arxiv: 2603.28733 · v2 · submitted 2026-03-30 · 💻 cs.LG

Recognition: no theorem link

See it to Place it: Evolving Macro Placements with Vision-Language Models

Ikechukwu Uchendu , Swati Goel , Karly Hou , Ebrahim Songhori , Kuang-Huei Lee , Joe Wenjie Jiang , Vijay Janapa Reddi , Vincent Zhuang

Authors on Pith no claims yet

Pith reviewed 2026-05-14 21:52 UTC · model grok-4.3

classification 💻 cs.LG

keywords macro placementvision-language modelschip floorplanningevolutionary optimizationwirelength reductionelectronic design automationfoundation modelsphysical design

0 comments

The pith

A pre-trained vision-language model can guide evolutionary search to produce better macro placements on a chip canvas without any fine-tuning.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper shows that a general-purpose vision-language model can propose useful subregions on the chip canvas to constrain where macros should be placed. These proposals are fed into a base placer and then refined over multiple rounds of evolutionary optimization that scores each full placement by wirelength and other quality measures. On standard open-source benchmarks the resulting VeoPlace system beats the strongest prior learning-based placer on nine of ten cases, with the largest wirelength drop exceeding 32 percent. The same VLM guidance also improves an existing analytical placer on every benchmark tested. The work therefore demonstrates that visual reasoning already present in foundation models can be applied directly to a core electronic-design task.

Core claim

VeoPlace uses an unmodified vision-language model to generate subregion constraints that limit the placement actions of a base placer; these constraints are iteratively updated by an evolutionary search that keeps only the proposals leading to lower final wirelength. The method requires no training or adaptation of the language model to chip data.

What carries the argument

VeoPlace framework that couples a vision-language model’s subregion proposals with evolutionary search to iteratively refine macro placements.

If this is right

VeoPlace outperforms the best prior learning-based placer on 9 of 10 open-source benchmarks.
Largest wirelength reductions exceed 32 percent relative to that prior method.
The same VLM guidance improves the analytical placer DREAMPlace on all 8 tested benchmarks, with gains reaching 4.3 percent.
The approach works with both learning-based and analytical base placers.
Foundation models can be inserted into existing electronic-design flows without retraining.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same VLM-constraint pattern could be tested on routing or timing-optimization subproblems in physical design.
If the method scales, the need for large chip-specific training sets in placement tools could shrink.
Designers might review and edit the VLM-proposed subregions in an interactive loop.
The evolutionary loop could be replaced by other search methods such as Monte-Carlo tree search while keeping the VLM proposals fixed.

Load-bearing premise

A general-purpose vision-language model already possesses enough spatial reasoning to suggest useful placement constraints on a chip canvas without any exposure to chip-design examples.

What would settle it

On a fresh macro-placement benchmark, if placements produced with the VLM proposals consistently show higher wirelength than the same base placer run without those proposals, the claim that the VLM supplies useful guidance would be refuted.

read the original abstract

We propose using Vision-Language Models (VLMs) for macro placement in chip floorplanning, a complex optimization task that has recently shown promising advancements through machine learning methods. Because human designers rely heavily on spatial reasoning to arrange components on the chip canvas, we hypothesize that VLMs with strong visual reasoning abilities can effectively complement existing learning-based approaches. We introduce VeoPlace (Visual Evolutionary Optimization Placement), a novel framework that uses a VLM, without any fine-tuning, to guide the actions of a base placer by constraining them to subregions of the chip canvas. The VLM proposals are iteratively optimized through an evolutionary search strategy with respect to resulting placement quality. On open-source benchmarks, VeoPlace outperforms the best prior learning-based approach on 9 of 10 benchmarks with peak wirelength reductions exceeding 32%. We further demonstrate that VeoPlace generalizes to analytical placers, improving DREAMPlace performance on all 8 evaluated benchmarks with gains up to 4.3%. Our approach opens new possibilities for electronic design automation tools that leverage foundation models to solve complex physical design problems.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

VeoPlace gets benchmark wins by feeding VLM subregion suggestions into evolutionary placement, but the gains could come from the search loop rather than the VLM itself.

read the letter

The core takeaway is that this paper pairs an unfine-tuned VLM with evolutionary search to constrain macro placement to suggested subregions, and it claims solid improvements over prior learning-based placers on 9 of 10 benchmarks plus consistent lifts when wrapped around DREAMPlace. That combination is new enough to notice, and the results look practically useful for EDA work if they hold up. The approach does well by showing the same framework works on both ML and analytical base placers without any VLM fine-tuning, which keeps the method lightweight and points to a broader way foundation models might slot into physical design loops. The numbers on wirelength reduction are concrete and the generalization claim is straightforward to check. The main soft spot is the missing ablation the stress test flags: nothing replaces the VLM proposals with random, center-biased, or simple heuristic subregions while holding the evolutionary budget and fitness function fixed. Without that control, we cannot separate the VLM's spatial reasoning from the benefits of any iterative constrained search. The abstract also gives almost no detail on prompting, proposal count per round, or how subregions are encoded, which leaves reproducibility thin. This paper is aimed at people already working on ML for floorplanning and chip placement who are curious about injecting vision-language models. A reader tracking foundation-model uses in optimization would find it worth a look, but only after seeing the methods section. It deserves peer review because the idea is grounded in a real workflow and the benchmark results are specific enough to evaluate, even if the current version needs tighter controls to pin down the VLM contribution.

Referee Report

2 major / 2 minor

Summary. The paper introduces VeoPlace, a framework that uses an off-the-shelf Vision-Language Model (VLM) without fine-tuning to propose subregions of the chip canvas for macro placement. These VLM proposals are iteratively refined via an evolutionary search strategy to optimize placement quality. The authors claim that VeoPlace outperforms the best prior learning-based macro placer on 9 of 10 open-source benchmarks (with peak wirelength reductions exceeding 32%) and generalizes to improve the analytical placer DREAMPlace on all 8 evaluated benchmarks (gains up to 4.3%).

Significance. If the reported gains are robust and attributable to the VLM's spatial reasoning rather than the evolutionary loop alone, this work would demonstrate a practical zero-shot application of general-purpose foundation models to a core EDA optimization task. It could open avenues for integrating visual reasoning into physical design tools without domain-specific training or large labeled datasets.

major comments (2)

[Results / Experimental Evaluation] The manuscript does not report an ablation that replaces VLM-generated subregion proposals with random, center-biased, or heuristic sampling while keeping the identical evolutionary search loop, fitness function, iteration budget, and base placer. Without this control, the central hypothesis that the VLM supplies useful spatial constraints cannot be isolated from the contribution of iterative evolutionary optimization itself (see Abstract and Results sections).
[Experimental Setup] Benchmark comparisons lack details on baseline implementations, hyperparameter matching, statistical significance testing (e.g., multiple runs with standard deviations), and exact wirelength/half-perimeter wirelength metrics used for the 9/10 win claim and the 32% peak reduction. This makes it difficult to assess whether the gains are reproducible or load-bearing for the generalization claim to DREAMPlace.

minor comments (2)

[Abstract] The abstract states benchmark wins but supplies no experimental details, baseline descriptions, or ablation results; these should be summarized briefly even in the abstract for clarity.
[Methodology] Notation for subregion proposal generation, VLM prompting template, and the exact evolutionary operators (mutation, crossover, selection) is not fully specified, hindering reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and insightful comments, which help strengthen the experimental validation of our work. We address each major comment below and will incorporate the suggested improvements in the revised manuscript.

read point-by-point responses

Referee: [Results / Experimental Evaluation] The manuscript does not report an ablation that replaces VLM-generated subregion proposals with random, center-biased, or heuristic sampling while keeping the identical evolutionary search loop, fitness function, iteration budget, and base placer. Without this control, the central hypothesis that the VLM supplies useful spatial constraints cannot be isolated from the contribution of iterative evolutionary optimization itself (see Abstract and Results sections).

Authors: We agree that isolating the contribution of the VLM-generated proposals from the evolutionary optimization loop is essential to substantiate our central hypothesis. In the revised manuscript, we will add a dedicated ablation study that replaces VLM proposals with random sampling, center-biased sampling, and heuristic-based sampling while keeping the evolutionary search loop, fitness function, iteration budget, and base placer identical. This will quantify the incremental benefit attributable to the VLM's spatial reasoning. revision: yes
Referee: [Experimental Setup] Benchmark comparisons lack details on baseline implementations, hyperparameter matching, statistical significance testing (e.g., multiple runs with standard deviations), and exact wirelength/half-perimeter wirelength metrics used for the 9/10 win claim and the 32% peak reduction. This makes it difficult to assess whether the gains are reproducible or load-bearing for the generalization claim to DREAMPlace.

Authors: We acknowledge that additional transparency is needed for reproducibility. In the revised manuscript, we will expand the Experimental Setup and Results sections to include: full details on baseline implementations and reproduction steps; explicit hyperparameter values and matching procedures for fair comparison; results from multiple independent runs (with means and standard deviations) to support statistical significance; and confirmation that all reported wirelength metrics use half-perimeter wirelength (HPWL). These additions will directly support the 9/10 benchmark wins, the 32% peak reduction, and the generalization claims to DREAMPlace. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical benchmark-driven framework

full rationale

The paper proposes VeoPlace as an empirical combination of an off-the-shelf VLM for subregion proposals inside an evolutionary search loop for macro placement. All reported results consist of direct benchmark comparisons (wirelength, etc.) against prior methods on fixed open-source suites. No equations, fitted parameters renamed as predictions, or self-citation chains appear in the derivation; the central hypothesis is tested via external performance metrics rather than being defined into existence. The work is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract provides no explicit free parameters, axioms, or invented entities; the method rests on the pre-trained capabilities of an existing VLM and standard evolutionary search.

pith-pipeline@v0.9.0 · 5520 in / 991 out tokens · 40253 ms · 2026-05-14T21:52:25.526441+00:00 · methodology