acceptodds
Under review as a conference paper at ICLR 2027

Trinity: One Differentiable Physics for Training, Refining and Scoring Generative Floorplanners

Abstract

Floorplanning arranges the blocks of a chip and decides their shapes under objectives that press blocks together, short wirelength and a small outline, and constraints that hold them apart, non-overlap, clusters, MIB shapes and boundary blocks. Recent diffusion placers train on reference layouts alone and leave this coupled system to guidance, post-hoc loops and a legalizer, reporting only the endpoint, which hides what the generator contributes. We re-implement four of them under one recipe on FloorSet, score raw, refined and legalized layouts on one scale, and propose Trinity, a flow-matching floorplanner whose six differentiable functions for the constraints and objectives are its training loss term, the energy of a closed-form refiner after sampling and the base of a soft cost for every stage. The network thus learns the correction prior placers apply in their samplers, and sampling needs no guidance. Stage by stage, the training term lowers a plain transformer's raw soft cost by 26% and matters most at short budgets, the shared refiner decides more of the final cost than the generator and matches a ported placer's loop in 16 to 660 times fewer steps, Trinity's refined soft cost is 36% below the best ported pipeline, the soft cost ranks settings as the contest's hard cost does, and on the FloorSet val set the pipeline reaches a mean hard cost of 1.014 in 1.63 s per case.

Primary area:
generative models
Keywords:
Floorplanning, Diffusion Models, Physics-Informed Generative Models, Constrained Generation, Electronic Design Automation

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_V4ZPOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Trinity is a flow-matching floorplanner for FloorSet with soft blocks and the full ICCAD 2026 constraint set. Six differentiable functions (overlap, cluster gap, MIB shape spread, boundary distance, wirelength, outline area) are used three times. They are an auxiliary loss on the clean prediction during training, the energy for an Adam refiner after sampling, and the basis of a "soft cost" that scores layouts before legalization. A sequence-pair LP legalizer produces the final feasible layout. The authors re-implement four diffusion placers under their own recipe and compare raw, refined and legalized layouts. On the FloorSet val set, the pipeline reaches a mean hard cost of 1.014 at 1.63 s per case with best-of-48 draws, against 1.222 for PARSAC with 16 workers. Short propositions motivate the clean-prediction head and the energy term.

Critical strengths

Strength 1: The stage-by-stage evaluation is a useful methodological point. Table 6 shows that once all generators share one refiner and legalizer, the first-draw hard cost spread shrinks from roughly 1.51 to 1.73 down to 1.23 to 1.36. This makes it clear that endpoint-only comparisons in prior diffusion placers mix generator quality with the strength of the correction loop. I think this observation is valuable for the community regardless of the method.

Strength 2: The full pipeline is practically strong. It handles soft blocks, pre-placed blocks, clusters, MIB and boundary constraints, and it reaches near-reference hard cost on the official FloorSet val set in about 1 to 2 s per case, which is far below PARSAC's CPU time (Table 7, Appendix L). The experimental effort is large (66 models, seeds in Appendix M), and code is provided as supplementary material.

Critical weaknesses

Weakness 1: The idea that putting the physics energy into the training loss has little effect on the final result. In Table 5, Trinity and Trinity at lambda = 0 both reach 1.07 soft cost after 400 refiner steps at NFE 1, and after only 25 steps they differ by 1.31 vs 1.35. On GSRC (Table 8), lambda = 0 is better on n100 and n300. Since 400 refiner steps are cheap (Table 9), the "short budget" advantage matters little in practice. Most of the gain over the ported pipelines comes from the refiner and legalizer, not from training. Also, the soft cost is built from the same six functions Trinity is trained on, so soft-cost comparisons favour Trinity by construction.

Weakness 2: The baseline comparison and the novelty claims are weaker than presented. All four diffusion baselines are ports into the authors' latent space, conditioning, sampler and overrides, and they were originally designed for fixed-shape macro placement. There is no check that the ports reproduce the original methods' behaviour. On the one external benchmark with published numbers (GSRC, Table 8), Trinity is 2 to 6% above PARSAC's published best on n100 and n200 and does not fit the n300 outline. On the method side, constraint losses on the clean prediction are already used by LACE and physics-informed diffusion, and the refiner is gradient descent on an analytical placement energy. The contribution is therefore mainly an integration and evaluation study, which the abstract and contributions list overstate.

Questions for authors

Clarification requests:

  1. What is the legalized hard cost (first draw and best of N) for Trinity at lambda = 0 through the same refiner and legalizer on the FloorSet val set? This is the direct test of whether the training term matters at the endpoint.
  2. Table 7 reports 1.014 with best of 48 draws selected by hard cost. What is the first-draw hard cost on the val set, and how do you normalize the comparison with PARSAC (GPU vs CPU-hours, 48 draws vs 16 workers)?
  3. Did any ported placer reproduce its published numbers on its original task before porting? Without this, how much of the gap in Tables 3 and 4 comes from the porting?
  4. Trinity does not fit the outline on GSRC n300, which is 2.5 times the largest training size. Is the limit the training range or the model? Would training on larger FloorSet instances fix it?

Suggestions for future work:

The main text is very dense and uses many local terms ("information mover", "rungs", "arms", "physics") that make it hard to read. Moving some tables to the appendix and stating the main claims in plain terms would help. A comparison on independent metrics, for example hard cost after a neutral legalizer, would make the generator comparison more convincing.

4 comments

We are grateful for the reviewer's time and for the comments on our work, and would like to respond to each comment as follows. The table and figure numbers in our responses follow the revised paper, and we kindly refer the reviewer to it.

Comment 1. The effect of the training term on the final result.

Response. We thank the reviewer for the comment, and respectfully clarify that the paper assigns a distinct role to each use of the energy: the training term improves the raw samples of the generator and matters most at short budgets, whereas the refiner decides more of the final cost, as the abstract states. To directly address the question of the reviewer, Table 39 reports Trinity with and without the training term through the same refiner and legalizer on the FloorSet validation set over three seeds. At one sampling step and after 400 refinement steps, the legalized cost of the first draw is 1.170 with the term and 1.167 without it, which differ by less than the spread across seeds, and the lowest cost among 48 draws is 1.035 and 1.044. The same holds at 32 sampling steps. The contribution of the training term is to the generator itself. On the same validation set at one sampling step, the term lowers the soft cost of the raw samples from 2.38 to 1.71, and it reduces the expansion that the legalizer applies to these samples from 3.09 to 2.74. On our held-out validation set of 12,000 cases, the raw samples reach 1.51 at 32 sampling steps with the term and without any guidance, whereas the guided samples of ChipDiffusion, MacroDiff+, and DiffPlace remain at 2.04 or above (Table 4). Moreover, the endpoint is set by the energy of the refiner: when this energy is exchanged, the legalized cost of every model moves together, from about 1.23 to about 1.34, irrespective of how the model was trained (Table 40). A sufficiently strong correction therefore conceals the difference between generators at the endpoint, which is the reason that the paper evaluates each stage separately. The revised Sections 1 and 4.3 state these roles explicitly. We hope that this response clarifies the role of the training term.

Comment 2. The results on GSRC.

Response. We thank the reviewer for the question, and respectfully clarify three points on GSRC: the two designs on which the variant trained without the energy term is shorter differ by less than the variation across training seeds, the design with 300 blocks (n300) fits the outline once the solver of the legalizer has sufficient time, and Trinity is shorter than our run of the released PARSAC. First, with 16 draws per seed, the variant trained without the term is shorter by 0.3% and 1.4% on the designs with 100 and 300 blocks (n100 and n300), Trinity is shorter on the designs with 10, 30, and 50 blocks (n10, n30, and n50), and two training seeds of Trinity differ from each other by up to 2.4% (Table 21). On the design with 200 blocks (n200), every draw of Trinity fits the outline, whereas at most 1% of the draws of the variant trained without the term fit it. Second, the design with 300 blocks did not fit the outline in the original submission since the solver of the legalizer was stopped after 2 seconds. With 60 seconds, both Trinity and this variant fit the outline, which indicates that the limit is the time of the solver rather than the model or the training range (Tables 8 and 21). Third, the wirelength of Trinity is 3% to 4% shorter than that of our run of the released PARSAC with 16 workers, and it remains 2.3% to 8.5% above the published best of PARSAC, which is obtained with 112 workers (Table 67). We hope that this response clarifies the results on GSRC.

(to be continued...)

(continued...)

Comment 3. The first-draw cost and the comparison of compute with PARSAC.

Response. We thank the reviewer for the question. On the FloorSet validation set, a single draw of Trinity, with no selection among draws, reaches a mean hard cost of 1.235 at 8 sampling steps, which is the setting of the reported 1.014. For comparison, PARSAC with 16 workers ends at 1.222 at its published budget (Appendix L.1). Regarding the normalization, we respectfully note that the time on a GPU and the time on CPU workers are different resources, for which no meaningful conversion exists. The paper therefore reports the time of each method as measured (Table 62). We hope that this response answers the question of the reviewer.

Comment 4. The soft cost and the comparisons that use it.

Response. We thank the reviewer for the question, and respectfully clarify that the soft cost is the hard cost extended to layouts before legalization, and that it measures the objectives and the constraints of the task rather than an objective specific to Trinity. Specifically, the hard cost applies only to legal layouts, and it is unable to score a layout in which blocks still overlap. The soft cost uses the same formula, and it replaces the count of violated constraints with their severity and penalizes the overlap, such that a layout before legalization receives a score on the same scale (Section 3.1). The two costs therefore describe the same quality from two viewpoints, and the soft cost is the more informative one, since it reveals differences that legalization conceals: at 32 sampling steps, the raw samples of the five generators range from 1.51 to 6.23 in the soft cost, whereas their hard cost after legalization ranges only from 1.51 to 1.73 (Tables 4 and 6). Two experimental results of the paper validate this measure. First, the soft cost and the hard cost place the settings in nearly the same order: their rank correlation is 0.94 for Trinity and 0.92 to 0.99 for the four prior placers, where 1 indicates an identical order (Table 10). Second, the comparison by the hard cost reaches the same conclusion. After legalization, the hard cost of Trinity is 1.234, whereas the lowest hard cost among the four prior placers is 1.525 (Table 3). Moreover, the refined layouts of Trinity remain ahead of those of the four prior placers when our legalizer is replaced with an independent legalizer built from PARSAC (Appendix K.7). We hope that this response clarifies the meaning of the soft cost.

Comment 5. The fidelity of the ported placers to the original methods.

Response. We thank the reviewer for the question. To directly address this question, we present an additional parity check of each ported placer against its released code. Since the original benchmarks of these placers fix the shape of every block, the paper verifies each port against the released code and weights (Appendix B.4). Specifically, each ported placer retains its released network and objective, together with its post-hoc loop where one is released, and adopts the latent representation, the conditioning, the sampler, and the overrides that Trinity uses (Section 4.1), and Appendix B.4 lists every adaptation. With the released weights, where these are available, each ported network reproduces the forward pass of its released code to within 1.4×10−61.4 \times 10^{-6}, and the ported guidance and loop energies match the released ones to within a relative difference of 4×10−124 \times 10^{-12}. For two settings in which a port departs from its release, namely the legality weight of the ChipDiffusion loop and the spacing of the global node in DiffPlace, the released settings change the legalized cost by at most 0.012. Moreover, we kindly highlight that Table 6 keeps each ported generator fixed and changes only the correction applied to its samples. Through its own pipeline, the legalized cost of the four ported placers is 1.525 to 1.567, and with our refiner on the same generators it decreases to 1.296 to 1.355 (Tables 3 and 6). Since the generator and its adaptation are unchanged in this comparison, the improvement arises from the correction alone. Furthermore, the training objective of FlowPlace on our larger backbone yields a raw soft cost of 2.60, compared with 2.46 on its own network (Appendix P), which places the gap in the training objective rather than in the network size. These results validate that the ports are faithful to the released methods, and that the gap to Trinity arises from the training objective and the correction of each placer rather than from its adaptation to FloorSet. We hope that these results answer the question of the reviewer. The revised paper reports the parity check in Appendix B.4.

(to be continued...)

(continued...)

Comment 6. The contribution of this paper.

Response. We thank the reviewer for the comment, and respectfully clarify that the main contribution of this paper is to place the coupled objectives and constraints of floorplanning in the training loss of a generative floorplanner, such that the network learns the constraints together with the objectives rather than leaving them to later corrections, and to establish the theory behind this choice. Specifically, a network trained on reference layouts alone estimates a conditional mean, and a mean of feasible layouts is not necessarily feasible, since the feasible set is nonconvex (Proposition D.2). Proposition 3.1 establishes that the minimizer of the training loss with the energy is a proximal step on this energy. The network therefore learns the correction that prior placers apply inside their samplers, and sampling runs without guidance. In contrast, Proposition 3.2 establishes that a correction with a fixed step inside the sampler moves even a layout that is already feasible. Moreover, the refiner resumes the descent of the same energy after sampling (Section 3.3). The design of Trinity therefore follows from this theory rather than from an integration of existing components.

Regarding LACE and physics-informed diffusion, we would like to kindly clarify that these two works address problems that differ in kind from floorplanning. LACE generates graphic layouts, such as documents and user interfaces, of at most 25 elements, and its two constraints, alignment and overlap, are aesthetic measures. In particular, LACE omits the overlap constraint on the dataset in which overlap is prevalent. Physics-informed diffusion generates fields on a regular grid, and its constraint is the residual of a partial differential equation. Floorplanning is a different and more complex problem: its input is a netlist that connects the blocks, its objectives are the wirelength over this netlist and the outline area, these objectives compete with several kinds of constraints, and a layout is feasible only when no two blocks overlap. Moreover, neither work analyzes how a constraint term in training relates to guidance during sampling, which is the subject of Propositions 3.1 and 3.2.

Furthermore, at 32 sampling steps, the same six functions do not improve any of the three prior placers that were retrained with them. To directly address this comment, we present this additional experiment in Table 34, in which FlowPlace, ChipDiffusion, and MacroDiff+ were retrained with the six functions applied to the clean estimate of each network. For FlowPlace, which predicts the velocity, these functions raise the soft cost of the raw samples from 2.46 to 2.57. For ChipDiffusion and MacroDiff+, which predict the noise, these functions raise it from 2.80 to 6.95 and from 3.85 to 6.35. After 400 refinement steps, the legalized cost of the first draw is higher for all three placers as well (Appendix H.9). In contrast, on the clean prediction of a plain transformer, the same functions lower the soft cost of the raw samples from 2.09 to 1.56 (Table 2). These results indicate that the effect of a constraint loss depends on the output on which it acts.

Regarding the refiner, the paper replaces the correction loops of prior placers with the descent of the training energy itself. On the samples of DiffPlace, MacroDiff+, and ChipDiffusion, our refiner reaches the soft cost of the released loop of each placer within 9, 31, and 8 steps, whereas these loops run 500, 500, and 5,000 iterations (Table 9). With this refiner, every ported generator reaches a lower legalized cost than through its own pipeline (1.296 to 1.355 against 1.525 to 1.567, Tables 3 and 6). We kindly refer the reviewer to Tables 1 and 2 for the ablations of the generator, and to Tables 5 and 6 for those of the refiner. We hope that this response clarifies the contribution of this paper.

(to be continued...)

(continued...)

Comment 7. The presentation of the paper and a comparison under a neutral legalizer.

Response. We sincerely thank the reviewer for the suggestions. With the additional page that ICLR allows in the rebuttal phase, we have adjusted the layout of the main text to provide more space, and the changes are highlighted in color. In particular, the main text now refers to the graph convolution and to the model sizes by these names. The camera-ready version will further improve the flow of the main text along the suggestions of the reviewer. Regarding a comparison on an independent metric, the revised paper includes the comparison that the reviewer suggests. On the FloorSet validation set at 32 sampling steps, when the refined layouts of every generator are legalized by a procedure built from the PARSAC annealer in place of our legalizer, Trinity remains ahead of the four prior placers: the lowest cost among four draws is 1.871 for Trinity, against 1.901 to 2.000 for these placers (Table 59 and Appendix K.7). We hope that these revisions address the suggestions of the reviewer.

Reviewer_hTHYOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

Trinity combines constraint supervision, generative modeling, gradient based refinement, and legalization for variable shape floorplanning. Its central proposal is to use shared geometric functions across training, refinement, and evaluation.

Critical strengths

The complete pipeline achieves competitive layout quality and short reported inference time on FloorSet relative to the evaluated baselines, demonstrating its potential for the studied task.

Critical weaknesses
  1. The motivation does not establish a clear limitation of existing methods. The absence of constraint losses during training does not explain why sampling guidance or postprocessing is inadequate. The theory and ablations support adding constraint supervision, but do not isolate the benefit of sharing the same energy across stages beyond the effect of that supervision. The demonstrated contribution therefore appears incremental, rather than resolving the claimed fundamental limitation of generative floorplanning.
  2. The comparisons with prior generative placers do not establish whether their poorer results arise from their constraint handling strategy or their adaptation to FloorSet. The main learned baselines are adapted from placement with fixed macro shapes to FloorSet, with modified representations and sampling procedures (Section 4.1). The comparison therefore mixes methodological improvements with differences in task adaptation and constraint coverage. Moreover, Table 9 measures speedups against each baseline’s own soft cost endpoint, rather than comparable final legalized quality. The reported speedups therefore do not establish an equivalent advantage in producing legal layouts over these generative baselines.
Questions for authors
  1. Does sharing the energy across stages still improve results when constraint supervision, shape freedom, and refinement capability are matched?
  2. How does Trinity compare with representative RL methods [1–4] and analytical placers such as DREAMPlace [5] and RePlAce [6] on compatible tasks using complete placement flows?
  3. On public designs from OpenROAD (https://github.com/The-OpenROAD-Project/OpenROAD) or MacroPlacement (https://github.com/TILOS-AI-Institute/MacroPlacement), how do congestion and timing compare with the open source OpenROAD flow, commercial tools, and manual macro floorplans?
  4. At matched legalized quality, what runtime advantage remains after preprocessing and initialization, and how many designs are needed to amortize data preparation and training?

References

[1] R. Cheng and J. Yan. On Joint Learning for Solving Placement and Routing in Chip Design. NeurIPS, 2021.

[2] Y. Lai, Y. Mu, and P. Luo. MaskPlace: Fast Chip Placement via Reinforced Visual Representation Learning. NeurIPS, 2022.

[3] Y. Lai et al. ChiPFormer: Transferable Chip Placement via Offline Decision Transformer. ICML, 2023.

[4] A. Mirhoseini et al. A graph placement methodology for fast chip design. Nature, 594:207–212, 2021.

[5] Y. Lin et al. DREAMPlace: Deep Learning Toolkit-Enabled GPU Acceleration for Modern VLSI Placement. DAC, 2019.

[6] C. K. Cheng, A. B. Kahng, I. Kang, and L. Wang. RePlAce: Advancing Solution Quality and Routability Validation in Global Placement. IEEE TCAD, 38(9):1717–1730, 2019.

3 comments

We appreciate the reviewer's time and effort spent on the review, and would like to respond as follows. The table and figure numbers in our responses follow the revised paper, and we kindly refer the reviewer to it.

Comment 1. The limitation of existing methods that motivates this paper.

Response. We thank the reviewer for the comment, and respectfully clarify that the limitation that motivates this paper lies in the corrections of existing generative placers, namely sampling guidance and post-hoc loops. Specifically, a network trained on reference layouts alone estimates a conditional mean, and a mean of feasible layouts is not necessarily feasible, since the feasible set is nonconvex (Section 1). Existing placers therefore correct their samples outside the network. However, Proposition 3.2 establishes that a correction with a fixed step moves even a layout that is already feasible, whereas the correction that Trinity learns in training vanishes as the sample approaches the data. The measurements on the 12,000 reference layouts confirm this result: these layouts have a mean soft cost of 1.065, and the released post-hoc loops of the prior placers move the blocks of these layouts by 0.025 to 0.029 of the layout scale and raise the soft cost of every layout, to a mean between 1.54 and 2.58 (Section 4.3 and Fig. 5). In contrast, our refiner moves the blocks by 0.0055 of the layout scale and lowers the soft cost in 85% of the cases. Moreover, at 32 sampling steps, the guidance of each guided placer increases the wirelength and the outline area of its samples, and the released guidance of DiffPlace even raises the overlap ratio from 0.113 to 0.328 (Appendix J). Even with guidance, the raw soft cost of every placer remains at 2.04 or above, whereas Trinity reaches 1.51 without guidance (Table 4). The revised introduction states this limitation explicitly. We hope that this response clarifies the motivation of this paper.

Comment 2. The benefit of one energy shared across training, refinement, and scoring.

Response. We thank the reviewer for the comment, and respectfully clarify that the shared energy allows the generator to learn the coupled system of constraints and objectives as a whole during training, the refiner to continue the same descent after sampling, and every stage to be scored on one scale. Specifically, with the energy in its training loss, the network learns the trade-off among the six functions before any correction is applied. In particular, the soft cost of the raw samples of Trinity is 1.51, against 2.02 for the same network trained without the energy (Table 1). Three of its constraint metrics, namely the boundary distance, the cluster gap, and the MIB variance, decrease at once by 73%, 18%, and 74%, while the overlap ratio stays level and the wirelength and the area are not sacrificed (Appendix H.6). Moreover, no guidance is required during sampling, since the network learns during training the correction that prior placers apply inside their samplers. After sampling, the refiner continues this descent. Specifically, Proposition 3.1 establishes that the correction learned in training is a proximal step on the energy, with a step size that vanishes at the end of sampling, and the refiner resumes the descent of the same energy from that point (Section 3.3). With this refiner, every ported generator reaches a lower legalized cost than through its own pipeline (1.296 to 1.355 against 1.525 to 1.567, Tables 3 and 6).

Furthermore, since training, refinement, and scoring share one set of functions, raw, refined, and legalized layouts are scored on one scale, and the soft cost ranks the settings as the hard cost does, with a Spearman correlation of 0.94 (Fig. 7). To directly address this question, we present an additional experiment, in which a second model was trained with an alternative energy that covers the same six terms in another functional form, and each training energy was crossed with each refiner energy under matched supervision, shape freedom, and refinement budget (Table 40). The training energy determines the raw sample: the raw soft cost is 1.51 for the model trained with our energy, 1.59 for the model trained with the alternative energy, and 2.02 for the model trained without an energy. The refiner energy determines the endpoint: with our energy in the refiner, the legalized cost of the first draw of these three models is 1.234, 1.242, and 1.226, against 1.337, 1.341, and 1.331 with the alternative energy in the refiner, and under either refiner energy these three models differ by at most 0.016. The revised paper includes this experiment in Appendix I.4. We hope that this response clarifies the role of the shared energy.

(to be continued...)

(continued...)

Comment 5. The comparison with RL methods and analytical placers, and the congestion and timing on public designs.

Response. We thank the reviewer for the comment, and respectfully clarify that this paper studies floorplanning, in which the position and the shape of every block are determined together under design constraints, whereas the RL methods and the analytical placers that the reviewer lists place macros or cells of fixed shapes. Specifically, FloorSet requires a shape for every soft block and imposes cluster, boundary, and multi-instance constraints, which the placement benchmarks do not contain (Section 2 and Table 11). On this task, the paper compares Trinity with the methods that are designed for it. On the FloorSet validation set, Trinity reaches a mean hard cost of 1.019 in 1.14 seconds per case, whereas PARSAC, the annealer of the group that built FloorSet, reaches 1.222 in about five CPU hours per case (Appendix L.1). On the GSRC benchmark with soft blocks, on which Trinity was not trained, its wirelength is below the published wirelength of GoodFloorplan, a learned floorplanning policy, at all three sizes (Table 8). Regarding congestion and timing, both quantities are measured after the complete downstream flow, whereas this paper works at the block level of the floorplan and evaluates the wirelength, the outline area, and the constraints of each layout (Section 3.1). We hope that this response clarifies the task that this paper addresses.

Comment 6. The runtime at matched legalized quality and the amortization of training.

Response. We thank the reviewer for the question. To directly address this question, we present an additional comparison at matched legalized quality in Table 65, which reports the time that each method requires to reach each level of the mean hard cost on the FloorSet validation set, and we kindly refer the reviewer to it. At each level, the pipeline of Trinity requires under one second per case, which covers sampling, refinement, and legalization. Regarding amortization, Trinity is trained once in about ten GPU hours. We respectfully note that the time on a GPU and the time on CPU workers are different resources, for which no meaningful conversion exists, and the paper therefore reports the time of each method as measured. We hope that these results answer the question of the reviewer. The revised paper includes Table 65 in Appendix L.1.

Thanks to the authors for their detailed responses and revisions. Most of my concerns have been addressed, so I'll raise my overall score to 3.

Reviewer_sfcrOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Trinity, a flow-matching floorplanner for the FloorSet benchmark (ICCAD 2026 Contest Problem C). It handles soft-block shapes and the full constraint set, including fixed and pre-placed blocks, clusters, multi-instance blocks and boundary blocks. Six differentiable functions describe the constraints and objectives: overlap, cluster gap, MIB shape deviation, boundary distance, wirelength and area. The same functions serve as an extra training loss, as the energy for a closed-form Adam refiner after sampling, and as the basis of a “soft cost” that scores raw, refined and legalized layouts on one scale. The generator is a DiT with a netlist attention bias, a one-hop graph-convolution residual and Laplacian positional encoding. It predicts the clean layout in a log-aspect-ratio latent that preserves block area by construction, and a sequence-pair LP legalizer produces the final feasible layouts. The authors retrain four diffusion placers (ChipDiffusion, DiffPlace, MacroDiff+, FlowPlace) under the same recipe for comparison. The physics term lowers raw soft cost by 26%, and the shared refiner matches each prior placer’s own correction loop in 16 to 660 times fewer steps. On the FloorSet validation set, the full pipeline reaches a mean hard cost of 1.014 in 1.63 seconds per case with 48 draws, while the PARSAC annealer reaches 1.222 after hours of CPU time.

Critical strengths

The comparison is careful and fair in setup. All four prior placers are retrained under the same recipe, latent and sampler, and every stage is scored on one scale, which makes it possible to separate the generator’s contribution from that of the refiner and legalizer. Prior placement papers report only the final result, and the paper shows convincingly that this hides how good the generator actually is. The experimental work is extensive: about 68 trained models, three seeds for the main ones, a model-size study, and an open account of limitations.

The practical results are strong. The full pipeline beats the retrained diffusion baselines on the official contest cost and gets close to the reference layouts in one to two seconds per case. PARSAC needs hours per case and ends worse. The refiner is simple, fast and helps every generator it is applied to.

Critical weaknesses
  1. The paper's central claim, that the physics should go into training, is not supported by its main results. The effect of the training term is reported only on raw samples and at short refinement budgets. Table 5 suggests the advantage disappears with refinement: at NFE 1, Trinity and Trinity at λ=0 both reach a soft cost of 1.07 after 400 refiner steps. Section 4.3 also states that the refiner decides most of the final cost, and Table 6 shows that the gap between generators shrinks considerably once they share the refiner. The main text never compares Trinity with and without the training term on the final legalized hard cost, which is the number that matters for the task. As presented, the evidence supports the refiner and the stage-by-stage evaluation as the main contributions, and the title and contributions should be adjusted unless the end-to-end effect of the training term can be shown.

  2. Several of the main comparisons favor the proposed method. The soft cost (Section 3.1) is built from the same functions Trinity is trained on and its refiner minimizes, so the headline result that Trinity's refined soft cost is 36% below the best ported pipeline partly measures how well each method optimizes Trinity's own objective. The prior methods are macro placers designed for fixed block shapes, ported into Trinity's latent, recipe and sampler, so they are evaluated on a task they were not built for. The headline FloorSet result of 1.014 selects the best of up to 48 draws by the hard cost, which is computed against the reference layout's wirelength and area, and a real design has no reference layout to select against. Since the refiner drives most of the final quality, the paper also needs a baseline that runs the same refiner and legalizer from a non-learned starting point, such as a spectral embedding of the netlist, to show what the generative model itself adds.

Questions for authors
  1. What is the final legalized hard cost of Trinity with and without the training term? If they are close, what does the training term contribute beyond short refinement budgets?
  2. What hard cost does the refiner plus legalizer reach from a non-learned starting point, such as a spectral embedding of the netlist or an analytical placement using the same energy?
  3. The 1.014 result selects among draws using a cost computed against the reference layout. What is the result when the draw is selected without the reference, for example by soft cost?
  4. How does performance change on designs larger than the training range, or on real designs?
4 comments

We thank the reviewer for the time dedicated to the review and for the comments, and would like to address each comment as follows. The table and figure numbers in our responses follow the revised paper, and we kindly refer the reviewer to it.

Comment 1. The effect of the training term on the final result.

Response. We thank the reviewer for the comment, and respectfully clarify that the paper assigns a distinct role to each use of the energy: the training term improves the raw samples of the generator and matters most at short budgets, whereas the refiner decides more of the final cost, as the abstract states. To directly address the question of the reviewer, Table 39 reports Trinity with and without the training term through the same refiner and legalizer on the FloorSet validation set over three seeds. At one sampling step and after 400 refinement steps, the legalized cost of the first draw is 1.170 with the term and 1.167 without it, which differ by less than the spread across seeds, and the lowest cost among 48 draws is 1.035 and 1.044. The same holds at 32 sampling steps. The contribution of the training term is to the generator itself. On the same validation set at one sampling step, the term lowers the soft cost of the raw samples from 2.38 to 1.71, and it reduces the expansion that the legalizer applies to these samples from 3.09 to 2.74. On our held-out validation set of 12,000 cases, the raw samples reach 1.51 at 32 sampling steps with the term and without any guidance, whereas the guided samples of ChipDiffusion, MacroDiff+, and DiffPlace remain at 2.04 or above (Table 4). Moreover, the endpoint is set by the energy of the refiner: when this energy is exchanged, the legalized cost of every model moves together, from about 1.23 to about 1.34, irrespective of how the model was trained (Table 40). A sufficiently strong correction therefore conceals the difference between generators at the endpoint, which is the reason that the paper evaluates each stage separately. The revised Sections 1 and 4.3 state these roles explicitly. We hope that this response clarifies the role of the training term.

Comment 2. The soft cost and the comparisons that use it.

Response. We thank the reviewer for the question, and respectfully clarify that the soft cost is the hard cost extended to layouts before legalization, and that it measures the objectives and the constraints of the task rather than an objective specific to Trinity. Specifically, the hard cost applies only to legal layouts, and it is unable to score a layout in which blocks still overlap. The soft cost uses the same formula, and it replaces the count of violated constraints with their severity and penalizes the overlap, such that a layout before legalization receives a score on the same scale (Section 3.1). The two costs therefore describe the same quality from two viewpoints, and the soft cost is the more informative one, since it reveals differences that legalization conceals: at 32 sampling steps, the raw samples of the five generators range from 1.51 to 6.23 in the soft cost, whereas their hard cost after legalization ranges only from 1.51 to 1.73 (Tables 4 and 6). Two experimental results of the paper validate this measure. First, the soft cost and the hard cost place the settings in nearly the same order: their rank correlation is 0.94 for Trinity and 0.92 to 0.99 for the four prior placers, where 1 indicates an identical order (Table 10). Second, the comparison by the hard cost reaches the same conclusion. After legalization, the hard cost of Trinity is 1.234, whereas the lowest hard cost among the four prior placers is 1.525 (Table 3). Moreover, the refined layouts of Trinity remain ahead of those of the four prior placers when our legalizer is replaced with an independent legalizer built from PARSAC (Appendix K.7). We hope that this response clarifies the meaning of the soft cost.

(to be continued...)

(continued...)

Comment 3. The fidelity of the ported placers to the original methods.

Response. We thank the reviewer for the question. To directly address this question, we present an additional parity check of each ported placer against its released code. Since the original benchmarks of these placers fix the shape of every block, the paper verifies each port against the released code and weights (Appendix B.4). Specifically, each ported placer retains its released network and objective, together with its post-hoc loop where one is released, and adopts the latent representation, the conditioning, the sampler, and the overrides that Trinity uses (Section 4.1), and Appendix B.4 lists every adaptation. With the released weights, where these are available, each ported network reproduces the forward pass of its released code to within 1.4×10−61.4 \times 10^{-6}, and the ported guidance and loop energies match the released ones to within a relative difference of 4×10−124 \times 10^{-12}. For two settings in which a port departs from its release, namely the legality weight of the ChipDiffusion loop and the spacing of the global node in DiffPlace, the released settings change the legalized cost by at most 0.012. Moreover, we kindly highlight that Table 6 keeps each ported generator fixed and changes only the correction applied to its samples. Through its own pipeline, the legalized cost of the four ported placers is 1.525 to 1.567, and with our refiner on the same generators it decreases to 1.296 to 1.355 (Tables 3 and 6). Since the generator and its adaptation are unchanged in this comparison, the improvement arises from the correction alone. Furthermore, the training objective of FlowPlace on our larger backbone yields a raw soft cost of 2.60, compared with 2.46 on its own network (Appendix P), which places the gap in the training objective rather than in the network size. These results validate that the ports are faithful to the released methods, and that the gap to Trinity arises from the training objective and the correction of each placer rather than from its adaptation to FloorSet. We hope that these results answer the question of the reviewer. The revised paper reports the parity check in Appendix B.4.

Comment 4. The selection among draws without the reference layout.

Response. We thank the reviewer for the question. To directly address this question, we present an additional comparison in Table 58, in which the draw of each case is selected by a rule that requires no reference layout. Specifically, the reference enters the hard cost only through the two relative gaps, in which the wirelength and the outline area of every draw of a case are divided by the same two reference values. The rule replaces these two values with the medians of the wirelength and the outline area over the draws of the case. On the FloorSet validation set at 32 sampling steps and 400 refinement steps, over three seeds, the draw selected by this rule among 48 draws reaches a mean hard cost of 1.016, compared with 1.015 for the selection by the hard cost, and the two selections choose the same draw in 90% of the cases. With four and 16 draws, both selections reach 1.047 and 1.025. These results validate that the selection among draws does not rely on the reference layout. We hope that these results answer the question of the reviewer. The revised paper reports this comparison in Appendix K.6.

(to be continued...)

(continued...)

Comment 5. The refiner and the legalizer from a non-learned starting point.

Response. We thank the reviewer for the question. To directly address this question, we present an additional experiment in Table 41, in which the same refiner and legalizer start from two layouts that no model generates, namely the spectral embedding of the netlist that the reviewer requested and uniformly random positions. On our held-out validation set of 12,000 cases, the legalized cost of the first draw from the spectral embedding is 17.38 without refinement, 10.68 after 400 refinement steps, and 3.61 after 5,000 steps. From the random positions, it is 23.80, 15.86, and 5.02. In contrast, the samples of Trinity at one sampling step reach 1.490 without refinement and 1.222 after 400 steps. The non-learned starts therefore remain far above the samples of Trinity, even with more than twelve times the refinement budget of our pipeline. These results validate that the generative model supplies the arrangement of the blocks, which the refiner and the legalizer do not recover from a non-learned start. We hope that these results answer the question of the reviewer. The revised paper reports this experiment in Appendix I.5.

Comment 6. The performance on designs beyond the training range.

Response. We thank the reviewer for the question, and respectfully clarify that Trinity transfers to designs beyond its training range without retraining. Specifically, Trinity was trained on FloorSet cases with 21 to 120 blocks, and the same pipeline was run unchanged on the GSRC and MCNC benchmarks of prior floorplanning work, which contain up to 300 blocks (Appendix L.2). The pipeline fits the outline on every case of the two benchmarks (Table 66). These cases include the designs with 200 and 300 blocks, and the design with 300 blocks, 2.5 times the largest training size, fits the outline with a time limit of 60 seconds for the solver of the legalizer. On the designs with 100, 200, and 300 blocks, the wirelength of Trinity is 3% to 4% shorter than that of our run of the released PARSAC with 16 workers, and it is below the published wirelength of GoodFloorplan, a learned floorplanning policy, at all three sizes (Table 8). Regarding real designs, MCNC, GSRC, and FloorSet are the floorplanning benchmarks with soft shapes that Table 11 lists, and the paper reports Trinity on each of them. We hope that this response clarifies the performance of Trinity beyond its training range.

Reviewer_sfcrOpenReview ↗

Post-rebuttal assessment

Thank you for the detailed responses and additional experiments. The reference-free draw-selection results, non-learned initialization baselines, and end-to-end training-loss ablation address my main concerns. I also appreciate the clearer distinction between improvements to raw samples and the refiner’s contribution to the final legalized result. The results beyond the training-size range provide useful additional evidence. I am maintaining my rating of 3 (weak acceptance).

Reviewer_1NCAOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

Trinity proposes a generative floorplanning framework that integrates a shared differentiable physics-based energy function into model training, post-generation refinement, and layout scoring. It combines rectified-flow generation with gradient-based refinement and an LP-based legalizer to produce high-quality, hard-feasible layouts. Experiments show that physics regularization and refinement improve layout quality and reduce the repair effort required during legalization.

Critical strengths
  • Novel representation: Introduces an area-preserving latent parameterization that enforces target block areas by construction while providing a compact representation for generative floorplanning.

  • Unified architecture: Combines rectified-flow generation, physics-based gradient refinement, and an LP-based legalizer in an integrated pipeline to improve layout quality and produce hard-feasible floorplans.

  • Experimental improvements: Demonstrates that physics-based refinement substantially improves legalized layout quality.

Critical weaknesses
  • Poor paper organization and readability: The paper is difficult to follow, with a weak narrative flow that makes it challenging to identify the central research problem, methodological contributions, and key takeaways.

  • Insufficient methodological detail in the main paper: Many important aspects of the proposed framework and its implementation are deferred to the appendix, requiring readers to repeatedly consult supplementary material to understand the method.

Equation (1) introduces ΔHPWL+\Delta_{\mathrm{HPWL}}^+ and Δarea+\Delta_{\mathrm{area}}^+ without defining the underlying relative gaps; these definitions are deferred to Appendix C. Since these quantities are central to the proposed evaluation metric, their definitions and the reference values used for normalization should be provided alongside Equation (1).

  • Insufficient experimental analysis: The main paper does not provide enough details to interpret the experimental results. Important analyses, such as layer-wise improvements in metrics and detailed ablations, are missing, making it difficult to understand where the reported gains originate.

Although using a common legalizer is a reasonable pipeline-level comparison, the paper should report more detailed legalization statistics for each method, including expansion, displacement, LP success rates, and fallback-stage frequencies. These statistics would help clarify whether Trinity's final advantage comes primarily from better initial layouts or from greater compatibility with the proposed legalizer.

  • Contributions are not clearly distinguished: The respective contributions of the generative model, physics-based refinement, and legalizer are not sufficiently disentangled, making it difficult to assess the effectiveness and novelty of each component independently.

  • Overly compressed presentation: The paper reads like a 30-page manuscript squeezed into a 10-page format. This compression comes at the expense of clarity, leaving readers to reconstruct the methodology, experimental setup, and main contributions rather than presenting a coherent and accessible argument.

Questions for authors

Why did you not directly describe the architecture first, followed by the need for the new representation, and then the learning losses?

What are the actual reported metrics? Are they actual gaps from the reference? How come the spectral encoding and information mover additions actually increase the gaps? The lower the gaps, the better, right?

2 comments

We thank the reviewer for the feedback and the effort dedicated to the review, and would like to respond to each comment as follows. The table and figure numbers in our responses follow the revised paper, and we kindly refer the reviewer to it.

Comment 1. The organization and the density of the paper.

Response. We sincerely thank the reviewer for the suggestions on the presentation. With the additional page that ICLR allows in the rebuttal phase, we have adjusted the layout of the main text to provide more space, and the changes are highlighted in color. In particular, Section 3 now opens with an overview of its three parts, and the two relative gaps of Eq. (1) and their reference values are defined directly after Eq. (1). The camera-ready version will further improve the flow of the main text along the suggestions of the reviewer. We hope that these revisions improve the readability of the paper.

Comment 2. The legalization statistics, and whether the advantage arises from the layouts or from the legalizer.

Response. We thank the reviewer for the question. To directly address this question, we present a comparison under a second legalizer that shares no code or energy with ours, together with the legalization statistics of every pipeline. First, when the refined layouts of every generator are legalized by a procedure built from the PARSAC annealer in place of our legalizer, Trinity remains ahead of the generators of all four prior placers, with a cost of 1.871 against 1.901 to 2.000 (Table 59 and Appendix K.7). Second, Table 61 reports the expansion that our legalizer applies, the share of draws that its first linear program completes, the share that reaches the fallback stages, and the displacement of the blocks. For the refined layouts of Trinity, the median expansion is 1.32, the first linear program completes 24.5% of the draws, and the mean displacement is 0.017 of the layout scale. For the four prior placers with their own post-hoc loops, the median expansion is 1.39 to 1.47, the first linear program completes 5.4% to 9.4% of the draws, and the mean displacement is 0.035 to 0.038. The legalizer therefore displaces the blocks of Trinity less than those of the prior placers. These results validate that the advantage of Trinity holds with a different legalizer, and that its blocks are displaced less during legalization. We hope that these results answer the question of the reviewer. The revised paper includes Table 61 in Appendix K.9.

Comment 3. The contribution of each component.

Response. We thank the reviewer for the comment, and kindly highlight that the paper reports an ablation for each of the three components, namely Tables 1, 2, and 6 for the generator, Tables 5 and 6 for the refiner, and Tables 59 and 61 for the legalizer. First, on the generator, Table 1 evaluates its components one at a time, Table 2 evaluates the six functions of the energy one at a time, and Table 6 compares generators under the same refiner and the same legalizer, where Trinity reaches a legalized cost of 1.234 and the generators of the four prior placers reach 1.296 to 1.355. Second, on the refiner, Table 5 traces the soft cost along the refinement, and Table 6 compares every generator with and without the refiner, where the refiner lowers the legalized cost of every generator, for example from 1.513 to 1.234 for Trinity. Third, on the legalizer, which every pipeline shares, Table 59 replaces it with a second legalizer, under which Trinity remains ahead of the four prior placers, with a cost of 1.871 against 1.901 to 2.000, and Table 61 reports its statistics for each pipeline. Furthermore, as an additional ablation of the generator, Table 41 replaces it with a non-learned start, and the final cost after the same refiner and legalizer is 10.68, whereas it is 1.222 with the generator of Trinity. This result validates the importance of the generator. We hope that this response clarifies the contribution of each component.

(to be continued...)

(continued...)

Comment 4. The reported metrics and the gaps in Table 1.

Response. We thank the reviewer for the question, and are pleased to clarify the metrics. The wirelength and area columns report the relative gaps to the reference layout of each case, in percent, and a lower gap is better. A negative gap indicates a wirelength or an outline below that of the reference, and the cost counts only the positive part of each gap (Section 3.1). Regarding the two graph components, their effect is a trade-off between the overlap and the two gaps. Without the energy in training, the two components together lower the overlap ratio from 0.026 to 0.022, while the wirelength gap moves from -0.21% to 0.22% and the area gap from 1.04% to 1.81%. This trade-off reflects the coupling between the constraints and the objectives that Section 1 describes, in which a reduction of overlap may sacrifice the wirelength and the outline area. The soft cost, which weighs all of these quantities, decreases from 2.09 to 2.02. With the energy in training, the complete model keeps the overlap ratio at 0.023, the two gaps return to -0.07% and 0.22%, and the soft cost reaches 1.51, the lowest in Table 1. We hope that this response clarifies the metrics.

Comment 5. The flag for ethics review.

Response. We respectfully note that the review flags the submission for ethics review, and would like to provide the relevant facts. This work uses only public benchmarks, namely FloorSet, GSRC, and MCNC, involves no human subjects or personal data, and states its use of AI tools in the AI use statement of the paper. We would appreciate it if the reviewer could indicate the aspect behind the flag, and we would be glad to address it.

Comments on the paper

1 comment

Dear Reviewers,

The authors have posted their rebuttals. As the author-reviewer discussion period is coming to a close soon, please read their responses and engage in the discussion. Thank you for your time and effort.

Best,

AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.