acceptodds
Under review as a conference paper at ICLR 2027

EAPlace-VLM: Evolutionary Macro Placement via Vision-Language Topological Diagnosis and Grid-Free Continuous Refinement

Abstract

Macro placement is a fundamental yet highly challenging problem in modern chip design, characterized by large-scale combinatorial complexity, high-dimensional constraints, and strong sensitivity to early decisions. Existing methods, including analytical optimization, reinforcement learning, and heuristic-based approaches, have achieved notable progress but still suffer from inherent limitations such as difficulty in handling discrete constraints, sample inefficiency, and susceptibility to local optima caused by sequential decision-making. In particular, heuristic-driven relocation strategies often exhibit limited global awareness, leading to suboptimal solutions. To address these challenges, this paper proposes EAPlace-VLM, a novel dual-engine evolutionary framework that integrates vision-language models (VLMs) into macro placement. The proposed method extends traditional heuristic optimization into a unified paradigm combining multimodal semantic reasoning and continuous physical refinement. Specifically, a VLM is incorporated as a low-frequency topological diagnostician within the evolutionary loop to guide macro relocation decisions, thereby alleviating the spatial myopia of conventional heuristics. To enable effective cross-modal reasoning, we introduce a topology-to-color visual encoding scheme that transforms hypergraph connectivity into perceptually meaningful patterns, allowing the VLM to perform zero-shot spatial analysis and identify problematic macro clusters. Furthermore, a grid-free continuous physical refinement module is developed to eliminate discretization artifacts via legality-aware sub-pixel optimization in continuous space, improving layout compactness and physical fidelity. Experimental results on the standard ISPD 2005 benchmarks demonstrate that EAPlace-VLM achieves the state-of-the-art performance. Compared to the leading baseline EGPlace, our framework significantly reduces wirelength by up to 25% while maintaining a negligible refinement computational overhead of less than 2.5%, securing the best average rank in both pure macro and mixed-size global placement tasks.

Primary area:
optimization
Keywords:
Macro Placement, Vision-Language Model, Evolutionary Algorithm, Electronic Design Automation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_vVbmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes EAPlace-VLM for macro placement. It builds on a WireMask-style evolutionary search where a replace-set of macros is removed and greedily reinserted. Most iterations choose the replace-set with a scalar heuristic (HPWL, RUDY, overlap). Every tau iterations, a VLM (GPT-4o or Qwen) looks at a rendered placement, in which each macro is coloured by its highest-fanout net, and returns up to K macro IDs. After the search, a local search in continuous coordinates removes the whitespace created by ceiling-based grid quantization. On ISPD2005, the full method has the best average rank in macro-only HPWL (Table 1) and, followed by DREAMPlace, in mixed-size HPWL (Table 2). When EGPlace gets the same refiner, the average gain drops to about 3.1%, with wins on 6 of 8 benchmarks (Table 4).

Critical strengths

Strength 1: The VLM only picks macro IDs, and the deterministic mask-guided insertion handles coordinates and legality. This is a sensible way to use a VLM in a placement loop, since a wrong answer can only waste one iteration and never produces an illegal layout. The invalid-output fallback also keeps the search robust.

Strength 2: The continuous refiner is simple and cheap (3 to 24 minutes against a 1000-minute search, Table 9). It helps both search methods: an 8.7% average HPWL reduction for EGPlace and 7.4% for EAPlace-VLM (Table 4). I also appreciate that the paper reports the matched-refiner comparison and the congestion increase openly instead of hiding them.

Critical weaknesses

Weakness 1: The central claim, that VLM diagnosis improves the search, is not isolated. There is no run of the same framework with the VLM branch replaced by the heuristic, or by a random K-subset on the same schedule. The comparison against EGPlace does not settle this, because the paper does not state whether its heuristic branch, parent selection and population settings are identical to EGPlace. GPT-4o and Qwen give almost the same numbers, and the Appendix A rationale describes a "same-color group" that does not match the image. So a fair reading is that any periodic perturbation might give a similar effect. In addition, tau is never reported, so the reader cannot tell how many VLM calls a run actually makes.

Weakness 2: Most of the gain comes from the refiner, not the VLM. Table 1 reports up to 25.1% against EGPlace, but against EGPlace with the same refiner the mean gain is about 3.1%. That mean is driven by bigblue3 (14.2%) and adaptec2 (9.5%), and EAPlace-VLM loses on adaptec3 and bigblue2. On adaptec1, adaptec4 and bigblue4, the differences are smaller than one standard deviation over 5 seeds, and no significance test is given. The refiner also raises peak RUDY by up to 32% (Table 6), and the mixed-size gains are not uniform (Table 2).

Questions for authors

Clarification requests:

  1. What value of tau was used, how many VLM calls happen in a 1000-minute run, and is VLM latency counted inside the 1000-minute budget?
  2. Table 10 reports a constant 337 prompt tokens for every benchmark. The prompt is said to include a macro map with 543 to 1,329 entries plus an image. How does that fit in 337 tokens? Is the full macro map actually sent?
  3. Eq. 17 computes HPWL from macro lower-left coordinates and does not mention pin offsets or I/O ports. Is this the metric behind Tables 1 to 4? If so, are the copied baseline numbers from BBOPlace-Bench computed the same way? Were WireMask-EA and EGPlace rerun with the same 1000-minute budget and hardware?
  4. Several Table 6 values differ from Tables 1 and 4. For example, EGPlace adaptec2 is 37.80 vs 39.92, and EAPlace-VLM bigblue3 is 36.02 vs 39.74. Which runs does Table 6 use?
  5. In Table 2, MaskPlace + DP has the identical entry 26.14 ± 0.07 on both adaptec3 and adaptec4. Is this a copy error?
  6. K = 6 and N = 5 were chosen on the same benchmarks used for the final comparison (Tables 7 and 8). K = 6 is clearly better than K = 5 and K = 7 on adaptec3 and bigblue3, which looks more like seed noise than a smooth trend. Can you comment?
  7. Which Qwen model and version is used by default? Also, the Figure 3 caption refers to "Appendix ??".

Suggestions for future work:

A random-subset control and a text-only control (macro map without the image) would make the VLM claim much more convincing. Results on a benchmark with timing or routing evaluation would also strengthen the practical case, since ISPD2005 HPWL alone says little about downstream quality.

Reviewer_BaeDOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes EAPlace VLM, which guides VLM to select macroblocks through topological coloring, combined with evolutionary search and fine-grained optimization. On ISPD2005, the average relative EGPlace line length decreased by 12.2%, and after unified post-processing, it decreased by 3.1%, but congestion may increase.

Critical strengths

1.The proposed method is of some kind novelty. 2.Conducted experiments show effectiveness.

Critical weaknesses

1.The tesed benchmarks are old 2.Some important results are missed (i.e. DREAMPlace 4.1.0 in https://ieeexplore.ieee.org/document/11037516) 3.The ablation is insufficient. (i.e. not ablating vlm int EAPlace to validate the usefulness of vlm) 4.Some experimental hyperparameters are not disclosed. (like the version of used qwen, default value of \tau) 5.The proposed algorithm 2 has obvious defects, a simple counterexample is two 2 * 1 macroblocks can fill a 4 * 1 canvas. If the first one is placed in the horizontal range [1,3], the second one will have no feasible position.

Questions for authors

No questions.

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

This paper presents EAPlace-VLM, an evolutionary framework for macro placement that incorporates periodic VLM-guided macro selection. Given a rendered layout and structured macro information, the VLM proposes a subset of macros for relocation, while a deterministic greedy insertion operator assigns coordinates and enforces geometric legality. The layout visualization encodes partial connectivity information by coloring each macro according to its dominant incident net. After evolutionary search, a continuous physical refinement stage reduces whitespace introduced by grid quantization.

Critical strengths

The framework separates VLM-based macro selection from deterministic coordinate assignment. This design integrates model-generated proposals into a search procedure with explicit geometric constraints, without relying on the VLM to generate legal coordinates. Periodic VLM intervention provides an additional selection mechanism that complements the existing heuristic search.

Critical weaknesses
  1. The independent contribution of VLM-guided selection is not sufficiently isolated. The experiments do not clearly provide a controlled comparison that replaces periodic VLM selection with heuristic selection while preserving the remaining search procedure. Consequently, the extent to which the search improvements arise specifically from the VLM branch remains uncertain. The reported results also do not distinguish informed macro selection from the benefits of periodically introducing a different selection strategy. Furthermore, the additional value of visual input and topology-to-color encoding has not been directly established. These limitations affect the assessment of the paper’s central VLM-guided contribution.
  2. The cumulative cost of VLM intervention and the timing boundaries of the search budget are insufficiently specified. Appendix D.6 reports average per-invocation latency, but does not fully specify the number of invocations per run, their cumulative duration, or whether the 1000-minute search budget includes input preparation, request waiting time, and output processing. The reported refinement overhead of less than 2.5% of the search budget does not establish the overhead of VLM intervention or of the complete method. This ambiguity limits the assessment of overall efficiency and the fixed-budget comparisons, although it does not by itself demonstrate that the comparisons are unfair.
Questions for authors
  1. Could the author provide a controlled comparison that replaces scheduled VLM calls with heuristic selection while keeping initialization, insertion, population management, and continuous refinement unchanged? A complementary comparison with random macro selection at the same intervention frequency and with matched relocation-set sizes would help determine whether the VLM provides benefits beyond additional search perturbation. Please also identify any existing experiments that directly isolate the contributions of visual input or topology-to-color encoding.
  2. Does the 1000-minute search budget include image and input preparation, VLM request waiting time, and output processing? Please clarify the timing boundaries and report the actual number of invocations and cumulative invocation time per run, so that the extent to which these costs are reflected in the fixed-time comparisons can be assessed.
  3. The caption of Figure 3 contains an unresolved cross-reference, “Appendix ??,” at line 398. Please replace it with the correct appendix reference.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.