, enabling small base models like Llama3.2-3B to improve by over 33 percentage points and approach the performance of much larger instruction-tuned models.
  • Semantic retrieval over an RF knowledge base gives larger accuracy improvements for frontier LLMs than keyword or hybrid retrieval, indicating that embedding-based context alignment fits RF reasoning better than lexical or fused retrieval.
  • Mixed-format training (mcQTSA plus ndQTSA) outperforms either format alone, showing that multiple-choice selection and open-ended explanatory reasoning provide complementary supervision signals.
  • Fine-tuning thinking-mode models both improves accuracy and cuts inference latency, suggesting domain supervision streamlines the internal reasoning process rather than merely adding output length.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the benchmark is valid and non-leaky, the same QTSA distillation recipe could bootstrap reasoning datasets for other expert domains with authoritative textbooks (analog, power, RF test), enabling rapid domain adaptation without expert labeling.
  • The semantic-over-hybrid finding hints that in technical domains, retrieval precision matters more than recall, and that injecting lower-rank lexical candidates actively degrades answers; a semantic-only top-5 or top-10 configuration might test this directly.
  • Because fine-tuning gains diminish below 2.5% for a 4B model while retrieval still adds ~3-4 points, the two strategies could be combined: SFT for parametric knowledge plus semantic RAG for verification, which the paper does not test jointly.
  • A natural stress test: generate QTSA questions that require multi-step derivations (e.g., S-parameter cascades, matching-network synthesis) and see whether fine-tuned small models can reproduce the derivation in free-form, not just choose among options.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper introduces RF-Agent, a framework that uses an LLM multi-agent pipeline (QTSA) to distill seven RF textbooks into over 11,000 reasoning samples, including a 1,000-question multiple-choice benchmark (mcQTSA). It then studies two adaptation strategies: supervised fine-tuning (SFT) on the distilled data and retrieval-augmented generation (RAG) over a knowledge base built from the same seven textbooks plus 950 papers, with semantic, keyword, and hybrid retrieval configurations. The authors report that domain-specific SFT substantially improves RF multiple-choice accuracy across several small and medium-sized LLMs, and that semantic retrieval outperforms keyword and hybrid RAG on state-of-the-art models. The dataset and code are publicly released.

Significance. If the benchmark is valid and independent, this is a useful contribution: it provides the first RF-domain reasoning dataset and benchmark, a systematic comparison of SFT and RAG strategies, and a reusable pipeline. The paper also deserves credit for releasing code/data, for evaluating multiple model families, and for designing a hit-and-miss control experiment. However, the central empirical claims currently rest on a benchmark whose gold answers are generated by the same type of model pipeline as the training data and whose source textbooks overlap with the RAG knowledge base. This creates a real risk that the reported SFT and RAG gains are inflated by source leakage rather than by general RF reasoning ability. The contribution is significant if these threats can be addressed, but the manuscript as written does not yet establish that.

major comments (4)
  1. [§II-A, §III-B1, §IV-B, App. B] The benchmark is not independent of the resources used to train and ground the models. The mcQTSA questions are generated from 1,108 subsections of seven textbooks (§II-A, §II-B1), and the RAG knowledge base contains the same seven textbooks (§III-B1). The hit-and-miss example in Fig. 4 shows the semantic hit retrieving the exact textbook derivation that 'directly supports option A.' Thus the Table II RAG gains may reflect retrieval of the generation source rather than improved RF reasoning. The SFT results (Table I) are similarly exposed, since the QTSA training data is distilled from the same subsections. The manual review in §II-B1 checked only diversity, not factual correctness or source independence. Please quantify leakage (e.g., fraction of benchmark questions whose generating subsection is retrieved), run RAG with the seven textbooks excluded from the knowledge base, and evaluate
  2. [§II-B] All gold answers and CoT traces are produced by GPT-4.1-mini and GPT-5-mini without expert verification. The paper calls the benchmark 'grounded in expert RF knowledge' (§II-C), but no mechanism ensures the generated answers are correct; the review in §II-B1 only verifies question diversity. Since every accuracy number and every SFT/RAG comparison in Tables I–III is computed against these labels, an unknown error rate in the gold answers directly biases all claims. I recommend reporting a human-expert review of a random sample of benchmark questions and their gold answers, with agreement statistics and an error analysis, and assessing whether the qualitative conclusions are robust to plausible label noise.
  3. [§IV-B, Tables II–III] The RAG ranking claim rests on small differences without uncertainty quantification. In Table II the gap between semantic and keyword is 0.9–1.1 points (e.g., GPT-4o: 93.0 vs 92.1) and between keyword and hybrid is 0.1–1.4 points; with 1,000 questions a 1-point difference is 10 items and may be within sampling variability. The hit-and-miss experiment (Table III) is run on only 100 questions. No confidence intervals or significance tests are reported. To support 'semantic retrieval performs best' (§IV-B3), please report bootstrap intervals or a paired test (e.g., McNemar) on the 1,000-item benchmark.
  4. [§IV, §IV-A1] The benchmark split is underspecified. Experiments are run on a 'fixed 1,000-sample mcQTSA benchmark' (§IV), and SFT training uses the 'non-benchmark partition of mcQTSA' (§IV-A1), but the paper does not state how the 1,000 samples were selected, whether the partition is random, or whether overlap between training and benchmark questions was checked (e.g., near-duplicate questions from the same subsection). This matters for the SFT results in Table I, because near-duplicate leakage would inflate fine-tuning gains. Please describe the split procedure and report a duplicate/near-duplicate analysis between the training partition and the 1,000 benchmark questions.
minor comments (5)
  1. [§II-A] The seven textbooks are used throughout but never listed. Please provide full bibliographic entries in the paper or appendix.
  2. [§II-B1] Minor wording: 'It is motivated by findings...' reads awkwardly; also 'adaption' in the introduction should be 'adaptation.'
  3. [§IV-A1] The construction of the 'randomly mixed' 5M-token subset (ratio of mcQTSA to ndQTSA, random seed) is not described.
  4. [§IV] Table I reports wall-clock inference times, but hardware/API details are not given. Specify the evaluation environment for reproducibility.
  5. [§IV-B3] The explanation for hybrid underperformance is plausible but not verified, e.g., by ablating the re-ranker or measuring retrieved chunk precision per configuration. Consider adding retrieval-quality metrics (hit rate, MRR) to support the qualitative explanation.

Circularity Check

2 steps flagged · score 6.0 of 10

RAG and SFT evaluations are not independent of the QTSA generation pipeline: the benchmark, training data, and RAG knowledge base all derive from the same seven textbooks, so reported gains may reflect source-passage retrieval or memorization rather than general RF reasoning.

  1. self definitional [Section II-A, II-B1, III-B1, Appendix B (Fig. 4)]
    "we ground our dataset construction in seven canonical RF textbooks ... Each book is parsed at the subsection level into self-contained conceptual units, yielding a corpus of 1,108 subsections. ... The Question Agent (GPT-4.1-mini) generates mcQTSA and ndQTSA questions strictly grounded in the provided subsection. ... Our retrieval corpus consists of 950 RF-related peer-reviewed papers and 7 canonical RF textbooks, forming a dense RF knowledge base. ... For one benchmark question, the hit condition uses the top-3 retrieved chunks and the miss condition uses ranks 4–6 from the same run: the hit"

    RAG accuracy is measured on mcQTSA questions generated strictly from the seven textbooks, and the retrieval corpus contains those same seven textbooks. A semantic top-3 hit can return the exact source subsection from which a question was generated; Appendix B's own example shows the hit chunk containing the derivation that directly supports the correct option. Thus the claim that 'semantic retrieval performs best' is confounded by construction: the evaluation target is definitionally tied to the retrieval source, so high hit accuracy may only mean the embedder finds the original passage, not that RAG improves general RF reasoning.

  2. other [Section III-A1, IV-A1, IV (benchmark setup)]
    "We train directly on the QTSA samples from Section II. ... all subsequent fine-tuning experiments adopt a mixed-format training set of 17M tokens, comprising the full ndQTSA corpus and the non-benchmark partition of mcQTSA. ... All models are evaluated on a fixed 1,000-sample mcQTSA benchmark with deterministic greedy decoding."

    The SFT training corpus and the 1,000-sample benchmark are both outputs of the same three-agent QTSA pipeline applied to the same 1,108 textbook subsections; the only separation reported is a 'non-benchmark partition' of mcQTSA, with no subsection-level disjointness guarantee. Consequently, fine-tuning could memorize the source-subsection-to-question mapping rather than acquire transferable RF reasoning, so the reported SFT gains over the base models are not demonstrated to measure general domain adaptation independent of the dataset-construction pipeline.

full rationale

No load-bearing self-citation chain or imported uniqueness theorem appears: the paper cites prior QTSA work [19] and retrieval work [17,18] as inspiration, but the central claims rest on the new QTSA pipeline rather than on an unverified self-citation. The main circularity is benchmark dependence: the mcQTSA benchmark is generated from the same seven textbooks that form the RAG knowledge base and that seed the QTSA training data. The RAG result in particular reduces, at least in part, to retrieving the exact source passages used to create the questions; the Appendix B hit example demonstrates this mechanism. The paper's manual review checked only question diversity, not factual correctness or source independence, and the conclusion itself concedes the benchmark 'may limit discrimination among stronger models.' The SFT results are exposed similarly because train and test share the generation pipeline and source corpus, with no reported disjoint split by subsection. These issues do not make every number meaningless—absolute accuracies and cross-model comparisons could still be informative if the auto-generated gold answers are correct—but the advertised conclusions about 'RF reasoning' improvement are partially circular relative to the construction inputs. Score 6 reflects one central claim (RAG) reducing by construction and a second central claim (SFT) sharing the same non-independent evaluation.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

No new physical or conceptual entities are introduced. The free parameters are retrieval depths, training-token budget, and benchmark subset size, all chosen by the authors without full ablations. The key unstated assumption is that the auto-generated, self-referential benchmark and retrieval corpus measure genuine RF reasoning.

free parameters (4)
  • Top-3 retrieved chunks (semantic/keyword/hybrid) = 3
    Retrieval depth is fixed at three for all RAG configurations without an ablation justifying that choice; it directly affects the reported RAG accuracy.
  • Hybrid candidate pool size (15 per method) = 15
    Hybrid RAG merges top-15 candidates from each retrieval method before re-ranking; this pool size is a design choice not derived from data or theory.
  • 17M-token mixed SFT training set = 17M tokens
    The final training set size was chosen after a scaling study showed saturation near 9M tokens; the exact 17M budget is a tuning decision, not a derived quantity.
  • 1000-sample benchmark subset = 1000
    A fixed 1,000-sample mcQTSA evaluation set is used, but the paper does not specify how this subset was selected or stratified, making it a free choice that affects all reported numbers.
assumptions (5)
  • domain assumption LLM-generated QTSA answer traces are factually correct
    The Answer Agent (GPT-5-mini) produces Thinking/Solution/Answer fields without expert verification; the manual review in Sec. II-B1 only checked question diversity, not factual correctness.
  • domain assumption Multiple-choice accuracy on mcQTSA measures RF reasoning
    The benchmark is treated as a proxy for RF reasoning ability, but no external validation shows that answering these auto-generated questions correlates with actual RF design competence.
  • domain assumption The RAG knowledge base and the benchmark are independent
    Both derive from the same seven canonical textbooks (Sec. II-A vs Sec. III-B1), so retrieval can return the exact source passage used to generate a given question, undermining the independence assumption.
  • domain assumption CoT distillation from teacher LLMs transfers reasoning
    The paper relies on prior findings [34], [35] that chain-of-thought distillation improves downstream performance; no within-paper evidence separates this effect from benchmark-specific overfitting.
  • domain assumption Textbook-quality data is sufficient for domain adaptation
    The choice of seven textbooks is motivated by [38] ('Textbooks Are All You Need'), but the paper does not demonstrate that this corpus is representative of the breadth of RF reasoning.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RF-Agent: A Practical Framework for Building Language Agents for RFIC Design." pith.science (2026). https://pith.science/paper/VNAMYTAT

@misc{pith2026260718772,
  author       = {Pith},
  title        = {Pith review of: RF-Agent: A Practical Framework for Building Language Agents for RFIC Design},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VNAMYTAT}},
  note         = {Machine review of arXiv:2607.18772}
}
read the original abstract

Large language models (LLMs) have driven rapid progress in electronic design automation (EDA), yet their application to radio-frequency (RF) circuit design remains limited by the scarcity of domain-specific datasets and standardized benchmarks. We present RF-Agent, which addresses this gap through textbook-driven knowledge distillation. A multi-agent Question-Thinking-Solution-Answer (QTSA) pipeline converts a subsection-level corpus from seven canonical RF textbooks into the first-of-its-kind RF-domain reasoning dataset (over 11,000 samples) with a dedicated multiple-choice benchmark. On this benchmark we study two adaptation strategies: supervised fine-tuning (SFT) and three retrieval-augmented generation (RAG) configurations (semantic, keyword, hybrid). Across multiple LLM families, domain-specific SFT significantly improves RF reasoning, especially for small and medium-sized models; among RAG configurations, semantic retrieval performs best, indicating embedding-based context alignment suits RF reasoning better than naive fusion. The dataset and benchmark provide a reusable foundation for future work on LLM-aided RF circuit design.

Figures

Figures reproduced from arXiv: 2607.18772 by the authors.

Figure 1
Figure 1. Multi-agent QTSA distillation pipeline. The Question Agent generates perspective-diverse mcQTSA and ndQTSA questions from each original [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Hybrid RAG pipeline for RF knowledge retrieval. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. A representative mcQTSA sample for mixers. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: A representative hit-and-miss example. The hit chunk (top) retrieves [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 7 linked inside Pith

  1. [1]

    AnalogGenie: A generative engine for automatic discovery of analog circuit topologies,

    J. Gao, W. Cao, J. Yang, and X. Zhang, “AnalogGenie: A generative engine for automatic discovery of analog circuit topologies,” in Proceedings of the 13th International Conference on Learning Representations (ICLR), 2025. [Online]. Available: https://openreview. net/forum?id=jCPak79Kev

    2025

  2. [2]

    LaMAGIC: Language-model-based topology generation for analog integrated circuits,

    C.-C. Chang, Y . Shen, S. Fan, J. Li, S. Zhang, N. Cao, Y . Chen, and X. Zhang, “LaMAGIC: Language-model-based topology generation for analog integrated circuits,” inProceedings of the 41st International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, vol. 235. PMLR, 2024. [Online]. Available: https://dl.acm.org/doi/10.5555/...

    arXiv 2024

  3. [3]

    AnalogCoder: Analog Circuit Design via Training-Free Code Genera- tion,

    Y . Lai, S. Lee, G. Chen, S. Poddar, M. Hu, D. Z. Pan, and P. Luo, “AnalogCoder: Analog Circuit Design via Training-Free Code Genera- tion,” inProc. AAAI Conference on Artificial Intelligence, vol. 39, no. 1, 2025, pp. 379–387

    2025

  4. [4]

    AnalogXpert: Automating Analog Topology Synthesis by Incorporating Circuit Design Expertise into Large Language Models,

    H. Zhang, S. Sun, Y . Lin, R. Wang, and J. Bian, “AnalogXpert: Automating Analog Topology Synthesis by Incorporating Circuit Design Expertise into Large Language Models,” in2025 International Sympo- sium of Electronics Design Automation (ISEDA). IEEE, 2025

    2025

  5. [5]

    Invited Paper: Multi-Agent Generative Synthesis for Ana- log/RF Circuit: from Scalable Topology Generation to Efficient Inverse Design,

    S. Wang, Q. Li, H. He, J. Gao, Z. Wang, Y . Sun, X. Zhang, T. Chi, and W. Cao, “Invited Paper: Multi-Agent Generative Synthesis for Ana- log/RF Circuit: from Scalable Topology Generation to Efficient Inverse Design,” in2025 IEEE/ACM International Conference On Computer Aided Design (ICCAD), 2025, pp. 1–9

    2025

  6. [6]

    ADO-LLM: Analog Design Bayesian Optimization with In-Context Learning of Large Language Models,

    Y . Yin, Y . Wang, B. Xu, and P. Li, “ADO-LLM: Analog Design Bayesian Optimization with In-Context Learning of Large Language Models,” inProceedings of the 43rd IEEE/ACM International Conference on Computer-Aided Design (ICCAD). ACM, 2024. [Online]. Available: https://doi.org/10.1145/3676536.3676816

    arXiv 2024

  7. [7]

    LEDRO: LLM-Enhanced Design Space Reduction and Optimization for Analog Circuits,

    D. V . Kochar, H. Wang, A. P. Chandrakasan, and X. Zhang, “LEDRO: LLM-Enhanced Design Space Reduction and Optimization for Analog Circuits,” in2025 IEEE International Conference on LLM-Aided Design (ICLAD). IEEE, 2025, pp. 141–148

    2025

  8. [8]

    A Large Language Model-based Multi-Agent Framework for Analog Circuits’ Sizing Relationships Extraction,

    C. Liu, W. Chen, H. Xu, Y . Du, J. Yang, and L. Du, “A Large Language Model-based Multi-Agent Framework for Analog Circuits’ Sizing Relationships Extraction,” in2025 International Symposium of Electronics Design Automation (ISEDA), 2025, pp. 181–187

    2025

Show all 46 references
  1. [9]

    TopoSizing: An LLM-aided framework of topology-based understanding and sizing for AMS circuits,

    Z. Wei, Z. Kong, Y . Wang, D. Z. Pan, and X. Tang, “TopoSizing: An LLM-aided framework of topology-based understanding and sizing for AMS circuits,”arXiv preprint arXiv:2509.14169, 2025. [Online]. Available: https://arxiv.org/abs/2509.14169

    2025

  2. [10]

    RTLCoder: Fully Open-Source and Efficient LLM-Assisted RTL Code Generation Technique,

    S. Liu, W. Fang, Y . Lu, J. Wang, Q. Zhang, H. Zhang, and Z. Xie, “RTLCoder: Fully Open-Source and Efficient LLM-Assisted RTL Code Generation Technique,”IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 2024

    2024

  3. [11]

    VeriGen: A Large Language Model for Verilog Code Generation,

    S. Thakur, B. Ahmad, H. Pearce, B. Tan, B. Dolan-Gavitt, R. Karri, and S. Garg, “VeriGen: A Large Language Model for Verilog Code Generation,”ACM Trans. Des. Autom. Electron. Syst., vol. 29, no. 3, Apr. 2024. [Online]. Available: https://doi.org/10.1145/3643681

    2024 doi

  4. [12]

    ChipNeMo: Domain-Adapted LLMs for Chip Design,

    M. Liu, T.-D. Ene, R. Kirbyet al., “ChipNeMo: Domain-Adapted LLMs for Chip Design,”arXiv preprint arXiv:2311.00176, 2023

    2023 arXiv

  5. [13]

    Analog Design Experiments With AI—Part 1 [The Analog Mind],

    B. Razavi, “Analog Design Experiments With AI—Part 1 [The Analog Mind],”IEEE Solid-State Circuits Magazine, vol. 17, no. 4, pp. 11–15, 2025

    2025

  6. [14]

    Analog Design Experiments With AI—Part 2 [The Analog Mind],

    B. Razavi, “Analog Design Experiments With AI—Part 2 [The Analog Mind],”IEEE Solid-State Circuits Magazine, vol. 18, no. 2, pp. 8–13, 2026

    2026

  7. [15]

    Invited Paper: Verilo- gEval: Evaluating Large Language Models for Verilog Code Genera- tion,

    M. Liu, N. Pinckney, B. Khailany, and H. Ren, “Invited Paper: Verilo- gEval: Evaluating Large Language Models for Verilog Code Genera- tion,” in2023 IEEE/ACM International Conference on Computer Aided Design (ICCAD), 2023, pp. 1–8

    2023

  8. [16]

    AmpAgent: An LLM-Based Multi-Agent System for Multi-Stage Amplifier Schematic Design from Literature for Process and Performance Porting,

    C. Liu, W. Chen, A. Peng, Y . Du, L. Du, and J. Yang, “AmpAgent: An LLM-Based Multi-Agent System for Multi-Stage Amplifier Schematic Design from Literature for Process and Performance Porting,”arXiv preprint arXiv:2409.14739, 2024. [Online]. Available: https://arxiv.org/abs/2409.14739

    2024 arXiv

  9. [17]

    Ask-EDA: A Design Assistant Empowered by LLM, Hybrid RAG and Abbreviation De-hallucination,

    L. Shi, M. Kazda, B. Sears, N. Shropshire, and R. Puri, “Ask-EDA: A Design Assistant Empowered by LLM, Hybrid RAG and Abbreviation De-hallucination,” in2024 IEEE LLM Aided Design Workshop (LAD), 2024, pp. 1–5

    2024

  10. [18]

    MuaLLM: A Multimodal Large Language Model Agent for Circuit Design Assistance with Hybrid Contextual Retrieval-Augmented Generation,

    P. Abbineni, S. Aldowaish, C. Liechty, S. Noorzad, A. G. Ghalati, and M. Fayazi, “MuaLLM: A Multimodal Large Language Model Agent for Circuit Design Assistance with Hybrid Contextual Retrieval-Augmented Generation,” in2026 31st Asia and South Pacific Design Automation Conferen...

    2026

  11. [19]

    AnalogSeeker: An Open-source Foundation Language Model for Analog Circuit Design,

    Z. Chen, J. Zhuang, J. Shen, X. Ke, X. Yang, M. Zhou, Z. Du, X. Yan, Z. Wu, Z. Xu, J. Huang, L. Shang, X. Zeng, and F. Yang, “AnalogSeeker: An Open-source Foundation Language Model for Analog Circuit Design,” 2025. [Online]. Available: https://arxiv.org/abs/2508.10409

    2025 arXiv

  12. [20]

    Improving LLM- Powered EDA Assistants with RAFT,

    L. Shi, M. Kazda, C. Schmitter, and H. Gupta, “Improving LLM- Powered EDA Assistants with RAFT,” in2025 IEEE International Conference on LLM-Aided Design (ICLAD), 2025, pp. 9–15

    2025

  13. [21]

    AMSnet-KG: A Netlist Dataset for LLM-based AMS Circuit Auto-design Using Knowledge Graph RAG,

    Y . Shi, Z. Tao, Y . Gao, T. Zhou, C. Chang, Y . Wang, B. Chen, G. Zhang, A. Liu, Z. Yu, T.-J. Lin, and L. He, “AMSnet-KG: A Netlist Dataset for LLM-based AMS Circuit Auto-design Using Knowledge Graph RAG,”ACM Trans. Des. Autom. Electron. Syst., vol. 30, no. 6, Oct. 2025. [Onl...

    2025 doi

  14. [22]

    A 24- to-29GHz Compact Transmit/Receive Front-End Module Featuring an Asymmetric Doherty Power Amplifier and 0.22mm2 Area,

    X. Zhang, R. Wang, Q. Zhou, H. Guo, C. Shi, and T. Chi, “A 24- to-29GHz Compact Transmit/Receive Front-End Module Featuring an Asymmetric Doherty Power Amplifier and 0.22mm2 Area,” in2025 IEEE International Solid-State Circuits Conference (ISSCC), vol. 68, 2025, pp. 1–3

    2025

  15. [23]

    A 28GHz Frequency-Diverse Sub-Array TX with Secret Phase Keys and Antenna Subset Modulation for Eavesdropping-Resilient Wireless Communi- cation,

    Q. Zhou, Y . Su, H. Guo, Y . Hu, K. Yang, and T. Chi, “A 28GHz Frequency-Diverse Sub-Array TX with Secret Phase Keys and Antenna Subset Modulation for Eavesdropping-Resilient Wireless Communi- cation,” in2026 IEEE International Solid-State Circuits Conference (ISSCC), vol. 69,...

    2026

  16. [24]

    A Packaged D-Band Transmit- ter with a Multifeed Lens Antenna Achieving 25.3dBm Single-Element EIRP for 2-D Scalable Arrays,

    H. Wang, H. Guo, X. Zhang, and T. Chi, “A Packaged D-Band Transmit- ter with a Multifeed Lens Antenna Achieving 25.3dBm Single-Element EIRP for 2-D Scalable Arrays,” in2025 IEEE Custom Integrated Circuits Conference (CICC), 2025, pp. 1–3

    2025

  17. [25]

    A Millimeter-Wave Four-Way Doherty Power Amplifier With Over-GHz Modulation Bandwidth,

    X. Zhang, H. Guo, and T. Chi, “A Millimeter-Wave Four-Way Doherty Power Amplifier With Over-GHz Modulation Bandwidth,”IEEE Journal of Solid-State Circuits, vol. 59, no. 12, pp. 3898–3914, 2024

    2024

  18. [26]

    A 9.05-to-37.0GHz LO Generator with Magnetic Mode Switching and Tuning-Free Octave-Bandwidth Common-Mode Resonator Achieving>190.7dBc/Hz FoM,

    H. Guo, Y . Hu, and T. Chi, “A 9.05-to-37.0GHz LO Generator with Magnetic Mode Switching and Tuning-Free Octave-Bandwidth Common-Mode Resonator Achieving>190.7dBc/Hz FoM,” in2025 IEEE International Solid-State Circuits Conference (ISSCC), vol. 68, 2025, pp. 560–562

    2025

  19. [27]

    Pushing the Performance Boundaries of MmWave and Sub-THz Transceiver Circuits Through Passive Network Design Innovations,

    T. Chi, Y . Hu, X. Zhang, and H. Guo, “Pushing the Performance Boundaries of MmWave and Sub-THz Transceiver Circuits Through Passive Network Design Innovations,” inProc. IEEE Int. Midwest Symp. Circuits Syst. (MWSCAS), 2024, pp. 759–763

    2024

  20. [28]

    AI-Assisted Data-Driven RFIC Designs: A Review of Recent Progress, Design Frameworks, and Challenges,

    C. Chu, Y . Xu, S. Fu, T.-Y . Huang, K. Manetakis, P. A. D. Fabbro, and H. Wang, “AI-Assisted Data-Driven RFIC Designs: A Review of Recent Progress, Design Frameworks, and Challenges,”IEEE Journal of Selected Topics in Electromagnetics, Antennas and Propagation, vol. 2, pp. 1–25, 2026

    2026

  21. [29]

    Deep-Learning Enabled Generalized Inverse Design of Multi-Port Radio-Frequency and Sub-Terahertz Passives and Integrated Circuits,

    E. A. Karahanet al., “Deep-Learning Enabled Generalized Inverse Design of Multi-Port Radio-Frequency and Sub-Terahertz Passives and Integrated Circuits,”Nature Communications, vol. 15, no. 1, Dec 2024

    2024

  22. [30]

    ML-Assisted RF IC Design Enablement: the New Frontier of AI for EDA,

    H. Chae, S. Chai, T. Chi, S. Li, and D. Z. Pan, “ML-Assisted RF IC Design Enablement: the New Frontier of AI for EDA,” inProc. Asia and South Pacific Design Automation Conf. (ASP-DAC), Jan. 2025, pp. 683–689

    2025

  23. [31]

    MOTIF-RF: Multi-template On-chip Transformer Synthesis Incorporating Frequency- domain Self-transfer Learning for RFIC Design Automation,

    H. He, Y . Xu, L. Xia, Y . Hu, F. Cai, and T. Chi, “MOTIF-RF: Multi-template On-chip Transformer Synthesis Incorporating Frequency- domain Self-transfer Learning for RFIC Design Automation,” in2026 31st Asia and South Pacific Design Automation Conference (ASP-DAC), 2026, pp. 1138–1144

    2026

  24. [32]

    AdreamDCO: AI- Driven Robust and Efficient Design Automation for Digitally Controlled Oscillators,

    Y . Hu, H. Guo, S. Wang, J. Liu, W. Cao, and T. Chi, “AdreamDCO: AI- Driven Robust and Efficient Design Automation for Digitally Controlled Oscillators,” in2025 62nd ACM/IEEE Design Automation Conference (DAC), 2025, pp. 1–7

    2025

  25. [33]

    Invited Paper: Towards Generative AI for Analog and RF IC Design: From Spec to Layout,

    H. Chae, S. Kim, S. Poddar, X. Gao, S. Li, and D. Z. Pan, “Invited Paper: Towards Generative AI for Analog and RF IC Design: From Spec to Layout,” in2025 IEEE/ACM International Conference On Computer Aided Design (ICCAD), 2025, pp. 1–9

    2025

  26. [34]

    Large Language Models are Reasoning Teachers,

    N. Ho, L. Schmid, and S.-Y . Yun, “Large Language Models are Reasoning Teachers,” inProceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Toronto, Canada: Association for Computational Linguistics, 2023, pp. 14 852–14 ...

    2023

  27. [35]

    Keypoint-based Progressive Chain-of-Thought Distillation for LLMs,

    K. Feng, C. Li, X. Zhang, J. Zhou, Y . Yuan, and G. Wang, “Keypoint-based Progressive Chain-of-Thought Distillation for LLMs,” inProceedings of the 41st International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, vol

  28. [36]

    AMSBench: A comprehensive benchmark for evaluating MLLM capabilities in AMS circuits,

    Y . Shi, Z. Zhang, H. Wang, Z. Tao, Z. Li, B. Chen, Y . Wang, Z. Yu, T.-J. Lin, and L. He, “AMSBench: A comprehensive benchmark for evaluating MLLM capabilities in AMS circuits,” arXiv preprint arXiv:2505.24138, 2025. [Online]. Available: https: //arxiv.org/abs/2505.24138

    2025

  29. [37]

    MMCircuitEval: A Comprehensive Multimodal Circuit-Focused Benchmark for Evaluat- ing LLMs,

    C. Zhao, Z. Shi, X. Wen, C. Liu, Y . Liu, Y . Zhou, Y . Zhao, H. Feng, Y . Zhu, G.-W. Wan, X. Cheng, W. Chen, Y . Fu, C. Chen, C. Xue, Y . Wang, Y . Lin, J. Yang, N. Xu, X. Wang, and Q. Xu, “MMCircuitEval: A Comprehensive Multimodal Circuit-Focused Benchmark for Evaluat- ing L...

    2025

  30. [38]

    Textbooks Are All You Need,

    S. Gunasekaret al., “Textbooks Are All You Need,”arXiv preprint arXiv:2306.11644, 2023. [Online]. Available: https://arxiv.org/abs/2306. 11644

    2023 arXiv

  31. [39]

    The Llama 3 Herd of Models,

    A. Grattafioriet al., “The Llama 3 Herd of Models,”arXiv preprint arXiv:2407.21783, 2024. [Online]. Available: https://arxiv.org/abs/2407. 21783

    2024 arXiv

  32. [40]

    Qwen3 technical report,

    A. Yanget al., “Qwen3 technical report,”arXiv preprint arXiv:2505.09388, 2025. [Online]. Available: https://arxiv.org/abs/ 2505.09388

    2025 arXiv

  33. [41]

    M3- Embedding: Multi-Linguality, Multi-Functionality, Multi-Granularity Text Embeddings Through Self-Knowledge Distillation,

    J. Chen, S. Xiao, P. Zhang, K. Luo, D. Lian, and Z. Liu, “M3- Embedding: Multi-Linguality, Multi-Functionality, Multi-Granularity Text Embeddings Through Self-Knowledge Distillation,” inFindings of the Association for Computational Linguistics: ACL 2024, L.-W. Ku, A. Martins, ...

    2024

  34. [42]

    MenTeR: A fully-automated Multi-agenT workflow for end-to-end RF/Analog Circuits Netlist Design,

    P.-H. Chen, Y .-S. Lin, W.-C. Lee, T.-Y . Leu, P.-H. Hsu, A. Dissanayake, S. Oh, and C.-S. Chiu, “MenTeR: A fully-automated Multi-agenT workflow for end-to-end RF/Analog Circuits Netlist Design,” in2025 IEEE International Conference on LLM-Aided Design (ICLAD), 2025, pp. 124–132

    2025

  35. [43]

    Okapi at TREC-3,

    S. E. Robertson, S. Walker, S. Jones, M. Hancock-Beaulieu, and M. Gat- ford, “Okapi at TREC-3,” inProceedings of the Third Text REtrieval Conference (TREC-3). National Institute of Standards and Technology (NIST), 1994

    1994

  36. [44]

    Reciprocal rank fusion outperforms Condorcet and individual rank learning methods,

    G. V . Cormack, C. L. A. Clarke, and S. B ¨uttcher, “Reciprocal rank fusion outperforms Condorcet and individual rank learning methods,” inProc. 32nd International ACM SIGIR Conference on Research and Development in Information Retrieval, 2009, pp. 758–759

    2009

  37. [45]

    Training Compute-Optimal Large Language Models,

    J. Hoffmannet al., “Training Compute-Optimal Large Language Models,”arXiv preprint arXiv:2203.15556, 2022. [Online]. Available: https://arxiv.org/abs/2203.15556 APPENDIXA A REPRESENTATIVE MCQTSA SAMPLE Fig. 3 shows a complete mcQTSA sample produced by our QTSA pipeline, includ...

    2022 arXiv

  38. [235]

    13 241–13 255

    PMLR, 2024, pp. 13 241–13 255. [Online]. Available: https: //proceedings.mlr.press/v235/feng24e.html

    2024

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.