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

arXiv:2603.28733 · cs.LG · Submitted 2026-03-30 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "See it to Place it".

Jane: The gist: VeoPlace is a novel framework that uses a Vision-Language Model to guide low-level chip placement by constraining it to promising regions,

Tom: First, who's behind it and why it matters.

Paper summary: Tom: So, wrapping up this look at "See it to Place it: Evolving Macro Placements with Vision-Language Models," the authors present VeoPlace as a way to use VLMs for macro placement by constraining the placer to subregions of the chip canvas.

Jane: It seems like they’ve successfully shown that you don't need to fine-tune these models to get good spatial guidance; you can use them out of the box with an evolutionary search strategy.

Lu: The authors show this framework outperforms prior learning-based approaches on nine of ten open-source benchmarks with peak wirelength reductions exceeding thirty-two percent <ref:2603.28733#pg1,benchmarks with peak wirelength reductions exceeding 32>. They also showed it improves DREAMPlace performance on all eight ICCAD two thousand fifteen Superblue benchmarks by up to four point three percent <ref:2603.28733#pg2,DREAMPlace performance on all eight ICCAD 2015 Superblue benchmarks>.

Meng: The paper also highlighted that the anchor weight lambda A controls how strongly the analytical placer is guided toward VLM-suggested positions, with lower weights performing better.

Lalam: The main contributions they list are the unified VLM-guided placement framework and their performance across different placement paradigms.

Tom: So, for anyone working on electronic design automation tools, this suggests a new way to leverage foundation models to tackle complex physical design problems by using them for high-level spatial reasoning.

Conclusion: Tom: So, VeoPlace is this new framework that uses a Vision-Language Model to guide where you put big components on a chip floorplan by telling the placer exactly which areas look good to place things in.

Jane: It’s about using those visual models—the VLMs—to give direction to the placement process without needing to retrain them, and it does that by letting them propose regions.

Lu: The authors are showing how this VLM proposal mechanism works iteratively through an evolutionary search strategy, where they keep testing different placements based on what the VLM suggests.

Meng: And they’re not just doing this for one type of placement tool; they proved it works whether you’re using a learning-based policy or an analytical placer.

Lalam: The core idea is that if you have a learning model, the VLM tells it where to focus its search, and if you have an analytical placer, the VLM suggestions become soft constraints in the math.

Tom: And on their numbers, they’re seeing big improvements. They're talking about peak reductions of over thirty-two percent in wirelength after standard placement and legalization steps.

Jane: That’s a lot of reduction for physical layout problems, especially when you look at how much better it is compared to the models they tested.

Lu: They showed that on nine out of ten benchmarks, the performance gains are massive, and they even improved an analytical placer called DREAMPlace by up to four point three percent on all eight Superblue tests.

Meng: But I want to know how much overhead this adds when you try to run this in a real design flow. How big is the context window size for these VLMs?

Lalam: The framework uses a structured prompt, which they found helps the VLM give better spatial reasoning, and they kept testing different ways to pick examples from their history buffer.

Jane: It sounds like the whole point is that you can leverage these large foundation models to solve complex physical design problems by just giving them a visual hint.

Tom: Exactly. We've seen how much these models can reason spatially, and now they are being used as an active guide in the optimization process itself.

Lu: The implication here is that we might see a shift where these models don't just analyze results after the fact, but actively influence the design choices made during the layout phase.

Jane: It opens up new possibilities for how electronic design automation tools can use these powerful vision-language capabilities to tackle those tricky spatial arrangement tasks.

Harvard University · Google DeepMind

cs.LG

Submitted: 2026-03-30

Updated: 2026-10-19

Comments: NeurIPS 2026 camera-ready. 33 pages, 9 figures, 18 tables

Code: https://github.com/laiyao1/chipformer

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 88/100

The gist: The gist: VeoPlace is a novel framework that uses a Vision-Language Model to guide low-level chip placement by constraining it to promising regions, achieving state-of-the-art results across both

Key concepts

VeoPlace: A framework that uses a Vision-Language Model (VLM) to guide chip floorplanning. It iteratively evolves placement options by having the VLM suggest promising regions on the chip canvas, which then constrains the low-level placement algorithm to those areas, improving overall design quality.
VLM Guidance for Placement: The VLM generates region proposals based on prior placements to provide spatial guidance. This guidance is incorporated differently depending on the placement method: it acts as soft anchor constraints for analytical placers or masks the action space for learning-based policies, directing them toward better locations.
Evolutionary Search Strategy: VeoPlace uses an evolutionary search strategy where a population of placements is iteratively optimized. The VLM acts as a variation operator within this search, suggesting new placement regions that help the system explore the design space and converge on higher-quality floorplans over time.
Global Half-Perimeter Wirelength (gHPWL): This is the primary evaluation metric used to measure placement quality. It calculates wirelength across the entire chip after standard cell placement and legalization. Lower gHPWL indicates a better, more compact physical arrangement of components.

Terminology

Benchmark — 3× in this paper · explained in 2 episodes
Vision-language models — 2× in this paper · explained in 2 episodes
Evolutionary Search — 1× in this paper · explained in 2 episodes
Fine-Tuning — 1× in this paper · explained in 5 episodes
Foundation Model — 1× in this paper · explained in 3 episodes
Foundation Models — 1× in this paper · explained in 7 episodes
Vision-Language Models (VLMs) — 1× in this paper · explained in 6 episodes

Summary

The gist: VeoPlace is a novel framework that uses a Vision-Language Model to guide low-level chip placement by constraining it to promising regions, achieving state-of-the-art results across both learning-based policies and analytical placers without requiring any fine-tuning of the VLM

How it works

VeoPlace is designed as an evolutionary framework that harnesses the spatial reasoning capabilities of VLMs for chip floorplanning by iteratively evolving a population of placements, using a VLM as a variation operator

The core mechanism involves the VLM generating region proposals based on prior attempts to provide spatial guidance to the low-level placer This guidance is incorporated differently depending on whether a learning-based policy or an analytical placer is being used For analytical placers, suggestions become soft anchor constraints in the optimization objective For learning-based policies, suggestions mask the action space to constrain sampling

The framework operates through a feedback loop where high-performing placements are stored in a history buffer This context is then used to query the VLM for region suggestions, which constrain the low-level placer as it generates a complete floorplan Highperforming placements are stored in the history buffer, creating an evolutionary feedback loop that continuously improves placement quality

VLM and Low-Level Placer Interface

The interface between the VLM suggestions and the low-level placer is tailored to the specific placement paradigm For analytical placers, VeoPlace interprets VLM suggestions as target macro locations and incorporates them as soft anchor constraints in the optimization objective This is achieved by adding an anchor term with weight λA to the standard DREAMPlace objective function

For learning-based policies, the VLM constraints are applied by masking the policy’s output logits outside of the suggested region This constrains π to place mt within the area identified as promising by the VLM, while still retaining control over the exact placement coordinates within the region

Structured Prompt

VeoPlace utilizes a structured prompt to elicit precise spatial reasoning from VLMs and guide them toward useful spatial guidance The prompt’s key characteristics are its structure, which elicits spatial reasoning, and the selection of in-context examples This structured prompting approach guides the VLM with clear objectives, constraints, and a standardized format The full multimodal prompt performs best on superblue1; removing the image causes only a small degradation, while removing the text produces a larger drop

Context Selection Strategies for Evolution

To manage the limited context budget, VeoPlace compares five strategies for selecting in-context examples from the history buffer The Top Stratified strategy is used as the default because it balances exploration with exploitation The results show that the Default prompt yields the best objective rather than extreme greedy or exploratory variants

Experimental Results

VeoPlace outperforms state-of-the-art learning-based methods and analytical placers on open-source benchmarks, beating DREAMPlace 4.3.0 on all eight ICCAD 2015 Superblue benchmarks with gains up to 4.3% Furthermore, VeoPlace outperforms ChiPFormer on 9 of 10 benchmarks with peak wirelength reductions exceeding 32%

Design Choice Sensitivity

A key finding is that the anchor weight λA controls how strongly DREAMPlace is guided toward VLM-suggested positions Lower weights perform better, with λA = 0.

Improvements for AI systems

Improved AI system can perform Unified VLM-guided placement framework by using a high-level VLM planner to guide a low-level placer by constraining it to promising regions, which is effective without requiring fine-tuning of the VLM.
Enhanced system can outperform prior methods by demonstrating that VeoPlace outperforms ChiPFormer on 9 of 10 benchmarks with peak wirelength reductions exceeding 32%, and generalizes to analytical placers, improving DREAMPlace performance on all eight evaluated benchmarks with gains up to 4.3%.
The system can be enhanced by leveraging structured prompting, as the authors find that a structured prompt guides the VLM with clear objectives, constraints, and a standardized format, transforming its general visual reasoning into useful spatial guidance.