Pith. sign in

REVIEW 3 major objections 6 minor 45 references

MicroEvo: Knowledge-Guided LLM Sampling for Efficient Microarchitecture Design Space Exploration

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read MicroEvo claims that a knowledge-guided LLM inside a Monte Carlo tree search loop turns scarce simulation budgets into much better Pareto fronts, beating NSGA-II by up to 36.2%.

desk verdict Big and fairly consistent hypervolume gains for LLM+MCTS in microarchitecture DSE, but the paper never isolates whether the LLM's internalized knowledge or the search scaffold is responsible. read the letter →

arxiv 2608.06183 v1 pith:OWKL3RMN submitted 2026-08-06 cs.AI

Jia Xiong , Runkai Li , Chenxu Niu , Guangyuan Gao , Changwen Xing , Yifan Zhang , Xinlai Wan , Jieran Cui
show 6 more authors
Chen Bai Yusheng Hua Ying Wang Ming Ling Xi Wang Tao Xie
This is my paper
classification cs.AI
keywords microarchitecturedesignspaceexplorationlargelanguagemodelsMonteCarloTreeSearchmulti-objectiveoptimizationParetofrontqualitysampleefficiencyprocessorknowledgeaccumulation
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

MicroEvo sets out to settle a practical question in processor design: when each performance, power, and area (PPA) simulation is expensive and only a few dozen can be run, how should the next configuration be chosen? The paper's answer is that a large language model, equipped with the structured search of Monte Carlo Tree Search and a memory of insights extracted from previous evaluations, can generate far more informative candidates than blind evolutionary or Bayesian search. The authors report up to 36.2% higher Pareto-front hypervolume than NSGA-II under a 20-evaluation budget, 10.6x faster time-to-target, and an 8.6% geometric-mean energy-efficiency gain over a hand-crafted design on the industrial-scale XiangShan Kunminghu core. If correct, this would make LLM-guided sampling a practical tool for microarchitecture design-space exploration under realistic simulation constraints.

What carries the argument

The load-bearing mechanism is an evolutionary Monte Carlo tree in which microarchitecture configurations are nodes and LLM-generated parameter modifications are edges. Selection uses Pareto-UCT, a multi-objective extension of UCT whose reward is hypervolume improvement, $Q_{\mathrm{HVI}}(v_i)=\mathrm{HV}(P\cup\{f(v_i)\})-\mathrm{HV}(P)$, augmented by a crowding-distance term that favors sparsely covered Pareto regions and a decayed exploration term. Expansion is performed by two LLM operators: the knowledge tuner, which makes local refinements guided by accumulated insights and Pareto-gap analysis, and the pattern explorer, which proposes structural alternatives from sibling-node context. Active Knowledge Accumulation converts each evaluated parent-child pair into reusable insights stored in a utility-scored memory, and the State-Aware Directive switches the LLM prompt among exploit, balance, and explore modes depending on stagnation count and operator effectiveness. Together these components convert the LLM's pretrained sense of microarchitectural coupling into a self-improving search that balances local refinement and global exploration.

What would settle it

Run MicroEvo on a design space whose parameter names and values have been randomly permuted so the LLM cannot bring semantic knowledge to bear, while keeping the tree search, Pareto-UCT, AKA, and SAD intact; if hypervolume gains over NSGA-II persist under this scrambled space, the claimed advantage is not caused by LLM knowledge but by the search machinery alone.

Watch

Extended reading notes

Core claim

The central claim is that the bottleneck in expensive microarchitecture design-space exploration is sampling, not prediction: existing methods waste simulations because they search blindly or use LLM feedback only as raw numerical hints. MicroEvo claims to fix this by making the LLM a knowledge-guided evolutionary operator inside an MCTS loop, where each node is a configuration, each edge is an LLM-proposed modification, and the reward is the node's marginal contribution to the global Pareto front. The paper reports that this design improves Pareto-front quality by up to 36.2% in hypervolume over NSGA-II, reaches a target hypervolume with 10.6x fewer evaluations, and on the industrial-scale XiangShan Kunminghu core the method's explored designs surpass a manual expert design by 8.6% in geometric mean energy efficiency. The ablation results are part of the claim: removing knowledge accumulation and state-aware directives drops hypervolume by roughly 7%, and a one-shot LLM generator is clearly worse than the iterative tree search, which the authors take as evidence that the gain comes from learning from the search itself rather than from memorized designs.

Load-bearing premise

The entire advantage rests on the LLM's pretrained knowledge of microarchitectural correlations transferring to the target processor's parameter space and remaining useful as the search accumulates its own insights; if that prior does not transfer, or if the accumulated knowledge degrades over iterations, MicroEvo would reduce to random mutation inside MCTS and lose the reported gains.

Editorial extensions

If this is right

  • Under a 20-evaluation budget, MicroEvo still produces broader Pareto coverage than baselines running 45 evaluations, which suggests scarce-budget design decisions can be guided by this kind of knowledge-driven sampling.
  • Time-to-target comparison shows MicroEvo reaching the target hypervolume with 10.6x fewer evaluations than NSGA-II and 8.3% of Boom-Explorer's execution time, so the practical cost is dominated by simulation, not LLM inference.
  • LLM inference accounts for only 7.31% of total runtime, with an API cost of about $0.024 for DeepSeek-V3.2 per search, meaning the method is economically viable for routine use.
  • On the XiangShan Kunminghu core, the method achieves the best hypervolume trajectory among all compared methods and surpasses the expert-crafted XS-GEM5 design by 8.6% in geometric mean energy efficiency.
  • Ablations show that removing both AKA and SAD, or either LLM operator alone, degrades hypervolume by 5-9%, indicating each component carries part of the gain and that the whole loop, not just the LLM prior, does the work.

Reading between the lines

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

  • If the LLM's pretrained microarchitecture knowledge transfers across core families, the same framework should transfer to other hardware search spaces, such as GPU pipelines, network-on-chip configurations, or high-level synthesis design spaces, with the knowledge memory replacing retraining.
  • The AKA memory produces human-readable design insights as a byproduct; these could be surfaced to architects as an automatically generated design guide, though the paper only uses them internally for prompting.
  • A direct test of how much of the gain comes from the LLM prior versus the tree mechanism would be to shuffle or anonymize parameter names so the LLM cannot exploit semantic correlations; if the gain persists, the search structure alone is responsible.
  • The framework's knowledge is accumulated per run; sharing the memory across runs on related cores could make each subsequent exploration cheaper, an extension the paper does not explore.
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

3 major / 6 minor

Summary. The paper proposes MicroEvo, a framework that couples off-the-shelf LLMs with Monte Carlo Tree Search for multi-objective microarchitecture design space exploration. The framework combines LLM-driven evolutionary operators (a knowledge tuner and a pattern explorer), a Pareto-aware tree policy (Pareto-UCT), an active knowledge accumulation mechanism with a utility-scored memory, and a state-aware directive controller that switches between exploit/balance/explore modes. Experiments on an Alpha21264-style core and the XiangShan Kunminghu processor compare MicroEvo against NSGA-II, MOTPE, LEMOE, Boom-Explorer, and RL-DSE, reporting hypervolume improvements up to 36.2%, ADRS reductions up to 57%, a 10.6x search-efficiency improvement over NSGA-II, and an 8.6% geometric-mean energy-efficiency gain over a hand-crafted expert design. The paper also reports ablations for each component and a token/cost breakdown.

Significance. The claimed gains, if confirmed, would be practically valuable because microarchitecture DSE is notoriously expensive and existing methods often ignore structural dependencies between parameters. The paper deserves credit for releasing code, evaluating on two very different design spaces (including an industrial-scale core), testing two LLMs, and including a per-component ablation study plus a cost analysis. However, the evaluation contains at least two load-bearing issues (the ADRS reference-set construction and the time-to-target metric) and a missing knowledge-free control that bears directly on the paper's central attribution of its gains to LLM internalized knowledge. If these issues are corrected, the framework is likely to be a useful contribution; as it stands, the quantitative claims need to be revisited.

major comments (3)
  1. [§4.1, Eq. (9)] The ADRS reference set P* is constructed by combining 5,000 randomly sampled design points with 'all solutions explored by the evaluated methods.' Because each method's own explored solutions are included in the reference front, a method that explores many diverse points will artificially lower its ADRS by being close to points it itself discovered, regardless of whether those points are on the true Pareto front. This situation biases the ADRS improvements reported in Table 1 (e.g., 53–57% reductions) and in Figure 10, and it is especially problematic because the reference set is not fixed across methods. I recommend recomputing ADRS with a reference front that is independent of the evaluated methods (e.g., a single large random sample, or a union of all methods' solutions excluding the method being evaluated) and reporting whether the relative rankings change.
  2. [§4.2, Eq. (11)] The time-to-target metric N_tau = min{t | HV(t) >= tau} is claimed to show a 10.6x search-efficiency improvement over NSGA-II. However, NSGA-II's final hypervolume at a budget of 45 evaluations is 0.663 (Table 1), below the threshold tau = 0.74 used in Figure 9. If a baseline never reaches the threshold within the allotted budget, N_tau is undefined or infinite, and the resulting ratio is not a valid finite speedup. The manuscript must clarify how N_tau was computed for methods that do not reach the threshold (e.g., by extrapolation, by capping at budget+1, or by using a lower threshold attained by all methods) and should report the raw N_tau values for every method. The 10.6x claim should be restated with this caveat.
  3. [§4.3, Table 3 and Figure 13] The paper attributes MicroEvo's improvements to 'knowledge-guided LLM sampling,' but no knowledge-free control is reported. Every ablation row in Table 3 retains the LLM operators; the only comparison without the LLM (single-objective UCT in Figure 13) simultaneously changes the selection policy, so the effect of removing the LLM's semantic knowledge is confounded. The observed gains could therefore stem entirely from the MCTS-based evolutionary loop and the active-learning heuristics rather than from the LLM's pretrained microarchitecture knowledge. Please add a control that replaces the LLM operators with a random mutation operator (e.g., uniformly random parameter perturbations with the same mutation rates) while keeping Pareto-UCT, AKA, and SAD fixed. This control is essential to support the central knowledge-attribution claim, and it is especially important for the XiangShan transfer results in Section 4.4, where the only direct evidence of LLM knowledge (Figure 2) is measured on a different, Alpha21264-style design space.
minor comments (6)
  1. [Throughout] The provided manuscript text contains many garbled Unicode placeholders (e.g., '/uni00000372/uni000001e4/...') in figures and body text; please ensure the camera-ready PDF renders all characters correctly.
  2. [Tables 1 and 3] Table 3 lists ADRS = 0.060 for the full MicroEvo at budget 45, while Table 1 reports ADRS = 0.056 for MicroEvo (DeepSeek) at the same budget; please reconcile this inconsistency.
  3. [Eq. (8)] The normalization of Q_HVI(vi | k) uses Q_min and Q_max but does not specify the set over which these extrema are computed (e.g., children of the same node, all historical expansions, or all Pareto-optimal points). Please define the normalization set explicitly.
  4. [§3.5] SAD's stagnation threshold and the criteria for 'weak local gains' versus 'strong local gains' are not quantified; please report the exact values used in the experiments so that the finite-state controller is reproducible.
  5. [§4.1] The ADRS reference set is described as 5,000 randomly sampled design points; please state the sampling distribution and whether the same random seed or point set is used for all compared methods.
  6. [Figure 2(b)] The z-axis label 'Hypervolume' is unclear because hypervolume is a set-level indicator, not a per-sample scalar. Please clarify how this value is computed and assigned to individual points.

Circularity Check

1 steps flagged · score 2.0 of 10

No derivation-level circularity; a self-referential ADRS reference set and a missing knowledge-free control are evaluation concerns, not equation-level circularity.

  1. other [Section 4.1, Eq. (9)-(10), ADRS definition]
    "For ADRS, we construct the reference Pareto front P∗ by combining 5,000 randomly sampled design points with all solutions explored by the evaluated methods: ADRS(P,P∗)= 1/|P∗| ∑_{x∗∈P∗} min_{x∈P} δ(x,x∗)"

    The reference set P* is built from the very outputs being compared. Because P* contains every solution explored by MicroEvo (and by each baseline), the min-distance from MicroEvo's own Pareto set to its own explored non-dominated points is zero by construction, and small for its dominated points. A method that explores a more diverse set is therefore automatically closer to a reference set that contains that same diverse set. This makes the reported 54-57% ADRS reductions partly attributable to a self-referential metric rather than to an externally fixed front. The HV results in Table 1 are computed against a fixed reference point and are not affected by this construction, so the central empirical claim retains independent support.

full rationale

MicroEvo is an empirical heuristic framework, not a mathematical derivation, and I find no equation-level circularity: Pareto-UCT, AKA, and SAD are all implemented as concrete algorithms whose outputs are evaluated by external GEM5/McPAT simulation, and the ablations in Table 3 test individual components. The main HV improvement (up to 36.2% over NSGA-II) is an external benchmark result, not a quantity forced by the method's definitions. The one genuine self-reference is the ADRS metric: its reference front includes every method's own explored solutions, so low ADRS is partially manufactured by exploring a diverse set of points that then appear in the reference. This is an evaluation artifact, not a derivation-level circularity. The search-efficiency metric N_tau is threshold-dependent, but the paper states the threshold (HV=0.74) explicitly; choosing a threshold that only MicroEvo reaches weakens the 10.6x claim but is not circular. The skeptic's strongest point is the absence of a knowledge-free control: every ablation still uses the LLM operator, so the specific attribution of gains to the LLM's internalized microarchitectural knowledge is untested. That is a real experimental gap and a correctness risk, but it is not a circularity under the hard rules. The self-citations [27]-[33] are contextual background about LLM use in chip design and are not load-bearing for any of MicroEvo's equations or claims. Overall, the paper's central HV comparison stands on independent external evaluation, so the circularity score is low.

Assumptions & free parameters 10 free parameters · 4 assumptions · 0 invented entities

The central claim rests on hand-set hyperparameters that are not sensitivity-analyzed, on the assumption that LLM pretrained knowledge transfers to the specific design spaces, and on an evaluation protocol whose reference set includes the method's own explored solutions. No new physical or conceptual entities are introduced.

free parameters (10)
  • e = 0.05
    Crowding distance weight in Pareto-UCT, set in Section 4.1; no sensitivity analysis.
  • lambda0 = 1.414
    Initial exploration coefficient in Eq. (6), set in Section 4.1.
  • omega = 0.3
    Visit penalty in knowledge memory utility, Eq. (7), set in Section 4.1.
  • HVI extraction threshold = 0.01
    Minimum HVI gain to trigger insight extraction in Algorithm 1.
  • Objective improvement threshold = 20%
    Alternative condition for insight extraction in Algorithm 1.
  • Semantic similarity threshold = 0.75
    Deduplication threshold for knowledge memory using MiniLM embeddings, Section 4.1.
  • Insights injected per iteration = 3
    Number of insights retrieved and injected into prompts, Section 4.1.
  • Operator children per stage = 2 per operator
    Each LLM operator generates 2 candidates per expansion, Section 4.1.
  • Initial root designs = 5
    Number of LLM-initialized configurations, Section 4.1.
  • Stagnation threshold = 1% of total HV
    SAD switches to explore mode when consecutive HVI gains fall below 1% of total HV, Section 3.5.
assumptions (4)
  • domain assumption LLMs possess internalized microarchitectural knowledge from pretraining
    Used in Section 2.2 to justify LLM-driven sampling; the entire method relies on the LLM generating structurally valid and high-quality designs.
  • domain assumption GEM5 and McPAT simulation outputs are accurate enough to serve as ground truth PPA for guiding the search
    All rewards in MCTS and all knowledge extraction use these simulation values (Section 3.1, Section 4.1).
  • ad hoc to paper The reference Pareto front P* for ADRS is a valid and unbiased baseline
    Constructed by combining random samples with all explored solutions of the evaluated methods (Section 4.1, Eq. 9), which can bias results.
  • standard math Hypervolume and crowding distance are reliable multi-objective indicators for guiding selection
    The Pareto-UCT criterion uses these standard metrics (Eq. 3-5).

how reviews work

0 comments
Cite this review

Pith. "Pith review of MicroEvo: Knowledge-Guided LLM Sampling for Efficient Microarchitecture Design Space Exploration." pith.science (2026). https://pith.science/paper/OWKL3RMN

@misc{pith2026260806183,
  author       = {Pith},
  title        = {Pith review of: MicroEvo: Knowledge-Guided LLM Sampling for Efficient Microarchitecture Design Space Exploration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OWKL3RMN}},
  note         = {Machine review of arXiv:2608.06183}
}
read the original abstract

Microarchitecture design space exploration suffers from expansive search spaces and expensive PPA evaluation, leaving only a small simulation budget for design decision-making. Existing methods perform blind search without considering microarchitectural dependencies and fail to learn from the iterative search effectively, leading to wasted evaluations and weak Pareto convergence. In this paper, we propose MicroEvo, a knowledge-guided framework that couples off-the-shelf LLMs with Monte Carlo Tree Search (MCTS) for multi-objective microarchitecture optimization. MicroEvo combines LLM-driven evolutionary operators, a Pareto-aware tree policy that balances Pareto contribution and diversity, an active knowledge accumulation mechanism that extracts and reuses optimization insights, and state-aware directives that adapt the search behavior online. Experiments show that MicroEvo improves Pareto-front quality by up to 36.2% over NSGA-II and achieves 10.6x higher search efficiency, and also demonstrates strong scalability to a complex industrial-scale core. The code repository is available at: https://github.com/GEAR-SEU/MicroEvo-ICCAD-26.

Figures

Figures reproduced from arXiv: 2608.06183 by the authors.

Figure 1
Figure 1. (a) The combinatorial explosion of the microarchitecture [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 3
Figure 3. Left: Comparison of the LLM-driven and EA-based DSE [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 5
Figure 5. Overview of MicroEvo. LLM first initializes root designs, Pareto-UCT then selects elite nodes in the MCTS tree [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗
Figures from the paper (7 more)
Figure 6
Figure 6. Figure 6: Prompt integration for LLM-driven expansion. [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 7
Figure 7. Figure 7: State-Aware Directive (SAD) workflow framework should favor minor refinements around promising de￾signs. In contrast, when the search stalls, repeated local modifica￾tions waste the simulation budget. To address this issue, we intro￾duce a state-aware directive (SAD) m…
Figure 8
Figure 8. Figure 8: Pareto front comparison of different DSE methods in the [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]
Figure 11
Figure 11. Figure 11: Token usage, cost, and total runtime breakdown of Mi [PITH_FULL_IMAGE:figures/full_fig_p007_11.png]
Figure 12
Figure 12. Figure 12: MCTS Search Progress over Iterations. initial sampling stage and ten rounds of iterative expansion, the total LLM invocation time is only 5.85 minutes, with an API cost of approximately $0.024 (34.52 minutes of LLM invocation and $1.32 API costs for Gemini-3-pro). LLM…
Figure 13
Figure 13. Figure 13: Effectiveness of the proposed Pareto-UCT. [PITH_FULL_IMAGE:figures/full_fig_p008_13.png]
Figure 15
Figure 15. Figure 15: Relative IPC/Power and IPC/Area improvement over the [PITH_FULL_IMAGE:figures/full_fig_p008_15.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 39 canonical work pages

  1. [1]

    Shaoshan Liu, Liangkai Liu, Jie Tang, Bo Yu, Yifan Wang, and Weisong Shi. 2019. Edge Computing for Autonomous Driving: Opportunities and Challenges.Proc. IEEE107, 8 (2019), 1697–1716

    work page 2019

  2. [2]

    Matthew Halpern, Yuhao Zhu, and Vijay Janapa Reddi. 2016. Mobile CPU’s rise to power: Quantifying the impact of generational mobile CPU design trends on performance, energy, and user satisfaction. In2016 IEEE International Symposium on High Performance Computer Architecture (HPCA). 64–76

    work page 2016

  3. [3]

    Jiménez, Tarun Nakra, Paul Kitchin, Ryan Hensley, Edward Brekelbaum, Vikas Sinha, and Ankit Ghiya

    Brian Grayson, Jeff Rupley, Gerald Zuraski Zuraski, Eric Quinnell, Daniel A. Jiménez, Tarun Nakra, Paul Kitchin, Ryan Hensley, Edward Brekelbaum, Vikas Sinha, and Ankit Ghiya. 2020. Evolution of the Samsung Exynos CPU Microar- chitecture. In2020 ACM/IEEE 47th Annual International Symposium on Computer Architecture (ISCA). 40–51

    work page 2020

  4. [4]

    Chen Bai, Jiayi Huang, Xuechao Wei, Yuzhe Ma, Sicheng Li, Hongzhong Zheng, Bei Yu, and Yuan Xie. 2023. ArchExplorer: Microarchitecture Exploration Via Bottleneck Analysis. InProceedings of the 56th Annual IEEE/ACM International Symposium on Microarchitecture (MICRO ’23). 268–282

    work page 2023

  5. [5]

    Xuyang Zhao, Tianning Gao, Zheng Wu, Zhaori Bi, Changhao Yan, Fan Yang, Sheng-Guo Wang, Dian Zhou, and Xuan Zeng. 2025. APPLE-DSE: Asynchronous Parallel Pareto Set Learning for Microarchitecture Design Space Exploration. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems 44, 7 (2025), 2765–2778

    work page 2025

  6. [6]

    Loh, Mahesh Subramony, and Sean White

    Samuel Naffziger, Noah Beck, Thomas Burd, Kevin Lepak, Gabriel H. Loh, Mahesh Subramony, and Sean White. 2021. Pioneering Chiplet Technology and Design for the AMD EPYC™and Ryzen™Processor Families : Industrial Product. In2021 ACM/IEEE 48th Annual International Symposium on Computer Architecture (ISCA). 57–70

    work page 2021

  7. [7]

    Karkhanis and James E

    Tejas S. Karkhanis and James E. Smith. 2007. Automated design of application specific superscalar processors: an analytical approach. InProceedings of the 34th Annual International Symposium on Computer Architecture (ISCA ’07). 402–411

    work page 2007

  8. [8]

    Guangyu Sun, Christopher Hughes, Changkyu Kim, Jishen Zhao, Cong Xu, Yuan Xie, and Yen-Kuang Chen. 2011. Moguls: A model to explore the memory hierar- chy for bandwidth improvements. In2011 38th Annual International Symposium on Computer Architecture (ISCA). 377–388

    work page 2011

Show all 45 references
  1. [9]

    Chen Bai, Qi Sun, Jianwang Zhai, Yuzhe Ma, Bei Yu, and Martin D.F. Wong. 2021. BOOM-Explorer: RISC-V BOOM Microarchitecture Design Space Exploration Framework. In2021 IEEE/ACM International Conference On Computer Aided Design (ICCAD). 1–9

    2021

  2. [10]

    Duo Wang, Mingyu Yan, Xin Liu, Mo Zou, Tianyu Liu, Wenming Li, Xiaochun Ye, and Dongrui Fan. 2023. A High-accurate Multi-objective Exploration Framework for Design Space of CPU. In2023 60th ACM/IEEE Design Automation Conference (DAC). 1–6

    2023

  3. [11]

    Runzhen Xue, Hao Wu, Mingyu Yan, Ziheng Xiao, Xiaochun Ye, and Dongrui Fan. 2025. MetaDSE: A Few-Shot Meta-Learning Framework for Cross-Workload CPU Design Space Exploration. InProceedings of the 62nd Annual ACM/IEEE Design Automation Conference (DAC ’25). 7 pages

    2025

  4. [12]

    Ziyang Yu, Chen Bai, Shoubo Hu, Ran Chen, Taohai He, Mingxuan Yuan, Bei Yu, and Martin Wong. 2023. IT-DSE: Invariance Risk Minimized Transfer Microar- chitecture Design Space Exploration. In2023 IEEE/ACM International Conference on Computer Aided Design (ICCAD). 1–9

    2023

  5. [13]

    Duo Wang, Mingyu Yan, Yihan Teng, Dengke Han, Haoran Dang, Xiaochun Ye, and Dongrui Fan. 2023. A Transfer Learning Framework for High-Accurate Cross- Workload Design Space Exploration of CPU. In2023 IEEE/ACM International Conference on Computer Aided Design (ICCAD). 1–9

    2023

  6. [14]

    Jianwang Zhai and Yici Cai. 2023. Microarchitecture Design Space Exploration via Pareto-Driven Active Learning.IEEE Transactions on Very Large Scale Integration (VLSI) Systems31, 11 (2023), 1727–1739

    2023

  7. [15]

    Xingyu Wu, Sheng-Hao Wu, Jibin Wu, Liang Feng, and Kay Chen Tan. 2025. Evolutionary Computation in the Era of Large Language Model: Survey and Roadmap.IEEE Transactions on Evolutionary Computation29, 2 (2025), 534–554

    2025

  8. [16]

    Jingyuan Li, Jianrong Zhang, Ye Li, Wenbo Yin, and Lingli Wang. 2025. LEMOE: LLM-Enhanced Multi-Objective Bayesian Optimization for Microarchitecture Exploration. In2025 62nd ACM/IEEE Design Automation Conference (DAC). 1–7

    2025

  9. [17]

    Zhantong Zhu, Kangbo Bai, and Tianyu Jia. 2026. Chat-A2: An LLM-aided Design Space Exploration Framework for High-Performance CPU Design. In2026 31st Asia and South Pacific Design Automation Conference (ASP-DAC). 540–546

    2026

  10. [18]

    Zheng Wu, Zhuochu Yang, Zhuoyuan Yang, Zihao Chen, Li Shang, and Fan Yang

  11. [19]

    Mingxin Tang, Wei Chen, Lizhou Wu, Libo Huang, and Kun Zeng. 2025. ChatDSE: A Zero-Shot Microarchitecture Design Space Explorer Powered by GPT4.0.ACM Trans. Des. Autom. Electron. Syst.30, 4 (2025), 24 pages

    2025

  12. [20]

    Fei Liu, Xialiang Tong, Mingxuan Yuan, Xi Lin, Fu Luo, Zhenkun Wang, Zhichao Lu, and Qingfu Zhang. 2024. Evolution of heuristics: towards efficient auto- matic algorithm design using large language model. InProceedings of the 41st International Conference on Machine Learning (...

    2024

  13. [21]

    Bernardino Romera-Paredes, Mohammadamin Barekatain, Alexander Novikov, Matej Balog, M Pawan Kumar, Emilien Dupont, Francisco JR Ruiz, Jordan S Ellenberg, Pengming Wang, Omar Fawzi, et al. 2024. Mathematical discoveries from program search with large language models.Nature625, ...

    2024

  14. [22]

    Zhi Zheng, Zhuoliang Xie, Zhenkun Wang, and Bryan Hooi. 2025. Monte Carlo Tree Search for Comprehensive Exploration in LLM-Based Automatic Heuristic Design. InInternational Conference on Machine Learning. 78338–78373

    2025

  15. [23]

    Haoran Ye, Jiarui Wang, Zhiguang Cao, Federico Berto, Chuanbo Hua, Haeyeon Kim, Jinkyoo Park, and Guojie Song. 2024. ReEvo: Large Language Models as Hyper-Heuristics with Reflective Evolution. InAdvances in Neural Information Processing Systems, Vol. 37. 43571–43608

    2024

  16. [24]

    Johnson Umeike, Neel Patel, Alex Manley, Amin Mamandipoor, Heechul Yun, and Mohammad Alian. 2023. Profiling gem5 Simulator. In2023 IEEE International Symposium on Performance Analysis of Systems and Software (ISPASS). 103–113

    2023

  17. [25]

    Guerreiro, Carlos M

    Andreia P. Guerreiro, Carlos M. Fonseca, and Luís Paquete. 2021. The Hypervol- ume Indicator: Computational Problems and Algorithms.ACM Comput. Surv.54, 6 (2021), 42 pages

    2021

  18. [26]

    Chen Bai, Jianwang Zhai, Yuzhe Ma, Bei Yu, and Martin D. F. Wong. 2024. To- wards Automated RISC-V Microarchitecture Design with Reinforcement Learning. Proceedings of the AAAI Conference on Artificial Intelligence38, 1 (2024), 12–20

    2024

  19. [27]

    Changwen Xing, SamZaak Wong, Xinlai Wan, Yanfeng Lu, Mengli Zhang, Zebin Ma, Lei Qi, Zhengxiong Li, Nan Guan, Zhe Jiang, Xi Wang, and Jun Yang. 2026. ChipMind: Retrieval-Augmented Reasoning for Long-Context Circuit Design Specifications.Proceedings of the AAAI Conference on Ar...

    2026

  20. [28]

    Xi Wang, Gwok-Waa Wan, Sam-Zaak Wong, Layton Zhang, Tianyang Liu, Qi Tian, and Jianmin Ye. 2024. ChatCPU: An Agile CPU Design and Verification Platform with LLM. InProceedings of the 61st ACM/IEEE Design Automation Conference (DAC). 6 pages

    2024

  21. [29]

    Lik Tung Fu, Jie Zhou, Shaokai Ren, Mengli Zhang, Jia Xiong, Hugo Jiang, Nan Guan, Xi Wang, and Jun Yang. 2026. ChatSVA: Bridging SVA Generation for Hardware Verification via Task-Specific LLMs.arXiv preprint arXiv:2604.02811 (2026)

    2026 arXiv

  22. [30]

    Runkai Li, Jia Xiong, Xiuyuan He, Jieru Zhao, Jiaqi Lv, Haowen Fang, Lei Qi, and Xi Wang. 2026. ChatHLS: Towards Systematic Design Automation and Optimization for High-Level Synthesis. InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics (V...

    2026

  23. [31]

    Gwok-Waa Wan, SamZaak Wong, Shengchu Su, Chenxu Niu, Ning Wang, Xin- lai Wan, Qixiang Chen, Mengnv Xing, Jingyi Zhang, Jianmin Ye, et al . 2026. FIXME: Towards End-to-End Benchmarking of LLM-Aided Design Verification. InProceedings of the AAAI Conference on Artificial Intellig...

    2026

  24. [32]

    Runkai Li, Jia Xiong, and Xi Wang. 2025. iDSE: Navigating Design Space Ex- ploration in High-Level Synthesis Using LLMs.arXiv preprint arXiv:2505.22086 (2025)

    2025 arXiv

  25. [33]

    Jianmin Ye, Tianyang Liu, Qi Tian, Shengchu Su, Zhe Jiang, and Xi Wang. 2025. ChatModel: Automating Reference Model Design and Verification with LLMs. arXiv preprint arXiv:2506.15066(2025)

    2025 arXiv

  26. [34]

    Maciej Świechowski, Konrad Godlewski, Bartosz Sawicki, and Jacek Mańdziuk

  27. [35]

    Yong Hu, Daniel Mueller-Gritschneder, and Ulf Schlichtmann. 2018. Wavefront- MCTS: Multi-objective Design Space Exploration of NoC Architectures based on Monte Carlo Tree Search. In2018 IEEE/ACM International Conference on Computer-Aided Design (ICCAD). 1–8

    2018

  28. [36]

    Hui Wang, Yang Liu, Xiaoyu Zhang, and Chaoxu Mu. 2025. CogMCTS: A Novel Cognitive-Guided Monte Carlo Tree Search Framework for Iterative Heuristic Evolution with Large Language Models.arXiv preprint arXiv:2512.08609(2025)

    2025

  29. [37]

    Chentong Chen, Mengyuan Zhong, Ye Fan, Jialong Shi, and Jianyong Sun. 2025. HiFo-Prompt: Prompting with Hindsight and Foresight for LLM-based Automatic Heuristic Design.arXiv preprint arXiv:2508.13333(2025)

    2025

  30. [38]

    Yoshihiko Ozaki, Yuki Tanigaki, Shuhei Watanabe, and Masaki Onishi. 2020. Multiobjective tree-structured parzen estimator for computationally expensive optimization problems. InProceedings of the 2020 Genetic and Evolutionary Com- putation Conference (GECCO ’20). 533–541

    2020

  31. [39]

    Reinhardt, Ali Saidi, Arkaprava Basu, Joel Hestness, Derek R

    Nathan Binkert, Bradford Beckmann, Gabriel Black, Steven K. Reinhardt, Ali Saidi, Arkaprava Basu, Joel Hestness, Derek R. Hower, Tushar Krishna, Somayeh Sardashti, Rathijit Sen, Korey Sewell, Muhammad Shoaib, Nilay Vaish, Mark D. Hill, and David A. Wood. 2011. The gem5 simulat...

    2011

  32. [40]

    Strong, Jay B

    Sheng Li, Jung Ho Ahn, Richard D. Strong, Jay B. Brockman, Dean M. Tullsen, and Norman P. Jouppi. 2009. McPAT: an integrated power, area, and timing modeling framework for multicore and manycore architectures. InProceedings of the 42nd Annual IEEE/ACM International Symposium o...

    2009

  33. [41]

    Official RISC-V Benchmark Suites

    2026. Official RISC-V Benchmark Suites. https://github.com/riscv-software- src/riscv-tests

    2026

  34. [42]

    Kaifan Wang, Jian Chen, Yinan Xu, Zihao Yu, Wei He, Dan Tang, Ninghui Sun, and Yungang Bao. 2025. XiangShan: An Open Source Project for High-Performance RISC-V Processors Meeting Industrial-Grade Standards.IEEE Micro45, 3 (2025), 49–57

    2025

  35. [43]

    2026. XS-GEM5. https://github.com/OpenXiangShan/GEM5

    2026

  36. [2023]

    Artificial Intelligence Review56, 3 (2023), 2497–2562

    Monte Carlo tree search: A review of recent modifications and applications. Artificial Intelligence Review56, 3 (2023), 2497–2562

    2023

  37. [2025]

    ChatArch: A Knowledge-driven Graph-of-thought LLM Framework for Processor Architecture Optimization.ACM Trans. Des. Autom. Electron. Syst.31, 2 (2025), 26 pages

    2025

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.