Paper deep dive
CoEvoP&R: Co-Evolving Placement Objectives with Routing Feedback via Large Language Models
Ruogu Chen, Weihua Xiao, Ramesh Karri, Jie Han
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 89%
Last extracted: 7/21/2026, 5:00:41 AM
Summary
The paper introduces CoEvoP&R, an LLM-based framework that automatically evolves analytical placement objective functions for electronic design automation. By integrating routing feedback and timing proxies into a generative loop, CoEvoP&R improves post-route wirelength, congestion, and timing metrics compared to native DREAMPlace and other baselines on ChiP-Bench and ICCAD 2015 Superblue benchmarks.
Entities (10)
Relation Signals (8)
CoEvoP&R → uses → LLM
confidence 95% · CoEvoP&R addresses these limitations with a large language model (LLM)-based framework
CoEvoP&R → validatesin → DREAMPlace
confidence 95% · The LLM proposes readable differentiable objectives, which are embedded and validated in DREAMPlace
CoEvoP&R → evaluateson → ChiP-Bench
confidence 90% · Across eight ChiP-Bench Nangate45 designs... CoEvoP&R reduces post-route routed wirelength
CoEvoP&R → evaluateson → ICCAD 2015 Superblue
confidence 90% · Across eight ICCAD 2015 Superblue designs, it reduces post-route routed wirelength
CoEvoP&R → optimizes → HPWL
confidence 85% · typically combining intermediate surrogate metrics such as half-perimeter wirelength (HPWL)
CoEvoP&R → optimizes → WNS
confidence 85% · gains of 0.70 ns in worst negative slack
CoEvoP&R → optimizes → TNS
confidence 85% · 912 ns reduction in total negative slack magnitude
FunSearch → similarto → CoEvoP&R
confidence 80% · FunSearch... realize this idea by using evaluator feedback to revise executable programs
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
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 this https URL.
Tags
Links
- Source: https://arxiv.org/abs/2607.17398v1
- Canonical: https://arxiv.org/abs/2607.17398v1
Trouble viewing inline? Open PDF directly →
Full Text
46,568 characters extracted from source content.
Expand or collapse full text
CoEvoP&R: Co-Evolving Placement Objectives with Routing Feedback via Large Language Models Ruogu Chen 0009-0009-1114-824X University of AlbertaEdmontonAlbertaCanada ruogu@ualberta.ca , Weihua Xiao 0000-0002-6232-4460 New York UniversityNew YorkNew YorkUnited States wx2356@nyu.edu , Ramesh Karri 0000-0001-7989-5617 New York UniversityNew YorkNew YorkUnited States rkarri@nyu.edu and Jie Han 0000-0002-8849-4994 University of AlbertaEdmontonAlbertaCanada jhan8@ualberta.ca 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 ChiPBench 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. electronic design automation, placement, routing, large language models, objective functions †copyright: none†ccs: Computing methodologies Machine learning†ccs: Hardware Electronic design automation†ccs: Hardware Physical design (EDA) 1. Introduction Modern analytical placers make global placement tractable by minimizing a differentiable placement objective. Existing objectives usually combine intermediate surrogates, such as smooth approximations to half-perimeter wirelength (HPWL) and cell-density penalties, which provide differentiable guidance for tools such as DREAMPlace (Lin et al., 2019). However, ChiPBench shows that placement-stage metrics can remain weakly aligned with downstream timing, including worst negative slack (WNS) and total negative slack (TNS) (Wang et al., 2025). Consequently, optimizing placement surrogates does not necessarily optimize the final power, performance, and area (PPA) metrics of the completed physical-design flow. Prior work narrows this gap by combining the placement objective with human-designed terms or learned black-box surrogates for the routing stage. Human-designed terms, such as Rectangular Uniform wire DensitY (RUDY), estimate routing demand from placement geometry (Spindler and Johannes, 2007), and subsequent methods incorporate related congestion and pin-density information. These terms improve routing awareness, but they rely on expert-designed approximations and often require retuning across designs or technologies. Learned predictors, post-route-informed density targets, and differentiable routability surrogates, as used by LaMPlace (Geng et al., 2025), GOALPlace (Agnesina et al., 2025), and RoutePlacer (Hou et al., 2024), respectively, improve cross-stage awareness. However, black-box surrogates can be difficult to explain, debug, and deploy inside existing analytical placement flows. This leaves a complementary path that searches directly over readable, differentiable objective forms. Large language models (LLMs) have shown potential in directing the search for programs that satisfy or optimize specified targets. FunSearch (Romera-Paredes et al., 2024), Eureka (Ma et al., 2024), LLM-SR (Shojaee et al., 2025), and AlphaEvolve (Novikov et al., 2025) realize this idea by using evaluator feedback to revise executable programs toward measured objectives. EvoPlace (Yao et al., 2026) and VeoPlace (Uchendu et al., 2026) further apply foundation models to placement by directing the search over optimizer components or macro-placement actions. Instead, we propose CoEvoP&R, an LLM-based framework that automatically evolves the analytical placement objective. For each generation, CoEvoP&R assembles a prompt from the restricted objective interface, baseline placement context, archive records of earlier objective variants, and measured feedback from placement-stage, timing-signal, and routed evaluations. The prompt guides the LLM to generate the next candidate objectives. CoEvoP&R embeds and validates each admitted candidate in the analytical placer, schedules downstream routed evaluation for selected candidates, and stores the paired objective and measured feedback in the archive for subsequent generations. CoEvoP&R makes four novel contributions. • LLM objective evolution. We present the first automated LLM-based framework that evolves analytical placement objective functions integrated with real placement and routing tools. • Automated validation. CoEvoP&R embeds and validates admitted objective programs inside DREAMPlace through a restricted interface, ensuring that generated objectives remain differentiable, bounded, and executable in an analytical placement flow. • Routing-aware feedback. CoEvoP&R evaluates evolved candidates with routing-tool feedback, including timing-related behavior, so the objective search is guided by downstream timing-related information. • Archive-guided iteration. CoEvoP&R stores each evolved candidate together with its placement, routing, timing, and failure feedback, then uses this archive to condition later prompts and guide subsequent generations. Across eight ChiPBench Nangate45 designs, CoEvoP&R reduces routed wirelength and congestion by 16.9% and 36.7% over native DREAMPlace, while improving post-route WNS by 0.70 ns and TNS by 912 ns. Across eight Superblue designs, CoEvoP&R reduce rWL and Cong. by 5.4% and 23.2% over the same baseline. 2. Related Work Analytical and Routability-Aware Placement. ePlace developed a nonlinear placement framework based on electrostatics (Lu et al., 2014). RePlAce advanced its optimization and routability validation, and DREAMPlace accelerated analytical placement with GPU-based PyTorch execution (Cheng et al., 2019; Lin et al., 2019). Timing-driven placers add timing guidance while retaining predefined objective structures (Liao et al., 2022; Ajayi et al., 2019). Routability-aware placers incorporate downstream information through routing estimators or learned guidance. RUDY estimates routing demand from placement geometry (Spindler and Johannes, 2007). LaMPlace learns cross-stage metric masks for sequential macro placement, GOALPlace derives density targets from post-route feedback, and RoutePlacer integrates a differentiable graph neural network routability surrogate into analytical placement (Geng et al., 2025; Agnesina et al., 2025; Hou et al., 2024). These methods improve downstream awareness through estimators, masks, targets, or surrogates, while leaving the complete symbolic objective form outside the search. Automated Placement Search. Automated placement methods differ in the artifact they search. Reinforcement-learning methods learn sequential macro actions, AutoDMP and HyperPlace tune placer parameters, and VeoPlace uses a vision-language model to guide evolutionary macro placement (Mirhoseini et al., 2021; Agnesina et al., 2023; Chen and Wang, 2025; Uchendu et al., 2026). EvoPlace is the closest placement-specific example of LLM-guided program evolution. It searches initialization, preconditioning, and optimizer-update components under HPWL feedback while retaining the placement objective (Yao et al., 2026). These works show that automated search can improve actions, parameters, and optimizer behavior, leaving differentiable objective-form search under downstream physical-design feedback as a distinct setting. Executable Program Evolution. FunSearch and Evolution of Heuristics evolve executable heuristics, LLM-SR searches symbolic equations, and AlphaEvolve extends evaluator-guided evolution to broader code artifacts (Romera-Paredes et al., 2024; Liu et al., 2024; Shojaee et al., 2025; Novikov et al., 2025). OpenEvolve provides an open implementation of this pattern (Sharma, 2025). Eureka is methodologically close because generated reward functions execute inside a learning loop and later proposals use training feedback (Ma et al., 2024). Differentiable placement introduces additional requirements because accepted expressions must produce finite values and gradients, adapt to placement trajectories, and justify expensive downstream evaluation. Figure 1. CoEvoP&R evolves complete placement objectives through archive-conditioned proposal, DREAMPlace embedding, cost-scaled evaluation, and archive feedback. A large language model proposes objective programs, validation embeds admitted candidates into DREAMPlace, cost-scaled evaluations measure placement, timing-signal, and routed evidence, and an archive reconstructs feedback for later proposals. 3. Preliminaries 3.1. Analytical Placers and Cross-Stage Metrics Chip placement assigns physical locations to circuit objects inside a fixed die. A netlist can be modeled as a hypergraph G=(V,E)G=(V,E), with vertices V representing macros, standard cells, and ports, and hyperedges E representing nets. A placement is the coordinate set =(xi,yi)i∈Vx=\(x_i,y_i)\_i∈ V after applying object sizes and fixed pin offsets. Analytical placers optimize these coordinates with differentiable objectives so that gradient-based updates can move many objects jointly. DREAMPlace performs global placement with a smooth wirelength and electrostatic-density objective (Lu et al., 2014; Lin et al., 2019). At placement step t, its native objective can be written as (1) ℒDP,t()=W~γt()+λtD(),L_DP,t(x)= W_ _t(x)+ _tD(x), where W~γt W_ _t is a differentiable wirelength approximation, D is the electrostatic density penalty, γt _t controls wirelength smoothing, and λt _t balances wire shortening against density spreading. The optimizer updates cell coordinates using the current gradient ∇ℒDP,t _xL_DP,t. Across placement, the native DREAMPlace objective forms a scheduled objective trajectory through the schedule variables γt _t and λt _t. After placement, HPWL is commonly used to evaluate the placement quality by: (2) HPWL()=∑e∈E[maxi∈exi−mini∈exi+maxi∈eyi−mini∈eyi].HPWL(x)= _e∈ E [ _i∈ ex_i- _i∈ ex_i+ _i∈ ey_i- _i∈ ey_i ]. HPWL decomposes over nets, tracks the bounding box of each net, and is fast to compute. Accordingly, it is widely used as a placement-stage metric. However, placement-stage metrics are useful but incomplete (Wang et al., 2025). The routing demand, pin access, timing paths, buffering, and layer assignment are measured more faithfully by downstream evaluations. Global routing reports routed wirelength and routing overflow. Timing analysis reports WNS and TNS. These cross-stage metrics are closer to final design quality but require much more expensive physical design flows. During analytical placement, the optimizer instead operates on differentiable observables that are available throughout the placement trajectory. 3.2. Differentiable Placement Observables A placement observable is a differentiable signal computed from the current placement. Analytical placers partition the layout region into bins ℬB, and b∈ℬb indexes one bin. An observable can be either a scalar or a bin map, (3) ϕ()∈ℝ,Φ()=Φb()b∈ℬ.φ(x) , (x)=\ _b(x)\_b . The gradients ∇ϕ _xφ and ∇Φb _x _b are available during placement, allowing the optimizer to update cell coordinates through these derivatives. Scalar observables include smooth wirelength W~() W(x) and density penalties D()D(x). Map observables include density maps Φden() ^den(x), RUDY-like routing demand maps Φroute() ^route(x) (Spindler and Johannes, 2007), pin-density maps Φpin() ^pin(x), and long-net pressure maps Φlong() ^long(x). These maps approximate physical pressure from placement geometry before routing is performed. Map observables are reduced to scalar terms before entering an analytical objective. For a nonnegative map M()=Mb()b∈ℬM(x)=\M_b(x)\_b , a common smooth reduction is (4) ρp(M)=(1|ℬ|∑b∈ℬ(Mb+ϵ)p)1/p, _p(M)= ( 1|B| _b (M_b+ε )^p )^1/p, where p>1p>1 controls hotspot sensitivity and ϵ>0ε>0 keeps the expression well defined. Larger p values place more weight on high-pressure bins, while smaller values behave closer to an average. A differentiable placement objective can therefore combine scalar observables and reduced map observables as (5) ℒ()=F(ϕ1(),…,ϕm(),ρp1(Φ1),…,ρpn(Φn)),L(x)=F ( _1(x),…, _m(x), _p_1( _1),…, _p_n( _n) ), where F is a differentiable composition of these terms. 4. Method CoEvoP&R searches for readable differentiable placement objectives that run inside an analytical placer while the placer, benchmarks, and downstream evaluation flow remain fixed. Each candidate θ is a compact symbolic objective that generalizes the static composition of Equation 5 into a scheduled one, (6) 0=Iθ(0),t+1=Uθ(t,t),ℒθ,t(t,d)=Fθ((t,d),t).h_0=I_θ(a_0), _t+1=U_θ(h_t,a_t), _θ,t(x_t,d)=F_θ\! (T(x_t,d),h_t ). Here (t,d)T(x_t,d) collects differentiable placement terms, th_t carries the time-varying coefficients we call schedule variables, and ta_t carries simple progress signals such as the current density overflow and the recent wirelength trend. A candidate is therefore fully specified by IθI_θ, UθU_θ, and FθF_θ. Across generations, evaluated candidates accumulate in an archive that supplies parents and local evidence for later proposals. Figure 1 organizes this closed loop into three steps: archive-conditioned objective proposal, candidate embedding with cost-scaled evaluation, and archive feedback for the next generation. Key method notation Symbol Meaning Symbol Meaning g evolution generation t placement iteration d design being optimized tx_t placement coordinates at iteration t ℐI objective interface B baseline context σ output schema gA_g MAP-Elites archive at generation g pgp_g parent objective HgH_g local archive evidence MgM_g measured feedback Πg _g prompt packet y proposed program c validated candidate ⊥ failed proposal θ candidate objective th_t schedule variables ta_t progress observables Iθ,Uθ,FθI_θ,U_θ,F_θ init, update, and loss maps of θ (t,d)T(x_t,d) differentiable placement terms (c)m(c) measured evidence for c ψ(c)ψ(c) archive feature mapping z MAP-Elites cell ⪯ Pareto evidence order K samples per generation N negative memory 4.1. Archive-Conditioned Objective Proposal At generation g, the current archive gA_g supplies a parent objective pgp_g and a local evidence set HgH_g. The evidence set contains neighboring elites, component trajectories, metric deltas, and failure lessons from earlier runs. CoEvoP&R pairs this evidence with measured feedback MgM_g and the objective interface ℐI to form the prompt packet Πg=(ℐ,B,pg,Hg,Mg,σ) _g=(I,B,p_g,H_g,M_g,σ), where B gives baseline descriptions and σ gives the required output format. Figure 2 shows this proposal contract. The prompt packet conditions the proposal model, and each response must provide the fields needed to instantiate a trajectory-aware objective. Archive conditioning guides proposals toward objective mechanisms that occupy useful behavioral cells and toward measurements that explain nearby placement improvements or degradations. Each response must therefore provide the three functions of Equation 6: an initialization IθI_θ that sets the schedule variables from design-specific calibration, an update UθU_θ that revises them once per placement iteration, and a loss FθF_θ over the placement terms. The progress vector ta_t exposes the iteration fraction, density overflow, HPWL trend, smoothing position, and per-term calibration constants, so proposed schedules react to measured optimization progress rather than to a fixed clock. The runtime objective embedded in DREAMPlace has the form (7) ℒθ,t _θ,t =W~γt(t)+htdenD(t)+htrouteR(t)+htpinP(t), = W_ _t(x_t)+h^den_tD(x_t)+h^route_tR(x_t)+h^pin_tP(x_t), γt _t =htγγtbase. =h^γ_tγ^base_t. The term W~γt W_ _t is the smoothed wirelength used by the placer, D is the density penalty, R is a differentiable routing pressure term, and P is a pin pressure term. The state variables htdenh^den_t, htrouteh^route_t, htpinh^pin_t, and htγh^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. Figure 2. Archive-conditioned proposal contract. The figure shows prompt packet construction, archive-conditioned proposal sampling, and the required fields of a sampled objective program. 4.2. Candidate Validation and DREAMPlace Embedding The proposal model produces symbolic objectives, while DREAMPlace requires terms, components, and state updates that are bound to placement quantities. CoEvoP&R performs this validation and embedding as the middle stage of Figure 1. Let EmbeddEmbed_d denote the embedding procedure for design d. For a proposed symbolic program y, this procedure either returns a candidate c that can be executed in DREAMPlace or marks the proposal as failed. Formally, (8) Embedd(y)=c,y∈d,⊥,y∈¯d.Embed_d(y)= casesc,&y _d,\\ ,&y∈ Y_d. cases Here dY_d is the set of proposed programs accepted by the grammar, physical-role, differentiability, and runtime checks for design d. The complement ¯d Y_d contains rejected programs, and ⊥ denotes an invalid candidate that is not executed by DREAMPlace. A candidate c contains the initialization map, update map, scalar loss, named components, and schedule variables required by Equations 6 and 7. After embedding, candidate c enters DREAMPlace through the physical roles in Equation 7. It provides the density, routing, pin, and smoothing schedules, while native wirelength and density operators form the differentiable placement backbone. The density state also updates the placer preconditioner, so objective changes affect both the scalar loss and the optimization trajectory. Timing-aware candidates may adjust net weights using placement-time criticality, span, and fanout features. 4.3. Cost-Scaled Evaluation and Evidence Ranking Downstream routing and timing tools provide faithful feedback at substantially higher cost than placement. CoEvoP&R therefore evaluates candidates through a cost-scaled evidence hierarchy. Tier A runs the analytical placer for every validated candidate and records HPWL, overflow, runtime, convergence behavior, and objective traces. Tier B applies a placement-stage timing proxy to selected candidates and records proxy WNS/TNS movement. Tier C schedules the downstream router for real routing and records routed wirelength, routing overflow, and post-route WNS and TNS. The evidence vector for a validated candidate c is (c)=mA(c)⊕aB(c)mB(c)⊕aC(c)mC(c),m(c)=m_A(c) a_B(c)m_B(c) a_C(c)m_C(c), where aBa_B and aCa_C indicate whether the candidate receives the corresponding downstream measurement. Archive selection uses Pareto evidence over the measurements available to each candidate pair. For candidate c, define (9) (c) (c) =(ΔHPWL(c),Δovf(c),−ΔWNS(c),−ΔTNS(c)), = ( _HPWL(c), _ovf(c),- _WNS(c),- _TNS(c) ), ci≺cj c_i c_j ⟺vk(ci)≤vk(cj)∀k∈ij v_k(c_i)≤ v_k(c_j)\ ∀ k _ij ∧vk(ci)<vk(cj)∃k∈ij. \ v_k(c_i)<v_k(c_j)\ ∃ k _ij. The vector (c)v(c) uses lower-is-better coordinates, with timing signs chosen so that WNS and TNS improvements align with HPWL and overflow improvements. The index set ijK_ij contains the coordinates measured for both candidates. We write ci⪯cjc_i c_j when ci≺cjc_i c_j holds or no measured coordinate differs. Structural failures, invalid placements, nonfinite traces, and robustness failures are removed before Pareto ranking. Candidates on the same nondominated front are ordered by design coverage and then by runtime. The timing proxy audit controls how placement-stage timing evidence influences the archive. CoEvoP&R periodically applies controlled perturbations to placement terms, net-weighting choices, and state schedules, then compares the resulting placement-stage WNS/TNS movement with downstream WNS/TNS movement using Pearson and Spearman redundancy checks. A timing proxy that adds stable information beyond HPWL and overflow receives more influence in archive updates. An unstable or redundant timing proxy remains available to objective construction but receives little authority in evidence ranking. Table 1. Main ChiPBench Nangate45 results and intra-family transfer. Entries report three-seed mean improvements over DREAMPlace, which is normalized to zero. rWL and Cong. are reductions in routed wirelength and post-route congestion. WNS and TNS are post-route timing gains in ns. Bold marks best means and underline marks second best. BlackParrot family Processor cores Peripheral and control Mean Method Metric bp_fe bp_be swerv_wrapper or1200 mor1kx ethernet dft68 vga_lcd Mean Analytical, expert, and search baselines DREAMPlace 4.0 (Liao et al., 2022) rWL (%) -5.3±0.3 -1.7±0.2 -3.9±0.3 -4.6±0.4 -2.1±0.2 -0.4±0.1 -1.8±0.2 0.3±0.1 -2.4±0.2 Cong. (%) -2.4±0.3 0.7±0.1 -1.8±0.2 -1.1±0.2 -1.2±0.2 -2.3±0.3 -0.4±0.1 0.4±0.1 -1.0±0.2 WNS (ns) +0.38±0.03 +0.48±0.04 +0.18±0.02 +1.32±0.09 +0.16±0.02 +0.36±0.03 +0.12±0.02 +0.28±0.03 +0.41±0.03 TNS (ns) +720±48 +480±32 +140±12 +940±62 +220±18 +340±24 +70±6 +640±42 +444±28 Best expert rWL (%) -2.4±0.4 3.7±0.3 -0.8±0.2 2.6±0.4 -2.3±0.3 1.9±0.2 -1.2±0.2 4.1±0.4 0.7±0.3 Cong. (%) 4.7±0.4 6.8±0.5 0.9±0.2 4.2±0.4 3.7±0.3 3.4±0.3 1.6±0.2 2.7±0.3 3.5±0.3 WNS (ns) -0.14±0.03 -0.21±0.03 -0.08±0.02 -0.28±0.04 -0.09±0.02 -0.18±0.03 -0.04±0.01 -0.03±0.01 -0.13±0.02 TNS (ns) -22.4±2.6 -34.6±3.4 -12.7±1.8 -178.6±14.2 -27.2±3.2 -68.4±6.4 -8.7±1.4 -13.4±1.8 -46±4.2 BO-DSL rWL (%) -0.7±0.8 3.8±0.7 1.2±0.6 2.9±0.7 -0.4±0.4 4.6±0.8 -0.3±0.4 5.7±0.9 2.1±0.6 Cong. (%) 6.3±0.9 5.4±0.7 2.8±0.6 6.1±0.8 5.4±0.7 4.9±0.6 2.7±0.5 4.2±0.6 4.7±0.7 WNS (ns) -0.06±0.04 +0.04±0.02 -0.09±0.03 -0.13±0.05 -0.01±0.02 -0.03±0.02 +0.01±0.01 +0.03±0.02 -0.03±0.03 TNS (ns) -12.4±3.6 +6.8±2.4 -18.7±4.2 -52.7±7.8 -6.4±2.8 -14.6±3.4 +4.2±1.4 +8.9±2.2 -10.6±3.6 AutoDMP (Agnesina et al., 2023) rWL (%) 7.8±0.5 8.6±0.4 11.3±0.7 6.9±0.4 9.7±0.6 12.7±0.8 5.4±0.3 14.8±0.9 9.7±0.6 Cong. (%) 9.4±0.6 11.7±0.7 7.8±0.5 10.7±0.6 12.9±0.8 13.6±0.8 8.4±0.5 14.2±0.9 11.1±0.7 WNS (ns) +0.27±0.03 +0.28±0.03 +0.08±0.02 +0.59±0.04 +0.14±0.02 +0.34±0.03 +0.05±0.01 +0.07±0.02 +0.23±0.02 TNS (ns) +112±10 +114±12 +30±4 +840±62 +80±8 +324±28 +14±3 +49±6 +195±16 EvoPlace (Yao et al., 2026) rWL (%) 2.4±0.7 6.8±0.6 5.1±0.5 4.3±0.5 9.5±0.7 9.7±0.9 0.9±0.3 14.6±1.2 6.7±0.7 Cong. (%) 1.8±0.6 2.9±0.4 -0.6±0.4 2.4±0.5 0.7±0.3 2.7±0.5 0.3±0.2 3.6±0.6 1.7±0.4 WNS (ns) +0.04±0.02 +0.06±0.02 -0.03±0.02 +0.07±0.02 +0.02±0.01 +0.03±0.02 +0.01±0.01 -0.04±0.02 +0.02±0.02 TNS (ns) +8.4±2.2 +14.7±3.2 -4.2±2.4 +38.4±5.6 +6.3±2.2 +22.6±3.8 +2.8±1.6 -9.7±2.8 +9.9±2.8 LaMPlace (Geng et al., 2025) rWL (%) 33.6±2.1 18.9±1.4 8.9±0.7 6.4±0.5 -2.4±0.3 -4.6±0.5 5.6±0.4 -8.4±0.7 7.3±0.8 Cong. (%) 33.1±2.1 20.6±1.3 4.4±0.3 24.6±1.6 18.4±1.2 22.8±1.4 14.6±0.9 26.4±1.7 20.6±1.3 WNS (ns) +0.05±0.01 +0.19±0.02 +0.04±0.01 +0.22±0.02 +0.13±0.02 +0.28±0.03 +0.06±0.01 +0.24±0.02 +0.15±0.02 TNS (ns) -407±28 +1323±88 +40±4 +680±46 +620±42 +380±26 +48±5 +540±36 +403±28 CoEvoP&R evolved objectives CoEvoP&R-E rWL (%) 32.4±2.2 17.4±1.3 18.4±1.3 13.6±1.0 9.4±0.7 17.4±1.2 5.2±0.4 21.4±1.5 16.9±1.0 Cong. (%) 31.4±2.0 33.6±2.2 68.4±3.8 43.6±2.6 28.4±1.8 36.4±2.3 27.2±1.7 24.6±1.6 36.7±2.0 WNS (ns) +0.62±0.05 +0.88±0.07 +0.42±0.04 +1.18±0.09 +0.38±0.03 +0.32±0.03 +0.72±0.06 +1.07±0.09 +0.70±0.06 TNS (ns) +940±68 +1180±88 +380±30 +1560±118 +540±42 +680±48 +190±16 +1830±136 +912±68 CoEvoP&R-L rWL (%) 25.4±1.9 13.4±1.2 14.4±1.1 10.4±0.8 8.4±0.7 13.4±1.0 4.4±0.3 15.8±1.2 13.2±0.9 Cong. (%) 28.4±2.1 32.4±2.1 56.4±3.6 35.4±2.3 25.4±1.8 30.4±2.0 20.4±1.4 22.4±1.6 31.4±1.9 WNS (ns) +0.46±0.05 +0.64±0.06 +0.32±0.04 +1.02±0.09 +0.21±0.03 +0.24±0.03 +0.48±0.04 +0.68±0.06 +0.51±0.04 TNS (ns) +530±42 +1100±86 +260±22 +1000±76 +380±32 +480±36 +128±12 +1220±92 +637±48 4.4. Archive Update and Closed-Loop Feedback The archive is organized as a MAP-Elites grid over objective behavior (Mouret and Clune, 2015). Each admitted objective is mapped to a feature cell z=ψ(c),z=ψ(c), where ψ summarizes complexity, mechanism family, HPWL behavior, overflow behavior, and timing behavior. To preserve diversity, CoEvoP&R uses an island model in which archive subpopulations exchange elites periodically. Within island r, the archive update is g+1(r)[z]=elite⪯(g(r)[z]∪g(r)[z]),A_g+1^(r)[z]=elite_ (A_g^(r)[z] _g^(r)[z] ), where g(r)[z]C_g^(r)[z] collects the admitted generation-g candidates of island r mapped to cell z, and ⪯ is the Pareto evidence order from Equation 9. Negative memory stores failed candidates and their failure reasons, so later prompt packets avoid recurring invalid mechanisms. Island migration periodically exchanges elites across subpopulations and exposes the proposal model to objectives that succeed under different behavioral niches. Algorithm 1 summarizes the loop that returns a readable objective with state variables, named components, lineage, and measured placement evidence. Algorithm 1 CoEvo objective evolution 1: 0,←Init(seeds,baselines),∅A_0,N (seeds,baselines), 2: for g=0,…,G−1g=0,…,G-1 do 3: (r,pg,Hg,Mg)←SelectParent(g,)(r,p_g,H_g,M_g) (A_g,N) 4: Πg←(ℐ,B,pg,Hg,Mg,σ) _g←(I,B,p_g,H_g,M_g,σ) 5: for k=1,…,Kk=1,…,K do 6: yg,k∼q(⋅∣Πg)y_g,k q(· _g) 7: cg,k←Embedd(yg,k)c_g,k _d(y_g,k) 8: if cg,k=⊥∨Admit(cg,k)=0c_g,k= (c_g,k)=0 then 9: ←∪(yg,k,cg,k)N ∪\(y_g,k,c_g,k)\ 10: else 11: g,k←Evaluate(cg,k)m_g,k (c_g,k) 12: g(r)←Insert(g(r),cg,k,g,k,ψ)A^(r)_g (A^(r)_g,c_g,k,m_g,k,ψ) 13: end if 14: end for 15: g+1←ParetoRefresh(Migrate(g))A_g+1 (Migrate(A_g)) 16: end for 17: return Best(G)Best(A_G) 5. Experiments and Results 5.1. Experimental Setup Benchmarks and exposure. We evaluate CoEvoP&R on eight ChiPBench Nangate45 designs (Wang et al., 2025), including bp_fe, bp_be, swerv_wrapper, ethernet, dft68, or1200, vga_lcd, and mor1kx. To test cross-technology generalization, we further use ICCAD 2015 Superblue designs 1, 3, 4, 5, 7, 10, 16, and 18 (Kim et al., 2015), together with gcd, ibex, and ariane implemented using the ASAP7 predictive 7 nm standard-cell library (Clark et al., 2016). CoEvoP&R-E uses feedback from the same design on which it is evaluated. CoEvoP&R-L follows leave-one-design-out (LODO) evaluation, where the target design is excluded from objective evolution and prompt evidence. CoEvoP&R-T zero-shot transfers the best candidate from the ChiPBench evolution for cross-family test. Baselines and evaluation flow. Baselines include analytical placement through DREAMPlace and DREAMPlace 4.0 (Lin et al., 2019; Liao et al., 2022), expert HPWL, density, RUDY (Spindler and Johannes, 2007), and pin-density objective mixes, same-grammar BO-DSL implemented with tree-structured Parzen estimator (TPE) (Bergstra et al., 2011), and shared-pipeline reruns of AutoDMP (Agnesina et al., 2023), RoutePlacer (Hou et al., 2024), EvoPlace (Yao et al., 2026), and LaMPlace (Geng et al., 2025). Comparable rows share coordinate conversion, legalization, OpenROAD routing (Ajayi et al., 2019), OpenTimer timing analysis (Huang and Wong, 2015), and metric extraction. Evaluation and statistics. All results in Tables 1 and 2 report improvements over DREAMPlace, which is normalized to zero. These results are averaged under reruns using three different seeds. The default LLM proposer is gpt-5.4. All evolution and experiments run on one NVIDIA RTX A6000 GPU with 48 GB memory and one Intel Xeon Gold 5218R central processing unit (CPU) using 16 OpenROAD/OpenTimer worker processes. Section 5.5 tests LLM model-family sensitivity. 5.2. Main Results on ChiPBench Nangate45 Table 1 is the primary same-technology result, covering CoEvoP&R-E and CoEvoP&R-L. CoEvoP&R-E improves mean routed wirelength and congestion by 16.9% and 36.7%, and improves WNS and TNS by 0.70 ns and 912 ns over DREAMPlace. The strongest non-CoEvo means reach 9.7% rWL, 20.6% congestion, 0.41 ns WNS, and 444 ns TNS, with the leading method varying by metric. At the design level, LaMPlace remains strong on bp_fe and bp_be wirelength, DREAMPlace 4.0 leads or1200 WNS, and CoEvoP&R-E supplies most of the best or second-best congestion and timing entries. CoEvoP&R-L keeps 13.2% rWL, 31.4% congestion, 0.51 ns WNS, and 637 ns TNS under the LODO protocol. The best candidate in this experiment is zero-shot transferred to the cross-family evaluation in Section 5.3. Table 2. Post-route transfer on Superblue and ASAP7. Entries report three-seed mean improvements over DREAMPlace, which is normalized to zero. rWL and Cong. are reductions in routed wirelength and post-route congestion. WNS and TNS are post-route timing gains in ns. Bold marks best means and underline marks second best. Superblue ASAP7 Method Metric sb1 sb3 sb4 sb5 sb7 sb10 sb16 sb18 gcd ibex ariane DREAMPlace 4.0 (Liao et al., 2022) rWL (%) -0.4 -0.6 +0.2 -0.8 -0.3 +0.1 -0.5 +0.3 +0.4 -0.3 -0.8 Cong. (%) +0.2 -0.3 -0.4 +0.1 -0.2 -0.7 -0.1 +0.4 +0.6 +0.2 -0.4 WNS (Δ ) +4.4 +16.8 +8.7 +21.7 +5.2 +1.9 +4.7 +8.3 +0.06 +0.18 +0.34 TNS (Δ ) +16700 +3370 +5210 +11300 +9810 +7040 +31400 +4320 +32 +187 +940 RoutePlacer (Hou et al., 2024) rWL (%) +0.1 -0.3 +0.1 +0.3 -0.1 +0.9 -0.1 +0.2 +0.4 -0.6 +1.8 Cong. (%) +3.1 +1.7 +4.6 +3.9 +3.9 +43.2 +1.4 +3.7 +6.4 +24.6 +28.4 WNS (Δ ) -0.6 -0.4 +0.2 -0.4 -0.8 -1.2 -0.3 -0.7 -0.02 -0.14 -0.34 TNS (Δ ) -1400 -800 +200 -1200 -1600 -3400 -800 -1400 -12 -87 -294 EvoPlace (Yao et al., 2026) rWL (%) -2.4 -3.6 +0.4 -2.8 +1.4 -3.2 -1.8 +0.6 +1.2 -8.4 -13.4 Cong. (%) -3.4 -8.6 -2.4 -4.2 +0.4 -12.4 -6.8 -1.2 +0.4 -14.2 -18.4 WNS (Δ ) +0.2 +0.4 +0.4 -0.4 +0.6 -0.8 -0.4 +0.4 +0.02 -0.24 -0.48 TNS (Δ ) -600 -1400 +400 -1200 +300 -2200 -1400 +200 +4 -124 -487 LaMPlace (Geng et al., 2025) rWL (%) -4.6 -6.8 +14.2 -8.4 +0.4 -2.4 +11.8 +22.4 -6.8 -18.4 -22.4 Cong. (%) +25.2 +19.9 +4.9 +19.8 +23.0 +24.8 +7.7 +28.2 +2.4 -8.7 -14.2 WNS (Δ ) +2.4 +6.8 +3.4 +13.4 +3.6 +1.2 +2.8 +0.4 +0.04 -0.16 -0.32 TNS (Δ ) +2400 +6800 +2400 +8400 +3800 +1400 +1800 +1200 -18 -94 -246 CoEvoP&R-E rWL (%) +5.4 +6.2 +4.8 +7.2 +5.4 +3.6 +6.4 +4.2 +5.4 +8.2 +10.6 Cong. (%) +26.4 +21.4 +15.8 +22.4 +24.4 +34.7 +9.8 +30.4 +14.6 +22.4 +26.8 WNS (Δ ) +5.8 +14.8 +9.4 +12.8 +7.6 +2.6 +6.2 +6.4 +0.24 +0.42 +0.68 TNS (Δ ) +18800 +38400 +18700 +67200 +36400 +8400 +14600 +6800 +148 +468 +824 CoEvoP&R-T rWL (%) +1.4 +3.8 +2.6 +4.2 +3.2 +1.8 +3.6 +2.4 +2.4 +3.8 +5.2 Cong. (%) +20.4 +20.4 +10.4 +20.4 +18.4 +25.4 +8.4 +24.6 +8.4 +19.4 +25.4 WNS (Δ ) +4.6 +8.2 +8.8 +7.4 +5.4 +2.2 +5.4 +4.2 +0.08 +0.22 +0.34 TNS (Δ ) +2400 +18400 +9600 +32400 +14200 +7400 +7600 +4600 +58 +204 +620 5.3. Generalization Table 2 reports Superblue and ASAP7 transfer results for CoEvoP&R-E and CoEvoP&R-T. On Superblue, E averages 5.4% rWL, 23.2% congestion reduction, 8.2 ns WNS, and 26,163 ns TNS gain. T retains 2.9%, 18.6%, 5.8 ns, and 12,075 ns, or 46% to 80% of E. On ASAP7, LaMPlace regresses while both CoEvoP&R variants remain positive, supporting symbolic cross-node transfer. Superblue1 is the clearest transfer-limit case. Figure 3 visualizes Superblue16 under matched placement rendering. CoEvoP&R-L improves rWL/Cong. by 3.6%/8.4% and gains 5.4 ns WNS and 12,800 ns TNS over DREAMPlace, while DREAMPlace 4.0’s timing-driven flow leads on TNS (+31,400 ns) at the cost of slightly worse rWL and congestion. rWL=0.0, Cong.=0.0 WNS=0.0, TNS=0 rWL=−0.5-0.5, Cong.=−0.1-0.1 WNS=+4.7, TNS=+31,400 rWL=+3.6, Cong.=+8.4 WNS=+5.4, TNS=+12,800 Figure 3. Superblue16 qualitative placement comparison. Metrics below each panel are improvements over the DREAMPlace reference. Three-panel Superblue16 placement comparison showing DREAMPlace, DREAMPlace 4.0 timing-driven, and CoEvoP&R-L with matching placement and density overlays plus rWL, Cong., WNS, and TNS values. 5.4. Timing proxy Audit Figure 4 audits bp_fe, swerv, ethernet, and or1200 under 100 to 2000 database unit (DBU) perturbations. The left panel measures HPWL redundancy, and the right panel counts admission outcomes. Admit uses the proxy in candidate selection, Tie uses it only to break close cases, and Reject excludes it for that metric. WNS has lower HPWL redundancy and is admitted on three of four designs, making it the primary timing proxy. TNS tracks HPWL more strongly and acts mainly as a tie-breaker. The ungated timing-proxy row in Table 3 tests this admission rule. Figure 4. Timing-proxy admission audit. Grouped bars show median Pearson and Spearman correlation with HPWL movement. Two-panel bar chart for timing proxy audit. Delta WNS has Pearson 0.58 and Spearman 0.54, with three Admit outcomes and one Tie outcome. Delta TNS has Pearson 0.82 and Spearman 0.78, with one Admit, two Tie, and one Reject outcome. 5.5. Ablation Studies and LLM Sensitivity Table 3 isolates the main sources of performance in CoEvoP&R. Random DSL evolution and one-shot LLM sampling lose most of the routing and timing gains, showing that archive-conditioned iteration is essential. Tier-A-only selection and direct Tier C evaluation both underperform Full CoEvoP&R, indicating that cost-scaled evidence allocation is more useful than either cheap-only or route-all evaluation. Scalar-history and metrics-only archives also degrade, showing that component traces and failure memory provide actionable search context. The model variants preserve the main advantage, with stronger LLMs improving validity and score under the same prompts, mutation schedules, and budgets. Table 3. Ablation and LLM sensitivity on the ChiPBench stress panel bp_fe, swerv_wrapper, ethernet, and or1200. R/C and W/T denote routed wirelength/congestion reductions and WNS/TNS gains. Valid/Acc. denote valid and admitted candidates. Evals and GPU-h denote evaluations and GPU-hours. Darker blue indicates larger degradation from Full CoEvoP&R. Configuration Valid/Acc. Score R/C W/T Tier-C Cost (%) (Δ ) evals (GPU-h) Full CoEvoP&R 68/12 4.87 20.5/45.0 +0.64/+890 24 32 Proposal and objective search Random DSL evolution 42/4 1.24 2.6/4.8 +0.08/+60 24 26 One-shot LLM sampling 58/8 2.14 7.2/13.4 +0.22/+264 24 16 Score-only LLM evolution 62/10 3.42 14.2/26.4 +0.40/+454 24 29 Evaluation and evidence policy Prompt-only free-form interface 24/4 1.86 5.8/12.4 +0.15/+164 24 27 Tier-A-only selection 68/18 2.68 9.4/15.4 +0.22/+268 0 5 Direct Tier C (all valid) 68/8 3.14 12.0/23.6 +0.32/+498 68 64 Ungated timing proxy 68/14 3.86 16.4/34.6 +0.46/+588 24 33 Archive state Champion-only scalar history 62/10 2.42 9.6/17.6 +0.26/+312 24 31 Top-k scalar archive 64/11 3.14 13.6/26.4 +0.36/+448 24 32 MAP-Elites, metrics only 66/12 3.68 17.0/34.4 +0.46/+588 24 32 Model family sensitivity gpt-5.4 (default) 68/12 4.87 20.5/45.0 +0.64/+890 24 32 gpt-4.1-mini 62/10 3.87 16.4/36.0 +0.49/+668 24 26 claude-opus-4-8 66/12 4.62 19.4/43.2 +0.60/+836 24 35 qwen3.7-max 58/9 3.24 13.8/28.4 +0.42/+528 24 28 6. Conclusion CoEvoP&R evolves executable placement objectives with bounded analytical placement and routed feedback, turning routing and timing evidence into readable differentiable objectives for DREAMPlace. Across ChiPBench, Superblue, and ASAP7, the evolved objectives improve routed wirelength, congestion, and timing. On ChiPBench, target evolution reaches 16.9% rWL reduction, 36.7% congestion reduction, 0.70 ns WNS gain, and 912 ns TNS gain over DREAMPlace. The results support LLM-guided objective evolution as an inspectable interface between differentiable placement and downstream evaluation. A natural next step is to extend the archive and validation interface toward multi-corner timing, power, and signoff-accurate routing feedback. References A. Agnesina, R. Liang, G. Pradipta, A. Rajaram, and H. Ren (2025) GOALPlace: begin with the end in mind. In Proceedings of the International Symposium on Physical Design, New York, NY, USA, p. 2–10. External Links: Document Cited by: §1, §2. A. Agnesina, P. Rajvanshi, T. Yang, G. Pradipta, A. Jiao, B. Keller, B. Khailany, and H. Ren (2023) AutoDMP: automated DREAMPlace-based macro placement. In Proceedings of the International Symposium on Physical Design, New York, NY, USA, p. 149–157. External Links: Document Cited by: §2, Table 1, §5.1. T. Ajayi, V. A. Chhabria, M. Fogaça, S. Hashemi, A. Hosny, A. B. Kahng, M. Kim, J. Lee, U. Mallappa, M. Neseem, G. Pradipta, S. Reda, M. Saligane, S. S. Sapatnekar, C. Sechen, M. Shalan, W. Swartz, L. Wang, Z. Wang, M. Woo, and B. Xu (2019) Toward an open-source digital flow: first learnings from the OpenROAD project. In Proceedings of the 56th Annual Design Automation Conference, New York, NY, USA, p. 76:1–76:4. External Links: Document Cited by: §2, §5.1. J. S. Bergstra, R. Bardenet, Y. Bengio, and B. Kégl (2011) Algorithms for hyper-parameter optimization. In Advances in Neural Information Processing Systems, Vol. 24, Granada, Spain, p. 2546–2554. Cited by: §5.1. M. Chen and T. Wang (2025) HyperPlace: harnessing a large language model for efficient hyperparameter optimization in GPU-accelerated VLSI placement. ACM Transactions on Design Automation of Electronic Systems 30 (6), p. 1–27. External Links: Document Cited by: §2. C. Cheng, A. B. Kahng, I. Kang, and L. Wang (2019) RePlAce: advancing solution quality and routability validation in global placement. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems 38 (9), p. 1717–1730. External Links: Document Cited by: §2. L. T. Clark, V. Vashishtha, L. Shifren, A. Gujja, S. Sinha, B. Cline, C. Ramamurthy, and G. Yeric (2016) ASAP7: a 7-nm FinFET predictive process design kit. Microelectronics Journal 53, p. 105–115. External Links: Document Cited by: §5.1. Z. Geng, J. Wang, Z. Liu, S. Xu, Z. Tang, S. Kai, M. Yuan, J. Hao, and F. Wu (2025) LaMPlace: learning to optimize cross-stage metrics in macro placement. In International Conference on Learning Representations, Singapore. Cited by: §1, §2, Table 1, §5.1, Table 2. Y. Hou, H. Ye, Y. Zhang, S. Xu, and G. Song (2024) RoutePlacer: an end-to-end routability-aware placer with graph neural network. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, New York, NY, USA, p. 1085–1095. External Links: Document Cited by: §1, §2, §5.1, Table 2. T. Huang and M. D. F. Wong (2015) OpenTimer: a high-performance timing analysis tool. In Proceedings of the 2015 IEEE/ACM International Conference on Computer-Aided Design, Austin, TX, USA, p. 895–902. External Links: Document Cited by: §5.1. M. Kim, J. Hu, J. Li, and N. Viswanathan (2015) ICCAD-2015 CAD contest in incremental timing-driven placement and benchmark suite. In Proceedings of the IEEE/ACM International Conference on Computer-Aided Design, Austin, TX, USA, p. 921–926. External Links: Document Cited by: §5.1. P. Liao, S. Liu, Z. Chen, W. Lv, Y. Lin, and B. Yu (2022) DREAMPlace 4.0: timing-driven global placement with momentum-based net weighting. In Proceedings of the Design, Automation and Test in Europe Conference and Exhibition, Antwerp, Belgium, p. 939–944. External Links: Document Cited by: §2, Table 1, §5.1, Table 2. Y. Lin, S. Dhar, W. Li, H. Ren, B. Khailany, and D. Z. Pan (2019) DREAMPlace: deep learning toolkit-enabled GPU acceleration for modern VLSI placement. In Proceedings of the 56th Annual Design Automation Conference, New York, NY, USA, p. 1–6. External Links: Document Cited by: §1, §2, §3.1, §5.1. F. Liu, X. Tong, M. Yuan, X. Lin, F. Luo, Z. Wang, Z. Lu, and Q. Zhang (2024) Evolution of heuristics: towards efficient automatic algorithm design using large language model. In Proceedings of the 41st International Conference on Machine Learning, Proceedings of Machine Learning Research, Vol. 235, Vienna, Austria, p. 32201–32223. Cited by: §2. J. Lu, P. Chen, C. Chang, L. Sha, D. J.-H. Huang, C. Teng, and C. Cheng (2014) ePlace: electrostatics based placement using nesterov’s method. In Proceedings of the 51st Annual Design Automation Conference, New York, NY, USA, p. 1–6. External Links: Document Cited by: §2, §3.1. Y. J. Ma, W. Liang, G. Wang, D. Huang, O. Bastani, D. Jayaraman, Y. Zhu, J. Fan, and A. Anandkumar (2024) Eureka: human-level reward design via coding large language models. In International Conference on Learning Representations, Vienna, Austria. Cited by: §1, §2. A. Mirhoseini, A. Goldie, M. Yazgan, J. W. Jiang, E. Songhori, S. Wang, Y. Lee, E. Johnson, O. Pathak, A. Nova, J. Pak, A. Tong, K. Srinivasa, W. Hang, E. Tuncer, Q. V. Le, J. Laudon, R. Ho, R. Carpenter, and J. Dean (2021) A graph placement methodology for fast chip design. Nature 594, p. 207–212. External Links: Document Cited by: §2. J. Mouret and J. Clune (2015) Illuminating search spaces by mapping elites. CoRR abs/1504.04909. External Links: Document Cited by: §4.4. A. Novikov, N. Vũ, M. Eisenberger, E. Dupont, P. Huang, A. Z. Wagner, S. Shirobokov, B. Kozlovskii, F. J. R. Ruiz, A. Mehrabian, M. P. Kumar, A. See, S. Chaudhuri, G. Holland, A. Davies, S. Nowozin, P. Kohli, and M. Balog (2025) AlphaEvolve: a coding agent for scientific and algorithmic discovery. Note: arXiv preprint arXiv:2506.13131 Cited by: §1, §2. B. Romera-Paredes, M. Barekatain, A. Novikov, M. Balog, M. P. Kumar, E. Dupont, F. J. R. Ruiz, J. S. Ellenberg, P. Wang, O. Fawzi, P. Kohli, and A. Fawzi (2024) Mathematical discoveries from program search with large language models. Nature 625, p. 468–475. External Links: Document Cited by: §1, §2. A. Sharma (2025) OpenEvolve: an open-source evolutionary coding agent. Note: https://github.com/algorithmicsuperintelligence/openevolve Cited by: §2. P. Shojaee, K. Meidani, S. Gupta, A. B. Farimani, and C. K. Reddy (2025) LLM-SR: scientific equation discovery via programming with large language models. In International Conference on Learning Representations, Singapore. Cited by: §1, §2. P. Spindler and F. M. Johannes (2007) Fast and accurate routing demand estimation for efficient routability-driven placement. In Proceedings of the Conference on Design, Automation and Test in Europe, Nice, France, p. 1226–1231. External Links: Document Cited by: §1, §2, §3.2, §5.1. I. Uchendu, S. Goel, K. Hou, E. Songhori, K. Lee, J. W. Jiang, V. J. Reddi, and V. Zhuang (2026) See it to place it: evolving macro placements with vision-language models. Note: arXiv preprint arXiv:2603.28733 Cited by: §1, §2. Z. Wang, Z. Geng, Z. Tu, J. Wang, Y. Qian, Z. Xu, Z. Liu, S. Xu, Z. Tang, S. Kai, M. Yuan, J. Hao, B. Li, and F. Wu (2025) Benchmarking end-to-end performance of AI-based chip placement algorithms. In Advances in Neural Information Processing Systems, Vol. 38. Note: Datasets and Benchmarks Track Cited by: §1, §3.1, §5.1. X. Yao, J. Jiang, Y. Zhao, P. Liao, Y. Lin, and B. 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. Note: Early access; arXiv:2504.17801 External Links: Document Cited by: §1, §2, Table 1, §5.1, Table 2.