- 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:
- 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.
- 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)?
- 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?
- 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
(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 , and the ported guidance and loop energies match the released ones to within a relative difference of . 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.
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...)