Pith. sign in

REVIEW 5 major objections 7 minor 55 references

PrefixAgent: An LLM-Powered Design Framework for Efficient Prefix Adder Optimization

T0 review · 5 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read PrefixAgent claims that a reasoning LLM fine-tuned on E-graph optimization traces can synthesize prefix adders with consistently smaller post-layout areas than search-based baselines and commercial synthesis tools, across 16-, 32-, and…

desk verdict PrefixAgent scales LLM-based prefix adder optimization to 64-bit with real physical wins, but the missing E-graph-only control leaves the LLM's contribution unproven. read the letter →

arxiv 2507.06127 v1 pith:56BRZ2R6 submitted 2025-07-08 cs.AR cs.AI

Dongsheng Zuo , Jiadong Zhu , Yang Luo , Yuzhe Ma This is my paper
classification cs.ARcs.AI
keywords prefixadderbackbonesynthesisE-graphequalitysaturationLLMfine-tuningtool-integratedreasoningarithmeticcircuitoptimizationarea-delaytrade-off
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

This paper claims that prefix-adder optimization can be decomposed into backbone synthesis and local structural refinement, and that this decomposition lets a large reasoning model, fine-tuned on E-graph-derived optimization traces, out-design both search-based baselines and a commercial synthesis tool. The central empirical claim is that the framework, PrefixAgent, consistently achieves smaller post-layout areas under equal delay targets, including an 11.4% area reduction over the best baseline at 64 bits and a smaller area than the commercial flow in every reported configuration. If correct, this establishes that carefully generated supervision traces, not brute-force search, can give LLM agents practical skill in a constrained hardware design task, and that the approach scales to 64-bit adders where prior LLM-based methods fail.

What carries the argument

The load-bearing object is the backbone, the subgraph that computes the most-significant-bit carry in a prefix adder; it is a full binary tree with $N-1$ nodes, and every valid prefix adder can be completed from it by adding auxiliary nodes. The paper shows a one-to-one correspondence between backbone regroup operations and the associativity rewrite rule $(o (o ?x ?y) ?z) \leftrightarrow (o ?x (o ?y ?z))$ on BackboneLang S-expressions, so an E-graph under equality saturation compactly represents the full space of backbones. Extraction is driven by the recursive timing cost $C(n)=\max(C(l),C(r))+d+\lambda$ over a linear node-delay model $y=kx+b$; the equality-saturation engine used by the paper returns optimal backbones and explanation traces, which become the training data. The trained model then uses tool-integrated reasoning with four tools—regroup, level-opt, fanout-opt, node clone—so the LLM makes high-level decisions while deterministic routines perform exact structural edits.

What would settle it

Run the E-graph extraction with $\lambda$ set to 0 and to twice its reported value, then push the resulting adders through the same commercial flow; if the area advantage over the commercial synthesis tool does not survive both extremes at 64 bits, the claim that the cost model generalizes fails.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that the exponential prefix-adder design space can be collapsed by optimizing only the backbone—the binary tree of nodes computing the most-significant-bit carry—and then completing the adder with auxiliary nodes so that the result is a zero-deficiency or low-deficiency adder in the sense of Snir's theorem ($s_C(n)+d_C(n)=2n-2$). Each backbone regroup corresponds exactly to an associativity rewrite, so equality saturation in an E-graph encodes all backbones compactly, and extraction under a linear delay cost produces both high-quality solutions and interpretable rewrite traces. The paper fine-tunes a 32-billion-parameter reasoning LLM on these traces (with a composite loss over chain-of-thought and function-call accuracy) and lets the model drive two phases: backbone regrouping, then local timing refinements via level-opt, fanout-opt, and node cloning. The reported consequence is consistent area reduction against reinforcement-learning, Monte Carlo tree search, and variational-autoencoder baselines and a commercial synthesis tool, with Pareto dominance at 32 and 64 bits.

Load-bearing premise

The whole pipeline relies on an assumption that a simple linear delay formula, $y = kx + b$ with tuning constants $d$ and $\lambda$, ranks adder designs in the same order as true post-layout area; if that ranking is wrong, the training data and the learned model inherit the error and the area advantage collapses.

Editorial extensions

If this is right

  • A designer can obtain a competitive 64-bit adder under a given arrival-time profile and delay target from a single fine-tuned-model run, rather than from tens of thousands of sampled candidates.
  • The two-phase decomposition sidesteps the known weakness of LLMs on graphs with more than about 20 nodes, because the model reasons about the $N-1$ backbone nodes and leaves exact local edits to deterministic tools.
  • The E-graph trace-generation recipe is portable: any prefix-computation or arithmetic-tree problem whose operations are associative and whose cost is decomposable can be turned into supervised reasoning traces for an LLM.
  • Because the framework produces one optimized design per delay target and beats baselines that use 5000 samples per target, it maintains quality while cutting average runtime by more than 10x.
  • The reported Pareto fronts at 32 and 64 bits dominate all baseline methods, so the framework does not trade scalability for quality as bit-width grows.

Reading between the lines

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

  • Beyond the paper: the associative-rewrite-to-E-graph-to-LLM-traces pipeline should transfer to other circuit families with associative operators, such as multiplier reduction trees or wide AND-reduction networks, provided a decomposable cost function exists.
  • Beyond the paper: the fixed constants $k$, $b$, $d$, and $\lambda$ are a testable weak point; learning them per technology node from post-layout data could either preserve the reported advantage at more delay targets or reveal which configurations the current cost model cannot rank.
  • Beyond the paper: because all structural edits are deterministic tool calls, the framework's demonstrated capability is a joint property of the tool set and the learned policy; improving the tools (e.g., a smarter regroup candidate generator) should yield area gains that are independent of the LLM.
  • Beyond the paper: the paper's Pareto fronts are built from a single objective (area under delay target); adapting the extraction to explicitly multi-objective costs could generate a denser set of Pareto-optimal backbones and would test whether the reported dominance persists at finer delay granularity.
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

5 major / 7 minor

Summary. PrefixAgent decomposes prefix-adder optimization into a backbone-synthesis phase and a local-refinement phase. In Phase I an LRM (QwQ-32B) repeatedly selects regroup operations on a backbone, with candidates generated by a tool; in Phase II it invokes level-opt, fanout-opt, and node-clone tools guided by an Enhanced Prefix Representation and critical-path analysis. The authors generate training traces by running egg equality saturation on a BackboneLang DSL with a single associativity rewrite rule and a timing cost function (Eq. (7)), then ask GPT-o1 and Claude 3.7 Sonnet to produce chain-of-thought annotations for the regroup traces, and fine-tune with a composite loss. Experiments under uniform and non-uniform arrival profiles compare against PrefixRL, MCTS, CircuitVAE, and a commercial synthesis tool across 16/32/64 bits, reporting smaller areas, Pareto dominance, and over 10x speedup. Physical implementation is done with Yosys/OpenROAD/NanGate45 and a commercial 32nm flow, and functional correctness is checked with ABC.

Significance. If the claims hold, the paper makes a useful contribution: it demonstrates an LLM/LRM-driven pipeline that scales to 64-bit prefix adders, reports actual physical-implementation results with equivalence checking, and gives a concrete mechanism for turning E-graph rewrites into interpretable optimization traces for fine-tuning. The backbone-plus-refinement decomposition is a sensible way to shrink the adder search space. However, the central attribution of the reported improvements to the fine-tuned LRM is not yet established: the evaluation lacks an E-graph-only control, and the cost model that generates both the extracted solutions and the training traces is not validated. The paper is therefore a promising system description whose headline claim needs additional evidence before it can be fully supported.

major comments (5)
  1. [Section III-C2, Eq. (7)] The timing cost function used by the E-graph extractor is presented as y = kx + b and C(n) = max(C(l), C(r)) + d + lambda, but the constants k, b, d, and lambda are never reported, and the loss weight lambda in Ltotal is also unreported. Because this cost function selects both the extracted backbone solutions and the rewrite traces that constitute the fine-tuning data, any systematic mis-ranking of backbones by Eq. (7) relative to post-layout area and delay will be inherited by the trained LRM and cannot be corrected by imitation learning. Please report the constants and the loss weight, and provide at least a rank-correlation or sensitivity analysis of Eq. (7) against the physical results, or a clear argument that the constants are fixed by prior work and were not tuned on the evaluation configurations.
  2. [Section IV-B, Tables II and III] No E-graph-only baseline appears anywhere in the evaluation. The paper compares PrefixAgent only against PrefixRL, MCTS, CircuitVAE, and the commercial synthesis tool; it never runs the egg extractor alone (with the Eq. (7) cost and the same Phase II construction and physical flow) on the evaluation configurations. As a result, the reported area reductions could be fully explained by the E-graph search, with the LRM functioning only as a wrapper around the extractor. Please add this control: run the extractor on the same benchmarks, implement the best extracted backbones (and preferably several suboptimal ones) through the same flow, and report their areas and delays. This is necessary to support the central claim that the fine-tuned LRM contributes to the results.
  3. [Section IV-A, Fig. 9] Experiments are reported without seeds, error bars, or run-to-run variation: each baseline samples 5000 designs, while PrefixAgent generates a single structure per target delay, and the numbers in Tables II and III appear to be single measurements. The text says PrefixAgent "consistently" achieves lower area, but no statistical evidence is provided. Please report variation over multiple seeds or runs and state whether the reported delay constraints were met exactly or with the same 1% tolerance used for the baselines. The runtime comparison in Fig. 10 would also benefit from the same transparency.
  4. [Section IV-C, Table III] The commercial-flow comparison is under-specified. The exact commercial EDA tool and version, synthesis effort and optimization options, how the input arrival profiles are imposed, and how target delays map to CST constraints are not given. CST is invoked on "y = a + b style Verilog," but it is not stated whether CST is allowed to restructure the adder freely or is limited to mapping the given RTL. Without these details the comparison may not be apples-to-apples; please document the flow in enough detail to be reproducible.
  5. [Section III-C4 and Section IV] The fine-tuning data generation covers only Phase I regroup decisions, since the E-graph traces are regroup traces. No analogous high-quality supervision is described for the Phase II tools (level-opt, fanout-opt, node clone), even though Phase II is part of the reported results. Either include Phase II supervision and traces in the data-generation discussion, or provide an ablation showing that the base model's Phase II behavior is adequate. Without this, it is unclear how much of the result relies on components that were actually fine-tuned.
minor comments (7)
  1. [Section IV-B, Table II] For the 16-bit LSB-first profile, PrefixAgent ties MCTS at 217 um2; the sentence saying PrefixAgent "consistently achieves the lowest area" should acknowledge this tie.
  2. [Section IV-B, Fig. 9] With only six target-delay points per bit-width, the claim that the Pareto front "dominates all baseline approaches" is a statement about the sampled points, not the continuous Pareto front; please phrase the claim accordingly.
  3. [Section III-C1] The statement that Rule R1 encodes the complete backbone solution space should be justified briefly, for example by noting that associativity rotations connect all binary tree bracketings of the leaves, or by demonstrating that the E-graph reaches all canonical forms.
  4. [Section III-D, Algorithm 1] FindCandidates returns one candidate pair per column; the relationship between this enumeration and the full set of E-graph rewrite steps used for training should be clarified, since it affects how faithfully the LRM can reproduce the extracted traces.
  5. [Eq. (7) and Section III-C4] The symbol lambda is used both for the timing margin in Eq. (7) and for the loss weight in Ltotal; please disambiguate the notation.
  6. [Section IV-C] Please identify the commercial EDA tool and version; "commercial EDA tools" is insufficient for reproducibility.
  7. [Figures 1, 5-7] Several figures in the submitted PDF contain overlapping or duplicated labels, for example the repeated "Regroup" callouts in Figs. 5-7 and the "CallFinish 1" label in Fig. 1; please provide clean vector figures.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the reported area/delay results come from physical implementation against external baselines (MCTS, PrefixRL, CircuitVAE, and a commercial tool), so Eq. (7) does not construct the outcome; the co-authored linear-delay precedent is peer-reviewed independent evidence, and the missing E-graph-only ablation is an attribution gap rather than a circular step.

full rationale

The derivation chain separates generation from evaluation. E-graph training data are produced by equality saturation under rewrite rule (R1) and extracted with the cost function of Eq. (7), C(n) = max(C(l), C(r)) + d + lambda, under a linear delay model y = kx + b (Section III-C2). The fine-tuned LRM imitates the resulting regroup traces. The claimed outcomes - the Pareto fronts of Fig. 9 and the area/delay numbers of Tables II and III - are post-layout measurements from OpenROAD and a commercial 32nm flow, with functional correctness verified by ABC; they are not computed from Eq. (7), so the headline result is not the cost function under another name. The one load-bearing premise with author overlap is the linear delay proxy, justified only by the sentence: 'Prior work [8], [16] has shown that a linear approximation provides accurate estimates of path delays in prefix adders.' References [8] and [16] are peer-reviewed publications with external co-authors, and nothing in this paper fits their constants to the evaluation configurations, so they count as independent evidence rather than a self-citation loop. The genuine weaknesses are: (i) the constants k, b, d, and lambda are never reported and no correlation between the Eq. (7) ranking and post-layout area is shown in this paper; (ii) Section IV-A describes no ablation that runs the E-graph extractor alone through Phase II and the physical flow, so the marginal contribution of the fine-tuned LRM over the extractor is not isolated. If the linear proxy systematically mis-ranks backbones, the training traces inherit the error, and the Tables II/III wins could be attributable to the E-graph search rather than to the LLM. That is a correctness and attribution risk, not a definitional reduction: the paper neither predicts Eq. (7) nor fits a parameter to the reported areas. The externally benchmarked physical results keep the core claim self-contained, so the circularity score stays at 1.

Assumptions & free parameters 7 free parameters · 7 assumptions · 3 invented entities

The central empirical claim rests on an E-graph extractor whose cost model has at least four undisclosed constants, and on a fine-tuning pipeline whose loss weight and data synthesis are also undisclosed. No code or data are released, so the paper's contribution cannot be separated from the E-graph optimizer that generated its training traces.

free parameters (7)
  • linear delay model coefficient k = not reported
    The delay model y = kx + b in Section III-C2 is used to rank backbone solutions; k is not given.
  • linear delay model intercept b = not reported
    The intercept b in the same delay model is not reported.
  • node delay d = not reported
    Eq. (7) defines C(n) = max(C(l), C(r)) + d + lambda; d affects all cost values and is not disclosed.
  • timing margin lambda = not reported
    Eq. (7) uses lambda to model fanout or layout uncertainty; if tuned on the evaluation configurations, the results are partly fitted.
  • loss weight lambda in Ltotal = not reported
    The composite objective Ltotal = LCoT + lambda * Lfunction_call in Section III-C4 is central to fine-tuning but the weight is not given.
  • extraction perturbation amount = not reported
    Suboptimal solutions are generated by controlled perturbations during extraction, but the perturbation distribution or magnitude is not specified.
  • max iterations K = not reported
    Algorithm 1 caps Phase I at K iterations; K is not reported and affects the quality of the final backbone.
assumptions (7)
  • standard math The prefix operator o is associative, as defined in Eq. (2), so regroup rewrites preserve meaning.
    Associativity is stated in Section II-A and is the basis for all E-graph rewrites in Section III-C.
  • domain assumption Snir's theorem characterizes zero-deficiency adders by S + L = 2n - 2.
    Invoked in Section II-A and III-B to justify that the backbone plus auxiliary-node construction yields zero- or low-deficiency adders.
  • domain assumption A linear path-length model accurately estimates prefix adder delays.
    Stated in Section III-C2 and attributed to prior work [8], [16]; the E-graph extraction ranking depends on this model.
  • ad hoc to paper The associativity rewrite rule R1 on BackboneLang S-expressions corresponds exactly to a backbone regroup operation and covers the full backbone solution space.
    This correspondence is asserted in Section III-C1 but not formally proven; it is load-bearing for the E-graph training data generation.
  • domain assumption Adding auxiliary nodes to a backbone as described produces a valid complete prefix adder with the stated level properties.
    The construction in Section III-B relies on Snir's necessary condition for zero-deficiency adders.
  • ad hoc to paper Chain-of-thought traces generated by GPT-o1 and Claude 3.7 Sonnet from E-graph rewrite steps are high-quality supervision for fine-tuning.
    Section III-C4 assumes that commercially generated CoT traces, combined with function calls, train the LRM effectively; this is not validated by ablation.
  • ad hoc to paper The composite loss with a function-call term improves structural decision accuracy over CoT-only supervision.
    Stated in Section III-C4 without experimental ablation showing the contribution of the function-call loss term.
invented entities (3)
  • Backbone subgraph
    purpose: Reduces prefix adder optimization to optimizing the binary tree that computes the MSB carry, cutting the design space from a product of Catalan numbers to a single Catalan number.
    This is a new decomposition concept introduced in Section III-B; it has no external falsifiable handle outside the framework.
  • BackboneLang DSL
    purpose: Encodes backbone structures as S-expressions for E-graph rewriting in the egg library.
    A domain-specific language introduced in Section III-C1 for internal use only.
  • Enhanced Prefix Representation (EPR)
    purpose: Provides node-level attributes such as logic level, parent relationships, and fanout to the LRM during Phase II refinement.
    A representation introduced in Section III-E; its benefit is asserted from preliminary experiments but not quantified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PrefixAgent: An LLM-Powered Design Framework for Efficient Prefix Adder Optimization." pith.science (2026). https://pith.science/paper/56BRZ2R6

@misc{pith2026250706127,
  author       = {Pith},
  title        = {Pith review of: PrefixAgent: An LLM-Powered Design Framework for Efficient Prefix Adder Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/56BRZ2R6}},
  note         = {Machine review of arXiv:2507.06127}
}
read the original abstract

Prefix adders are fundamental arithmetic circuits, but their design space grows exponentially with bit-width, posing significant optimization challenges. Previous works face limitations in performance, generalization, and scalability. To address these challenges, we propose PrefixAgent, a large language model (LLM)-powered framework that enables efficient prefix adder optimization. Specifically, PrefixAgent reformulates the problem into subtasks including backbone synthesis and structure refinement, which effectively reduces the search space. More importantly, this new design perspective enables us to efficiently collect enormous high-quality data and reasoning traces with E-graph, which further results in an effective fine-tuning of LLM. Experimental results show that PrefixAgent synthesizes prefix adders with consistently smaller areas compared to baseline methods, while maintaining scalability and generalization in commercial EDA flows.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

55 extracted references · 47 canonical work pages

  1. [1]

    A parallel algorithm for the efficient solution of a general class of recurrence equations,

    P. M. Kogge and H. S. Stone, “A parallel algorithm for the efficient solution of a general class of recurrence equations,” IEEE Transactions on Computers, 1973

    work page 1973

  2. [2]

    Conditional-sum addition logic,

    J. Sklansky, “Conditional-sum addition logic,” IRE Transactions on Elec- tronic Computers, 1960

    work page 1960

  3. [3]

    A regular layout for parallel adders,

    Brent and Kung, “A regular layout for parallel adders,” IEEE Transactions on Computers, 1982

    work page 1982

  4. [4]

    Fast area-efficient vlsi adders,

    T. Han and D. A. Carlson, “Fast area-efficient vlsi adders,” in 1987 IEEE 8th Symposium on Computer Arithmetic (ARITH) , 1987

    work page 1987

  5. [5]

    A depth-decreasing heuristic for combinational logic; or how to convert a ripple-carry adder into a carry-lookahead adder or anything in-between,

    J. Fishburn, “A depth-decreasing heuristic for combinational logic; or how to convert a ripple-carry adder into a carry-lookahead adder or anything in-between,” in Proc. DAC, 1990

    work page 1990

  6. [6]

    Non-heuristic optimization and synthesis of parallel- prefix adders,

    R. Zimmermann, “Non-heuristic optimization and synthesis of parallel- prefix adders,” in International Workshop on Logic and Architecture Synthesis, 1996

    work page 1996

  7. [7]

    Constructing zero-deficiency parallel prefix adder of minimum depth,

    H. Zhu, C.-K. Cheng, and R. Graham, “Constructing zero-deficiency parallel prefix adder of minimum depth,” in Proceedings of the 2005 Asia and South Pacific Design Automation Conference , 2005

    work page 2005

  8. [8]

    Ufo-mac: A unified framework for optimization of high-performance multipliers and multiply-accumulators,

    D. Zuo, J. Zhu, C. Li, and Y . Ma, “Ufo-mac: A unified framework for optimization of high-performance multipliers and multiply-accumulators,” in Proc. ICCAD, 2024

    work page 2024

Show all 55 references
  1. [9]

    An algorithmic approach for generic parallel adders,

    J. Liu, S. Zhou, H. Zhu, and C.-K. Cheng, “An algorithmic approach for generic parallel adders,” in Proc. ICCAD, 2003

    2003

  2. [10]

    Area minimization algorithm for parallel prefix adders under bitwise delay constraints,

    T. Matsunaga and Y . Matsunaga, “Area minimization algorithm for parallel prefix adders under bitwise delay constraints,” in Proceedings of the 17th ACM Great Lakes Symposium on VLSI , 2007

    2007

  3. [11]

    Size-optimized depth- constrained large parallel prefix circuits,

    S. Lin, B. Jiang, W. Sheng, and E. Young, “Size-optimized depth- constrained large parallel prefix circuits,” in Proceedings of the 61st ACM/IEEE Design Automation Conference , 2024

    2024

  4. [12]

    Towards optimal performance-area trade-off in adders by synthesis of parallel prefix struc- tures,

    S. Roy, M. Choudhury, R. Puri, and D. Z. Pan, “Towards optimal performance-area trade-off in adders by synthesis of parallel prefix struc- tures,” in Proc. DAC, 2013

    2013

  5. [13]

    Towards optimal performance-area trade-off in adders by synthe- sis of parallel prefix structures,

    ——, “Towards optimal performance-area trade-off in adders by synthe- sis of parallel prefix structures,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems , 2014

    2014

  6. [14]

    Polynomial time algorithm for area and power efficient adder synthesis in high-performance designs,

    ——, “Polynomial time algorithm for area and power efficient adder synthesis in high-performance designs,” IEEE Transactions on Computer- Aided Design of Integrated Circuits and Systems , 2016

    2016

  7. [15]

    A learning bridge from architectural synthesis to physical design for exploring power efficient high-performance adders,

    S. Roy, Y . Ma, J. Miao, and B. Yu, “A learning bridge from architectural synthesis to physical design for exploring power efficient high-performance adders,” in 2017 IEEE/ACM International Symposium on Low Power Electronics and Design (ISLPED) , 2017

    2017

  8. [16]

    Cross-layer optimization for high speed adders: A pareto driven machine learning approach,

    Y . Ma, S. Roy, J. Miao, J. Chen, and B. Yu, “Cross-layer optimization for high speed adders: A pareto driven machine learning approach,” IEEE TCAD, 2019

    2019

  9. [17]

    High-speed adder design space exploration via graph neural processes,

    H. Geng, Y . Ma, Q. Xu, J. Miao, S. Roy, and B. Yu, “High-speed adder design space exploration via graph neural processes,” IEEE TCAD, 2022

    2022

  10. [18]

    Prefixrl: Optimization of parallel prefix circuits using deep reinforcement learning,

    R. Roy, J. Raiman, N. Kant, I. Elkin, R. Kirby, M. Siu, S. Oberman, S. Godil, and B. Catanzaro, “Prefixrl: Optimization of parallel prefix circuits using deep reinforcement learning,” in Proc. DAC, 2021

    2021

  11. [19]

    Rl-mul: Multiplier design optimization with deep reinforcement learning,

    D. Zuo, Y . Ouyang, and Y . Ma, “Rl-mul: Multiplier design optimization with deep reinforcement learning,” in 2023 60th ACM/IEEE Design Au- tomation Conference (DAC), 2023

    2023

  12. [20]

    Rl-mul 2.0: Multiplier design op- timization with parallel deep reinforcement learning and space reduction,

    D. Zuo, J. Zhu, Y . Ouyang, and Y . Ma, “Rl-mul 2.0: Multiplier design op- timization with parallel deep reinforcement learning and space reduction,” ACM Trans. Des. Autom. Electron. Syst. , 2025

    2025

  13. [21]

    A hierarchical adaptive multi-task reinforcement learning framework for multiplier circuit design,

    Z. Wang, J. Wang, D. Zuo, J. Yunjie, X. Xia, Y . Ma, J. Hao, M. Yuan, Y . Zhang, and F. Wu, “A hierarchical adaptive multi-task reinforcement learning framework for multiplier circuit design,” in Proc. ICML, 2024

    2024

  14. [22]

    Computing circuits optimization via model-based circuit genetic evolution,

    Z. Wang, J. Wang, X. Xia, D. Zuo, L. Chen, Y . Ma, J. HAO, M. Yuan, and F. Wu, “Computing circuits optimization via model-based circuit genetic evolution,” in The Thirteenth International Conference on Learning Representations, 2025

    2025

  15. [23]

    Circuitvae: Efficient and scalable latent circuit optimiza- tion,

    J. Song et al. , “Circuitvae: Efficient and scalable latent circuit optimiza- tion,” in Proc. DAC, 2024

    2024

  16. [24]

    Scalable and effective arithmetic tree generation for adder and multiplier designs,

    Y . Lai, J. Liu, D. Z. Pan, and P. Luo, “Scalable and effective arithmetic tree generation for adder and multiplier designs,” in Advances in Neural Information Processing Systems , vol. 37, 2024

    2024

  17. [25]

    Learning to reason with llms,

    OpenAI, “Learning to reason with llms,” https://openai.com/index/ learning-to-reason-with-llms/

  18. [26]

    Openai o3,

    ——, “Openai o3,” https://openai.com/index/introducing-o3-and-o4-mini/ /

  19. [27]

    Qwq-max-preview,

    Alibaba, “Qwq-max-preview,” https://qwenlm.github.io/blog/ qwq-max-preview/

  20. [28]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning,

    DeepSeek-AI, D. Guo et al. , “Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning,” 2025. [Online]. Available: https://arxiv.org/abs/2501.12948

    2025 arXiv

  21. [29]

    Qwq-32b: Embracing the power of reinforcement learning,

    Alibaba, “Qwq-32b: Embracing the power of reinforcement learning,” https://qwenlm.github.io/blog/qwq-32b/

  22. [30]

    Analogcoder: Analog circuit design via training-free code generation,

    Y . Lai, S. Lee, G. Chen, S. Poddar, M. Hu, D. Z. Pan, and P. Luo, “Analogcoder: Analog circuit design via training-free code generation,” in Proc. AAAI, 2024

    2024

  23. [31]

    Lamagic: language-model-based topology generation for analog integrated circuits,

    C.-C. Chang et al., “Lamagic: language-model-based topology generation for analog integrated circuits,” in Proc. ICML, 2024

    2024

  24. [32]

    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 The Thirteenth International Conference on Learning Representations , 2025

    2025

  25. [33]

    Data is all you need: Finetuning llms for chip design via an automated design-data augmentation framework,

    K. Chang et al., “Data is all you need: Finetuning llms for chip design via an automated design-data augmentation framework,” in Proc. DAC, 2024

    2024

  26. [34]

    Graphwiz: An instruction-following language model for graph problems,

    N. Chen, Y . Li, J. Tang, and J. Li, “Graphwiz: An instruction-following language model for graph problems,” in Proc. KDD, 2024

    2024

  27. [35]

    How do large language models understand graph patterns? a benchmark for graph pattern comprehension,

    X. Dai, H. Qu, Y . Shen, B. Zhang, Q. Wen, W. Fan, D. Li, J. Tang, and C. Shan, “How do large language models understand graph patterns? a benchmark for graph pattern comprehension,” in The Thirteenth Interna- tional Conference on Learning Representations , 2025

    2025

  28. [36]

    Grapharena: Evaluating and exploring large language models on graph computation,

    J. Tang, Q. Zhang, Y . Li, N. Chen, and J. Li, “Grapharena: Evaluating and exploring large language models on graph computation,” in The Thirteenth International Conference on Learning Representations , 2025

    2025

  29. [37]

    Prefixllm: Llm-aided prefix circuit design,

    W. Xiao, V . S. C. Putrevu, R. V . Hemadri, S. Garg, and R. Karri, “Prefixllm: Llm-aided prefix circuit design,” 2024. [Online]. Available: https://arxiv.org/abs/2412.02594

    2024 arXiv

  30. [38]

    Limo: Less is more for reasoning,

    Y . Ye, Z. Huang, Y . Xiao, E. Chern, S. Xia, and P. Liu, “Limo: Less is more for reasoning,” 2025. [Online]. Available: https://arxiv.org/abs/2502.03387

    2025 arXiv

  31. [39]

    Limr: Less is more for rl scaling,

    X. Li, H. Zou, and P. Liu, “Limr: Less is more for rl scaling,” 2025. [Online]. Available: https://arxiv.org/abs/2502.11886

    2025 arXiv

  32. [40]

    s1: Simple test- time scaling,

    N. Muennighoff, Z. Yang, W. Shi, X. L. Li, L. Fei-Fei, H. Hajishirzi, L. Zettlemoyer, P. Liang, E. Cand `es, and T. Hashimoto, “s1: Simple test- time scaling,” 2025. [Online]. Available: https://arxiv.org/abs/2501.19393

    2025 arXiv

  33. [41]

    Depth-size trade-offs for parallel prefix computation,

    M. Snir, “Depth-size trade-offs for parallel prefix computation,” 1986

    1986

  34. [42]

    Tora: A tool-integrated reasoning agent for mathematical problem solving,

    Z. Gou, Z. Shao, Y . Gong, Y . Shen, Y . Yang, M. Huang, N. Duan, and W. Chen, “Tora: A tool-integrated reasoning agent for mathematical problem solving,” arXiv preprint arXiv:2309.17452 , 2023

    2023 arXiv

  35. [43]

    Torl: Scaling tool-integrated rl,

    X. Li, H. Zou, and P. Liu, “Torl: Scaling tool-integrated rl,” arXiv preprint arXiv:2503.23383, 2025

    2025 arXiv

  36. [44]

    Deepseekmath: Pushing the Limits of Mathematical Reasoning in Open Language Models,

    Z. Shao, P. Wang, Q. Zhu, R. Xu, J.-M. Song, M. Zhang, Y . K. Li, Y . Wu, and D. Guo, “Deepseekmath: Pushing the Limits of Mathematical Reasoning in Open Language Models,” arXiv.org, 2024

    2024

  37. [45]

    Techniques for program verification,

    C. G. Nelson, “Techniques for program verification,” Ph.D. dissertation, Stanford University, 1980

    1980

  38. [46]

    Equality saturation: a new approach to optimization,

    R. Tate, M. Stepp, Z. Tatlock, and S. Lerner, “Equality saturation: a new approach to optimization,” in Principles of Programming Languages (POPL), 2009

    2009

  39. [47]

    Constraint-aware e-graph rewriting for hardware performance optimization,

    S. Coward, T. Drane, and G. A. Constantinides, “Constraint-aware e-graph rewriting for hardware performance optimization,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems , 2025

    2025

  40. [48]

    E-syn: E-graph rewriting with technology-aware cost functions for logic synthesis,

    C. Chen, G. Hu, D. Zuo, C. Yu, Y . Ma, and H. Zhang, “E-syn: E-graph rewriting with technology-aware cost functions for logic synthesis,” in Proc. DAC, 2024

    2024

  41. [49]

    egg: Fast and extensible equality saturation,

    M. Willsey, C. Nandi, Y . R. Wang, O. Flatt, Z. Tatlock, and P. Panchekha, “egg: Fast and extensible equality saturation,” Proc. ACM Program. Lang., 2021

    2021

  42. [50]

    Claude 3.7 sonnet,

    Anthropic, “Claude 3.7 sonnet,” https://www.anthropic.com/claude/sonnet

  43. [51]

    Open Cell Library v2008 10 SP1,

    Nangate Inc., “Open Cell Library v2008 10 SP1,” 2008. [Online]. Available: http://www.nangate.com/openlibrary/

    2008

  44. [52]

    Yosys open synthesis suite,

    C. Wolf, “Yosys open synthesis suite,” https://yosyshq.net/yosys/

  45. [53]

    Toward an open-source digital flow: First learnings from the openroad project,

    T. Ajayi, V . A. Chhabria, M. Fogac ¸a, S. Hashemi, A. Hosny, A. B. Kahng, M. Kim, J. Lee, U. Mallappa, M. Neseem et al., “Toward an open-source digital flow: First learnings from the openroad project,” in Proc. DAC, 2019

    2019

  46. [54]

    ABC: A System for Sequential Synthesis and Verification,

    Berkeley Logic Synthesis and Verification Group, “ABC: A System for Sequential Synthesis and Verification,” http://www.eecs.berkeley.edu/ ∼alanmi/abc/

  47. [55]

    Sample-efficient optimization in the latent space of deep generative models via weighted retraining,

    A. Tripp et al., “Sample-efficient optimization in the latent space of deep generative models via weighted retraining,” in Proceedings of the 34th International Conference on Neural Information Processing Systems, 2020

    2020

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.