CoEvoP&R: Co-Evolving Placement Objectives with Routing Feedback via Large Language Models

Jie Han; Ramesh Karri; Ruogu Chen; Weihua Xiao

REVIEW 2 major objections 5 minor 26 references

CoEvoP&R: Co-Evolving Placement Objectives with Routing Feedback via Large Language Models

T0 review · 2 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 An LLM-based search loop evolves readable, differentiable placement objectives that improve routed wirelength, congestion, and timing over the native analytical placer.

desk verdict Solid, honestly-reported empirical paper with real routed gains, but the 'objective evolution' claim is softer than advertised because Eq. (7) pins every candidate to a fixed linear template and the champion programs are never shown. read the letter →

arxiv 2607.17398 v1 pith:R6S73UMF submitted 2026-07-19 cs.LG cs.AIcs.NE

Ruogu Chen , Weihua Xiao , Ramesh Karri , Jie Han This is my paper

classification cs.LGcs.AIcs.NE

keywords placementobjectiveslargelanguagemodelsprogramevolutionanalyticalroutabilitytimingclosureMAP-Eliteselectronicdesignautomation

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

The paper tries to establish that the differentiable objective an analytical placer minimizes does not have to be hand-designed or replaced by a black-box surrogate. It claims a closed loop can evolve readable, differentiable objectives automatically: a large language model proposes compact objective programs; each admitted program is embedded in the DREAMPlace placer, evaluated with placement, timing-proxy, and real-router feedback; and a MAP-Elites archive stores the survivors to guide the next round of proposals. On eight ChiPBench Nangate45 designs, the evolved objectives reduce routed wirelength by 16.9% and congestion by 36.7% over native DREAMPlace, with better worst negative slack and total negative slack; leave-one-design-out and zero-shot runs retain a large share of those gains. If true, this would mean downstream physical-design quality can be improved by searchable, inspectable objective programs rather than by expert retuning or opaque learned surrogates.

What carries the argument

The load-bearing mechanism is the scheduled objective template: at each placement iteration, the loss is smooth wirelength plus schedule-weighted density, routing-pressure, and pin-pressure terms, with the weights and the wirelength smoothing scale updated by a small state machine written by the LLM. Around this template, a MAP-Elites archive groups candidates by behavioral niche, and a three-tier evaluation pipeline gives cheap placement-only feedback to every candidate, timing-proxy feedback to selected candidates, and real-router feedback only to finalists. This combination keeps proposals differentiable and executable while still allowing genuinely different objective shapes to compete a

What would settle it

An experiment: take the final champion objectives, remove each evolved term one at a time, and replace the learned schedule variables with the best fixed schedule found by grid search; if wirelength, congestion, and timing gains persist unchanged, the evolved structure is not doing the work. A simpler calculation is to compare the champions term-by-term against RUDY-plus-density baselines: if they are scalar multiples of those known terms within numerical tolerance, the structural-discovery claim is not supported.

Watch

Extended reading notes

Core claim

The paper claims that a closed loop—LLM proposes a differentiable objective, DREAMPlace executes it, routing and timing tools report downstream quality, and an archive conditions the next proposal—can evolve analytical placement objectives that beat both the native placer and prior automated methods on downstream metrics. The search is restricted to a compact symbolic interface, so the output is a readable differentiable program with schedule variables, not a black-box predictor. The evidence: on eight ChiPBench Nangate45 designs, routed wirelength improves by 16.9%, congestion by 36.7%, worst negative slack by 0.70 ns, and total negative slack by 912 ns; leave-one-design-out evolution retai

Load-bearing premise

The gains are credited to the LLM-evolved objective structure, but the final objective programs are never disclosed; if those programs reduce to a scheduled, reweighted version of already-known routing-pressure, density, and wirelength terms, then the claimed objective discovery narrows to constrained coefficient search.

Editorial extensions

If this is right

Placement objectives can be treated as a searchable program space rather than a fixed human artifact, so each new design library or technology node can in principle get an objective evolved for it.
Because evolved objectives are readable and embed in an existing analytical placer, designers can inspect what the placer is optimizing instead of accepting a black-box surrogate.
The cost-scaled evidence policy shows that downstream-aware objective search is feasible without routing every candidate, since cheap placement feedback is used broadly and expensive routing feedback is reserved for selected candidates.
Timing proxies can be admitted or gated based on measured redundancy with HPWL, giving a principled way to decide when a cheap placement-stage timing estimate can be trusted.
Zero-shot transfer of the best objective to different designs and technology families suggests the discovered objectives encode general routing and timing regularities, not per-design overfitting.

Reading between the lines

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

The paper never shows the final champion objective programs, so a natural next step is publishing them; if they turn out to be scheduled versions of known RUDY/density penalties, the contribution would be better described as automated coefficient calibration than structural objective discovery.
The same archive-conditioned proposal loop could be tried on other hand-tuned stages in physical design, such as detailed routing parameters or clock-tree synthesis, where differentiable feedback is sparse.
A direct way to test the LLM's contribution is to fix the archive and feedback but replace the LLM with random grammar sampling; the reported ablation suggests a large gap, implying the proposal model's inductive bias matters, not just the archive.
One testable extension is to make the objective template less restrictive—allow the LLM to introduce new differentiable terms beyond the fixed density/routing/pin schedule—and see whether gains grow, which would reveal whether current results are capped by the template.

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.

Desk Editor's Note

read the letter

Quick take: this is a solid empirical paper that will likely reproduce its headline gains, but the strong framing—evolving placement objectives—is only partially earned. The search is over the time-varying weights/schedules in a fixed linear combination of wirelength, density, routing pressure, and pin pressure (Eq. 7). That is a real and useful contribution, but it is schedule evolution, not discovery of new objective structure.

What is genuinely new: prior work (EvoPlace) fixes the placement objective and searches optimizer components; learned surrogates are black boxes. CoEvoP&R is the first to put an LLM in the loop over the objective's scheduling coefficients, with downstream routing/timing feedback and a MAP-Elites archive. The experimental discipline is a cut above typical ML-for-EDA papers: three seeds with spreads, internal arithmetic checks out, a 12-configuration ablation isolates the archive, cost-scaled evaluation, timing-proxy audit, LODO and zero-shot transfer rows. I recomputed the headline means from Table 1; they are consistent. The gains over native DREAMPlace and the cited baselines are large and mostly meaningful.

Soft spots, in order of severity. First, the central artifact is never shown: the final champion programs are not disclosed, and Figure 2 is explicitly labeled illustrative. Combined with Eq. (7), readers cannot tell whether the champions are novel mechanisms or tuned combinations of known surrogates. Second, there is an internal tension: the Figure 2 illustrative program uses log1p(route), which is not representable in Eq. (7) unless the template is not the full lowering target. That needs one sentence of clarification. Third, Table 2 reports no error bars for Superblue/ASAP7 rows, and the abstract leads with the -E (same-design evolution) numbers while the LODO/zero-shot rows carry the generalization weight. Minor: reproduction depends on a proprietary LLM API and the repo lacks a commit hash.

None of these sink the paper. The empirical claim—LLM-guided search over scheduled objective weights with routing feedback beats strong baselines—stands. But the novelty claim should be reframed to match what the search actually explores. A serious referee should ask for the champion programs or a representative sample, and for the exact relationship between Eq. (7) and the grammar.

Who gets value: anyone working on LLM-driven program evolution for physical design, and EDA folks skeptical of black-box surrogates. I would send it to review; with those clarifications it would be a solid contribution.

Referee Report

2 major / 5 minor

Summary. The paper proposes CoEvoP&R, an LLM-driven evolutionary framework that searches for differentiable placement objective functions and embeds them into DREAMPlace. The loop is archive-conditioned: an LLM proposes candidates from a restricted grammar, candidates are validated and embedded, a cost-scaled evaluation tier estimates placement, timing-proxy, and routed metrics, and a MAP-Elites archive plus negative memory guides subsequent proposals. The headline claim is that, across eight ChiPBench Nangate45 designs, CoEvoP&R reduces post-route routed wirelength and congestion by 16.9% and 36.7% over native DREAMPlace while improving worst negative slack by 0.70 ns and total negative slack by 912 ns; out-of-sample support comes from leave-one-design-out results and zero-shot transfer to Superblue and ASAP7. The paper also reports ablations isolating archive conditioning, cost-scaled evaluation, timing-proxy gating, and LLM-model sensitivity.

Significance. If the results hold, the contribution is significant for both EDA and LLM-guided program evolution: it is, to my knowledge, the first system that evolves the analytical placement objective itself rather than hyperparameters, optimizer components, or macro actions, and it validates the evolved objectives against real downstream routing and timing. The empirical package is stronger than typical for this area: the ablation includes random-DSL and one-shot LLM controls; LODO and zero-shot transfer rows directly address overfitting concerns; and the main means in Tables 1 and 2 are internally consistent with the stated spreads. The framework's cost-scaled evidence hierarchy and archive-guided proposal design are also sensible and well motivated. The main limitation is a central ambiguity about what the LLM actually searches over: the fixed template in Eq. (7) may reduce the contribution from structural objective discovery to constrained coefficient-schedule tuning.

major comments (2)

[§4.1–§4.2, Eq. (7), Fig. 2] There is an unresolved contradiction between the general objective form in Eq. (6) and the runtime objective stated in Eq. (7). Eq. (7) fixes L_theta,t = eW + h_den*D + h_route*R + h_pin*P, with only the coefficient/smoothing schedules evolving. Under that template, every admitted LLM proposal is reduced to a time-varying linear combination of four fixed terms. Yet the illustrative program in Fig. 2 contains 'score = wl + den + 0.01*log1p(route)', which cannot be represented as h_route*R for a single fixed routing-pressure term R, and Eq. (6) allows arbitrary differentiable composition F_theta. These statements cannot all be true. If Eq. (7) is the actual lowering target, then the headline claim of 'evolving analytical placement objectives' should be reframed as 'evolving coefficient/smoothing schedules for a fixed objective template'—still useful, but a materially weaker contribution. I
[§5.2, Table 1 (CoEvoP&R-E row)] Even accepting the empirical gains, the absence of the discovered objective programs makes the 'readable, inspectable objective' claim unverifiable. The paper states that Figure 2's program is 'illustrative, not the selected final objective', and no final program is given in text, appendix, or code listing. If the champions are, say, scheduled RUDY-weighted density/HPWL mixes, then the framework demonstrates strong LLM-guided schedule search but not the discovery of structurally new objective mechanisms. I am not asking for a different contribution; I am asking that the claims in the abstract and Contribution 1 be made consistent with the actual search space, and that the selected programs be disclosed so the mechanism is falsifiable.

minor comments (5)

[§4.2, Eq. (8)] The complement notation appears as 'Y_d' and 'Y_d' without a visible complement symbol; the intended meaning is clear but the typesetting should be fixed.
[§4.4, Algorithm 1] Algorithm 1 returns Best(A_G), but the selection criterion for 'best' is not defined. The Pareto order in Eq. (9) is a partial order; clarify how a single champion is chosen (e.g., coverage, runtime, or a weighted scalarization).
[§3.2, Eq. (4)] The p-norm exponent p and epsilon are free parameters, but the paper does not state how they are set or whether they are part of the search. Please state the defaults and whether the LLM can propose them.
[§5.4] The timing-proxy audit in Fig. 4 reports only four designs and aggregate admission counts. Given that the gating rule is a component of the ablation study, a small table with per-design Pearson/Spearman correlations and admission decisions would be easier to verify.
[§5.5, Table 3] The text says that 'stronger LLMs improve validity and score under the same prompts, mutation schedules, and budgets', but no details on the mutation/schedule hyperparameters (K, G, island migration rate, seed count) are given anywhere. At least list the evolution budget and K/G values in the experimental setup.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the empirical claims are supported by external benchmarks, measured routed/timing metrics, and out-of-sample LODO/zero-shot evaluation.

full rationale

CoEvoP&R is an empirical search method rather than a derivation, and its main claims are supported by external benchmarks (ChiPBench, Superblue, ASAP7), external baseline tools, and actual downstream routing/timing evaluation through OpenROAD and OpenTimer. The reported improvements are measured outcomes after embedding candidate objectives in DREAMPlace, not quantities that are equal to the prompt inputs or to fitted parameters by construction. The LODO protocol (CoEvoP&R-L) and zero-shot transfer (CoEvoP&R-T) provide out-of-sample evidence, so the results are not merely re-reporting the same designs used to steer the search. There is no author self-citation chain: the cited methods and baselines are by other research groups, and no load-bearing uniqueness or existence theorem is imported from the authors' own prior work. The main limitation—that every embedded objective is expressed through the fixed linear template of Eq. (7), so the search may reduce to schedule/coefficient tuning over four known terms—weakens or narrows the contribution but is not a circularity: it does not make the measured post-route wirelength, congestion, WNS, or TNS gains identical to the inputs by construction. The undisclosed final champion programs and the illustrative log1p(route) program being inconsistent with Eq. (7) are verification concerns, not circular-reasoning steps. Therefore the paper exhibits no significant circularity.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The framework is a search procedure rather than a derivation, so the ledger records what the search assumes and what it fits. The evolved schedule coefficients are free parameters fitted by selection on benchmark designs — their values are simultaneously the core contribution and entirely undisclosed. The objective template (Eq. 7) and the adequacy of its terms (wirelength, density, routing pressure, pin pressure) are domain assumptions, partially checked by the ablation but not derived. The LLM's proposal quality (validity 68%, admission 12%) is an unstated external dependency. Benchmark representativeness and the OpenROAD/OpenTimer flow standing in for signoff are further domain assumptions, acknowledged in the conclusion. No new physical entities are introduced.

free parameters (4)

Evolved schedule coefficients h_den, h_route, h_pin, h_gamma (Eq. 7) = not disclosed — final evolved objectives never shown in the paper
The LLM proposes time-varying weights for density, routing-pressure, pin-pressure, and wirelength-smoothing terms; candidates are admitted via Pareto evidence on the benchmark designs. In CoEvoP&R-E mode these coefficients are fitted to the same designs whose results are reported. No values appear in the text (Figure 2's example is labeled illustrative).
Hotspot p-norm exponent p and epsilon in the reduction rho_p(M) (Eq. 4) = not reported
The p>1 hotspot-sensitivity and eps>0 well-definedness constants shape every map-observable reduction, but the paper gives no chosen values and no sensitivity analysis for them.
Evolution hyperparameters: K samples per generation, G generations, island-migration rate, Pareto tie-break order = not reported
Algorithm 1 references K and G but never states their values; only aggregate cost appears (24 Tier-C evals, 32 GPU-hours on the four-design stress panel). These choices control how much selection pressure the archive applies.
Timing-proxy perturbation amplitude (100–2000 DBU) = 100–2000 database units
The audit in §5.4 perturbs placement terms by 100–2000 DBU; the chosen amplitude range determines the reported Pearson/Spearman redundancy values and admission outcomes.

assumptions (6)

domain assumption Gradient descent on the smooth objective (Eq. 1) with the electrostatic density model and a RUDY-like routing-demand term adequately represents the physics that determines downstream routed wirelength, congestion, and timing.
The closed loop assumes placement-stage differentiable terms predict post-route outcomes well enough to guide search. The paper's own §5.4 audit shows the timing proxy is admitted on only 3 of 4 designs and correlates with HPWL (Pearson 0.58–0.82), so this assumption is only partially supported.
domain assumption The restricted objective template of Eq. (7) — a weighted sum of wirelength, density, routing pressure, pin pressure with schedule variables — spans objectives that dominate the baselines.
If true improvement mechanisms require terms outside this template (non-separable couplings, richer net-weighting than timing-criticality features), the search cannot find them. No final objective is disclosed to verify the template's products, so this premise is uncheckable from the text.
ad hoc to paper gpt-5.4 (and the weaker tested models) sample meaningful, diverse programs from the DSL grammar at a high enough rate for evolution to progress.
Valid/Acc. = 68/12 means most LLM proposals are invalid or rejected; the method's efficiency depends on unspecified abilities of one proprietary API, and the model-family ablation shows strong model dependence (score 4.87 for gpt-5.4 vs 3.24 for qwen3.7-max).
domain assumption ChiPBench Nangate45, ICCAD 2015 Superblue, and ASAP7 are representative of placement-routing quality for general industrial designs.
The claims are benchmark-scoped; cross-technology generalization is asserted from three node/benchmark families with 8+8+3 designs. Transfer retention of 46–80% on Superblue is evidence, not proof, of generality.
domain assumption OpenROAD routing plus OpenTimer timing analysis is a faithful proxy for signoff PPA.
The paper calls the outputs 'post-route' and uses them for all claims, but the flow is the open-source toolchain; the conclusion explicitly defers signoff-accurate routing and multi-corner timing to future work.
standard math The smoothed wirelength and density functions of DREAMPlace are differentiable with finite gradients throughout placement, and the density state can safely drive the placer preconditioner.
Invoked in §4.2 as the 'differentiable placement backbone'; the framework inherits these properties from the cited DREAMPlace and ePlace machinery rather than proving them.

how reviews work

0 comments

Cite this review

Pith. "Pith review of CoEvoP&R: Co-Evolving Placement Objectives with Routing Feedback via Large Language Models." pith.science (2026). https://pith.science/paper/R6S73UMF

@misc{pith2026260717398,
  author       = {Pith},
  title        = {Pith review of: CoEvoP&R: Co-Evolving Placement Objectives with Routing Feedback via Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/R6S73UMF}},
  note         = {Machine review of arXiv:2607.17398}
}

read the original abstract

Analytical placers rely on differentiable objective functions to guide placement, typically combining intermediate surrogate metrics such as half-perimeter wirelength (HPWL) and cell-density penalties. However, these placement-stage surrogates remain misaligned with downstream routed and timing quality. Prior work reduces this gap with human-designed terms or learned black-box surrogates, but the former requires expert retuning and the latter is difficult to explain, debug, or deploy in analytical placement flows. CoEvoP&R addresses these limitations with a large language model (LLM)-based framework that automatically evolves analytical placement objectives. At each generation, the prompt combines the restricted objective interface, baseline context, and archived prior candidates with routing-related feedback from placement, timing proxy, and routing tools. The LLM proposes readable differentiable objectives, which are embedded and validated in DREAMPlace, evaluated through a timing proxy and an actual router, and stored with their feedback to guide later generations. Across eight ChiP-Bench Nangate45 designs and three seeds, CoEvoP&R reduces post-route routed wirelength and congestion by 16.9% and 36.7%, with gains of 0.70 ns in worst negative slack and a 912 ns reduction in total negative slack magnitude over native DREAMPlace. Across eight ICCAD 2015 Superblue designs, it reduces post-route routed wirelength and congestion by 5.4% and 23.2%. Code is available at https://github.com/FCHXWH823/CoEvoP-R.git.

Figures

Figures reproduced from arXiv: 2607.17398 by the authors.

**Figure 1.** CoEvoP&R evolves complete placement objectives through archive-conditioned proposal, DREAMPlace embedding, cost-scaled [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

**Figure 2.** Archive-conditioned proposal contract. 𝑃 is a pin pressure term. The state variables ℎ den 𝑡 , ℎ route 𝑡 , ℎ pin 𝑡 , and ℎ 𝛾 𝑡 determine the relative weights and the active smoothing scale. This formulation makes archive evidence actionable through complete objectives whose coefficients and smoothing behavior evolve with placement progress. 4.2 Candidate Validation and DREAMPlace Embedding The proposal model produce… view at source ↗

**Figure 4.** Timing-proxy admission audit. Grouped bars show median [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗

Figures from the paper (1 more)

**Figure 3.** Figure 3: Superblue16 qualitative placement comparison. Metrics [PITH_FULL_IMAGE:figures/full_fig_p006_3.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

26 extracted references · 1 canonical work pages

[1]

Anthony Agnesina, Rongjian Liang, Geraldo Pradipta, Anand Rajaram, and Haox- ing Ren. 2025. GOALPlace: Begin with the End in Mind. InProceedings of the International Symposium on Physical Design. ACM, New York, NY, USA, 2–10. doi:10.1145/3698364.3705342

arXiv 2025
[2]

Anthony Agnesina, Puranjay Rajvanshi, Tian Yang, Geraldo Pradipta, Austin Jiao, Ben Keller, Brucek Khailany, and Haoxing Ren. 2023. AutoDMP: Au- tomated DREAMPlace-based Macro Placement. InProceedings of the Interna- tional Symposium on Physical Design. ACM, New York, NY, USA, 149–157. doi:10.1145/3569052.3578923

arXiv 2023
[3]

Chhabria, Mateus Fogaça, Soheil Hashemi, Abdelrahman Hosny, Andrew B

Tutu Ajayi, Vidya A. Chhabria, Mateus Fogaça, Soheil Hashemi, Abdelrahman Hosny, Andrew B. Kahng, Minsoo Kim, Jeongsup Lee, Uday Mallappa, Marina Neseem, Geraldo Pradipta, Sherief Reda, Mehdi Saligane, Sachin S. Sapatnekar, Carl Sechen, Mohamed Shalan, William Swartz, Lutong Wang, Zhehong Wang, Mingyu Woo, and Bangqi Xu. 2019. Toward an Open-Source Digita...

2019
[4]

Bergstra, Rémi Bardenet, Yoshua Bengio, and Balázs Kégl

James S. Bergstra, Rémi Bardenet, Yoshua Bengio, and Balázs Kégl. 2011. Algo- rithms for Hyper-Parameter Optimization. InAdvances in Neural Information Processing Systems, Vol. 24. Curran Associates, Inc., Granada, Spain, 2546–2554

2011
[5]

Magi Chen and Ting-Chi Wang. 2025. HyperPlace: Harnessing a Large Language Model for Efficient Hyperparameter Optimization in GPU-Accelerated VLSI Placement.ACM Transactions on Design Automation of Electronic Systems30, 6, Article 98 (2025), 27 pages. doi:10.1145/3733601

work page doi:10.1145/3733601 2025
[6]

Kahng, Ilgweon Kang, and Lutong Wang

Chung-Kuan Cheng, Andrew B. Kahng, Ilgweon Kang, and Lutong Wang. 2019. RePlAce: Advancing Solution Quality and Routability Validation in Global Place- ment.IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems38, 9 (2019), 1717–1730. doi:10.1109/TCAD.2018.2859220

arXiv 2019
[7]

Clark, Vinay Vashishtha, Lucian Shifren, Aditya Gujja, Saurabh Sinha, Brian Cline, Chandarasekaran Ramamurthy, and Greg Yeric

Lawrence T. Clark, Vinay Vashishtha, Lucian Shifren, Aditya Gujja, Saurabh Sinha, Brian Cline, Chandarasekaran Ramamurthy, and Greg Yeric. 2016. ASAP7: A 7-nm FinFET Predictive Process Design Kit.Microelectronics Journal53 (2016), 105–115. doi:10.1016/j.mejo.2016.04.006

doi:10.1016/j.mejo.2016.04.006 2016
[8]

Zijie Geng, Jie Wang, Ziyan Liu, Siyuan Xu, Zhentao Tang, Shixiong Kai, Mingx- uan Yuan, Jianye Hao, and Feng Wu. 2025. LaMPlace: Learning to Optimize Cross-Stage Metrics in Macro Placement. InInternational Conference on Learning Representations. OpenReview.net, Singapore

2025

Show all 26 references

[9]

Yunbo Hou, Haoran Ye, Yingxue Zhang, Siyuan Xu, and Guojie Song. 2024. RoutePlacer: An End-to-End Routability-Aware Placer with Graph Neural Net- work. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. ACM, New York, NY, USA, 1085–1095. do...

2024 doi
[11]

Myung-Chul Kim, Jin Hu, Jiajia Li, and Natarajan Viswanathan. 2015. ICCAD- 2015 CAD Contest in Incremental Timing-Driven Placement and Benchmark Suite. InProceedings of the IEEE/ACM International Conference on Computer-Aided Design. IEEE, Austin, TX, USA, 921–926. doi:10.1109/...

2015
[12]

Peiyu Liao, Siting Liu, Zhitang Chen, Wenlong Lv, Yibo Lin, and Bei Yu. 2022. DREAMPlace 4.0: Timing-Driven Global Placement with Momentum-Based Net Weighting. InProceedings of the Design, Automation and Test in Europe Conference and Exhibition. IEEE, Antwerp, Belgium, 939–944...

2022
[13]

Yibo Lin, Shounak Dhar, Wuxi Li, Haoxing Ren, Brucek Khailany, and David Z. Pan. 2019. DREAMPlace: Deep Learning Toolkit-Enabled GPU Acceleration for Modern VLSI Placement. InProceedings of the 56th Annual Design Automation Conference. ACM, New York, NY, USA, 1–6. doi:10.1145/...

2019
[14]

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
[15]

Huang, Chin- Chi Teng, and Chung-Kuan Cheng

Jingwei Lu, Pengwen Chen, Chin-Chih Chang, Lu Sha, Dennis J.-H. Huang, Chin- Chi Teng, and Chung-Kuan Cheng. 2014. ePlace: Electrostatics Based Placement Using Nesterov’s Method. InProceedings of the 51st Annual Design Automation Conference. ACM, New York, NY, USA, 1–6. doi:10...

2014
[16]

Yecheng Jason Ma, William Liang, Guanzhi Wang, De-An Huang, Osbert Bastani, Dinesh Jayaraman, Yuke Zhu, Jim Fan, and Anima Anandkumar. 2024. Eureka: Human-Level Reward Design via Coding Large Language Models. InInternational Conference on Learning Representations. OpenReview.n...

2024
[17]

Le, James Laudon, Richard Ho, Roger Carpenter, and Jeff Dean

Azalia Mirhoseini, Anna Goldie, Mustafa Yazgan, Joe Wenjie Jiang, Ebrahim Songhori, Shen Wang, Young-Joon Lee, Eric Johnson, Omkar Pathak, Azade Nova, Jiwoo Pak, Andy Tong, Kavya Srinivasa, William Hang, Emre Tuncer, Quoc V. Le, James Laudon, Richard Ho, Roger Carpenter, and J...

2021 doi
[18]

Jean-Baptiste Mouret and Jeff Clune. 2015. Illuminating Search Spaces by Mapping Elites.CoRRabs/1504.04909 (2015). doi:10.48550/arXiv.1504.04909

2015 doi arXiv
[19]

Alexander Novikov, Ngân V ˜u, Marvin Eisenberger, Emilien Dupont, Po-Sen Huang, Adam Zsolt Wagner, Sergey Shirobokov, Borislav Kozlovskii, Francisco J. R. Ruiz, Abbas Mehrabian, M. Pawan Kumar, Abigail See, Swarat Chaudhuri, George Holland, Alex Davies, Sebastian Nowozin, Push...

2025 arXiv
[20]

Pawan Kumar, Emilien Dupont, Francisco J

Bernardino Romera-Paredes, Mohammadamin Barekatain, Alexander Novikov, Matej Balog, M. Pawan Kumar, Emilien Dupont, Francisco J. R. Ruiz, Jordan S. Ellenberg, Pengming Wang, Omar Fawzi, Pushmeet Kohli, and Alhussein Fawzi
[21]

Asankhaya Sharma. 2025. OpenEvolve: An Open-Source Evolutionary Coding Agent. https://github.com/algorithmicsuperintelligence/openevolve

2025
[22]

Parshin Shojaee, Kazem Meidani, Shashank Gupta, Amir Barati Farimani, and Chandan K. Reddy. 2025. LLM-SR: Scientific Equation Discovery via Programming with Large Language Models. InInternational Conference on Learning Representa- tions. OpenReview.net, Singapore

2025
[23]

Johannes

Peter Spindler and Frank M. Johannes. 2007. Fast and Accurate Routing Demand Estimation for Efficient Routability-Driven Placement. InProceedings of the Con- ference on Design, Automation and Test in Europe. IEEE, Nice, France, 1226–1231. doi:10.1109/DATE.2007.364463

2007
[24]

Ikechukwu Uchendu, Swati Goel, Karly Hou, Ebrahim Songhori, Kuang-Huei Lee, Joe Wenjie Jiang, Vijay Janapa Reddi, and Vincent Zhuang. 2026. See It to Place It: Evolving Macro Placements with Vision-Language Models. arXiv preprint arXiv:2603.28733

2026 arXiv
[25]

Zhihai Wang, Zijie Geng, Zhaojie Tu, Jie Wang, Yuxi Qian, Zhexuan Xu, Ziyan Liu, Siyuan Xu, Zhentao Tang, Shixiong Kai, Mingxuan Yuan, Jianye Hao, Bin Li, and Feng Wu. 2025. Benchmarking End-To-End Performance of AI-Based Chip Placement Algorithms. InAdvances in Neural Informa...

2025
[26]

Xufeng Yao, Jiaxi Jiang, Yuxuan Zhao, Peiyu Liao, Yibo Lin, and Bei Yu. 2026. EvoPlace: Evolution of Optimization Algorithms for Global Placement via Large Language Models.IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems(2026). doi:10.1109/TCAD.202...

2026
[2024]

doi:10.1038/s41586-023-06924-6

Mathematical Discoveries from Program Search with Large Language Models.Nature625 (2024), 468–475. doi:10.1038/s41586-023-06924-6

2024 doi

Pith tools

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

[1] [1]

Anthony Agnesina, Rongjian Liang, Geraldo Pradipta, Anand Rajaram, and Haox- ing Ren. 2025. GOALPlace: Begin with the End in Mind. InProceedings of the International Symposium on Physical Design. ACM, New York, NY, USA, 2–10. doi:10.1145/3698364.3705342

arXiv 2025

[2] [2]

Anthony Agnesina, Puranjay Rajvanshi, Tian Yang, Geraldo Pradipta, Austin Jiao, Ben Keller, Brucek Khailany, and Haoxing Ren. 2023. AutoDMP: Au- tomated DREAMPlace-based Macro Placement. InProceedings of the Interna- tional Symposium on Physical Design. ACM, New York, NY, USA, 149–157. doi:10.1145/3569052.3578923

arXiv 2023

[3] [3]

Chhabria, Mateus Fogaça, Soheil Hashemi, Abdelrahman Hosny, Andrew B

Tutu Ajayi, Vidya A. Chhabria, Mateus Fogaça, Soheil Hashemi, Abdelrahman Hosny, Andrew B. Kahng, Minsoo Kim, Jeongsup Lee, Uday Mallappa, Marina Neseem, Geraldo Pradipta, Sherief Reda, Mehdi Saligane, Sachin S. Sapatnekar, Carl Sechen, Mohamed Shalan, William Swartz, Lutong Wang, Zhehong Wang, Mingyu Woo, and Bangqi Xu. 2019. Toward an Open-Source Digita...

2019

[4] [4]

Bergstra, Rémi Bardenet, Yoshua Bengio, and Balázs Kégl

James S. Bergstra, Rémi Bardenet, Yoshua Bengio, and Balázs Kégl. 2011. Algo- rithms for Hyper-Parameter Optimization. InAdvances in Neural Information Processing Systems, Vol. 24. Curran Associates, Inc., Granada, Spain, 2546–2554

2011

[5] [5]

Magi Chen and Ting-Chi Wang. 2025. HyperPlace: Harnessing a Large Language Model for Efficient Hyperparameter Optimization in GPU-Accelerated VLSI Placement.ACM Transactions on Design Automation of Electronic Systems30, 6, Article 98 (2025), 27 pages. doi:10.1145/3733601

work page doi:10.1145/3733601 2025

[6] [6]

Kahng, Ilgweon Kang, and Lutong Wang

Chung-Kuan Cheng, Andrew B. Kahng, Ilgweon Kang, and Lutong Wang. 2019. RePlAce: Advancing Solution Quality and Routability Validation in Global Place- ment.IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems38, 9 (2019), 1717–1730. doi:10.1109/TCAD.2018.2859220

arXiv 2019

[7] [7]

Clark, Vinay Vashishtha, Lucian Shifren, Aditya Gujja, Saurabh Sinha, Brian Cline, Chandarasekaran Ramamurthy, and Greg Yeric

Lawrence T. Clark, Vinay Vashishtha, Lucian Shifren, Aditya Gujja, Saurabh Sinha, Brian Cline, Chandarasekaran Ramamurthy, and Greg Yeric. 2016. ASAP7: A 7-nm FinFET Predictive Process Design Kit.Microelectronics Journal53 (2016), 105–115. doi:10.1016/j.mejo.2016.04.006

doi:10.1016/j.mejo.2016.04.006 2016

[8] [8]

Zijie Geng, Jie Wang, Ziyan Liu, Siyuan Xu, Zhentao Tang, Shixiong Kai, Mingx- uan Yuan, Jianye Hao, and Feng Wu. 2025. LaMPlace: Learning to Optimize Cross-Stage Metrics in Macro Placement. InInternational Conference on Learning Representations. OpenReview.net, Singapore

2025