THEIA: A Multimodal Dataset and Benchmark for Vision-Language Analysis of Layout

Giuseppe Chiari; Michele Piccoli; Federico Viola; Davide Zoni

cs.LGSep 28, 2026

THEIA: A Multimodal Dataset and Benchmark for Vision-Language Analysis of Layout

Authors: Giuseppe Chiari, Michele Piccoli, Federico Viola, Davide Zoni

Organizations: DEIB Politecnico di Milano Milan, Italy, 20133

Abstract

The integration of artificial intelligence into computer-aided design frameworks has sparked a shift in the design of analog integrated circuits (ICs), transitioning the field from using manual and algorithmic-based solutions to adopting automated and intelligent paradigms. In this scenario, the GDSII file represents the industry-standard database containing the ultimate and most accurate source of information of the analog circuit, encapsulating the complex physical geometries and parasitic realities that define tape out performance. This paper proposes THEIA, a novel dataset containing thousands of layout images paired with question-answer conversations, along with a benchmark that employs a fine-tuned vision-language model (VLM) to analyze GDSII files of analog circuits, enabling designers to interact with and query physical layouts as intuitive, meaningful entities. Experimental results using thousands of analog designs across five realistic tasks demonstrate that the proposed fine-tuned VLM outperforms state-of-the-art general-purpose VLMs by a significant margin (up to 73%), highlighting a fundamental gap between general-purpose multimodal reasoning and domain-specific layout understanding.

Figures & tables

Figure 3 : Methodology overview. It comprises three main phases: (i) Dataset Creation , (ii) VLM Fine-tuning , and (iii) VLM Deployment . Dataset Creation generates sets of task-specific questions and answers ( Q&As ) from a collection of GDSII layouts ( GDS ). VLM Fine-tuning uses training and validation splits to fine-tune a VLM to the target domain. VLM Deployment evaluates the fine-tuned model on the held-out test data, employing an LLM-as-a-judge to assess prediction correctness.

Panel of Figure 4 — Figure 4 : Detailed methodology stages: (i) Dataset Creation (top), (ii) VLM Fine-tuning (middle) and (iii) VLM Deployment (bottom).

Model	Family	Params	LB	VE	MM Train.	Inst-tuned	Res.	Sp. Aw.
GPT-5.2 OpenAI (2025)	OpenAI	–	Proprietary	Proprietary
Gemini-3-Flash Google (2025)	Google	–	Proprietary	Proprietary
GLM-4V-9B GLM et al. (2024)	Zhipu AI	$9$ B	GLM	ViT
InternLM-4KHD Dong et al. (2024)	InternLM	$7$ B	InternLM	ViT (4K)
InternVL2 Chen et al. (2024b)	InternVL	$8$ B	InternLM	Hybrid
InternVL3.5 Wang et al. (2025)	InternVL	$8$ B	InternLM	Hybrid

Table 1 : VLMs considered in this work, compared by backbone, vision encoder, multimodal and instruction tuning, supported resolution, and spatial awareness. InternLM-4KHD and Qwen2.5-VL-7B are fine-tuned for the controlled backbone-transfer study.

Category	Circuit Type	Variants (#)	Avg. Devices (#)	NMOS (#)	PMOS (#)	CAP (#)	RES (#)	Total Devices (#)
Single Device	Capacitor	$4\,997$	$1$	–	–	$1$	–	$4\,997$
	NMOS Transistor	$5\,000$	$1$	$1$	–	–	–	$5\,000$
	PMOS Transistor	$5\,000$	$1$	–	$1$	–	–	$5\,000$
	Resistor	$5\,000$	$1$	–	–	–	$1$	$5\,000$
	Subtotal	$19\,997$	–	–	–	–	–	$19\,997$
Base Circuits	Ahuja OTA	$995$	$15$	$10$	$4$	$1$	–	$14\,925$

Table 2 : Analog circuits dataset composition and statistics. For each category, the table reports the circuit types, the number of layout variants, the average number of devices per variant, and the average per-device-type counts (NMOS, PMOS, capacitors, and resistors).

Complexity	Task ID	Task Description	Q&As (#)
Easy	A	Identification of single component devices (capacitors, resistors, NMOS, PMOS)	$19\,997$
Medium	B	Identification of base circuit topologies (OTAs, filters, regulators, gate drivers)	$5\,894$
Medium	C	Component counting and enumeration in base circuits	$27\,475$
Hard	D	Component counting and enumeration in complex mixed circuits	$19\,848$
Hard	E	Identification of base circuit topologies within complex mixed circuits	$4\,140$
Total			$77\,354$

Table 3 : Task descriptions, organized by complexity (i) Easy, (ii) Medium, and (iii) Hard.

Panel of Figure 5 — Figure 5 : Examples of GDSII view of single-component devices layouts.

Complexity	Task	Count (#)	InternLM FT		InternLM no FT	GPT-5.2	Qwen2.5-VL-7B
Complexity	Task	Count (#)	Pass@1	Pass@5	InternLM no FT	GPT-5.2	no FT	FT
Easy	A	$1\,001$	100%	100%	$41\%$	$100\%$	$19\%$	100%
Medium	B	$300$	84%	92%	$16\%$	$83\%$	$17\%$	85%
Medium	C	$1\,399$	83%	94%	$18\%$	$20\%$	$23\%$	93%
Hard	D	$1\,003$	63%	87%	$24\%$	$26\%$	$24\%$	89%
Hard	E	$207$	81%	89%	$7\%$	$10\%$	$20\%$	91%

Table 4 : Accuracy on the held-out THEIA test splits. The Qwen2.5-VL-7B transfer uses the same splits, prompts, answer format, and deterministic decoding with and without task-specific fine-tuning.

Task	THEIA test	ALIGN test	$\Delta$
B: topology identification	$84.0\%$ ( $252/300$ )	$89.5\%$ ( $34/38$ )	$+5.5$ pp
C: component counting	$83.0\%$ ( $1\,161/1\,399$ )	$73.7\%$ ( $84/114$ )	$-9.3$ pp

Table 5 : Cross-generator evaluation of InternLM-4KHD fine-tuned on THEIA and evaluated on held-out ALIGN Kunal et al. (2019) layouts.

Task	Judge Pass@1	EM	NSM	Count Acc.	Topology-set F1
A	$100\%$	$100\%$	$100\%$	–	–
B	$84\%$	$84\%$	$84\%$	–	–
C	$83\%$	$83\%$	$83\%$	$83\%$	–
D	$63\%$	$63\%$	$63\%$	$63\%$	–
E	$81\%$	$81\%$	$81\%$	–	$89\%$

Table 6 : Deterministic validation metrics for fine-tuned InternLM-4KHD on held-out THEIA samples. LLM-judge Pass@1 is shown for comparison.

Task	Count	Cohen’s $\kappa$	Accuracy	Precision	Recall / F1
A	$1\,001$	$1.0$	$100\%$	$100\%$	$100\%$ / $100\%$
B	$300$	$1.0$	$100\%$	$100\%$	$100\%$ / $100\%$
C	$1\,399$	$1.0$	$100\%$	$100\%$	$100\%$ / $100\%$
D	$1\,003$	$1.0$	$100\%$	$100\%$	$100\%$ / $100\%$
E	$207$	$1.0$	$100\%$	$100\%$	$100\%$ / $100\%$

Table 7 : Agreement of Qwen3-32B LLM judge with deterministic evaluation on held-out samples.

Figure 6 : Accuracy trends as the amount of task-specific fine-tuning data is progressively increased.

Task	Multiple choice	Open ended
A	$100\%$	$100\%$
B	$84\%$	$83\%$
C	$83\%$	$99\%$
D	$63\%$	$75\%$

Table 8 : Fine-tuned InternLM-4KHD accuracy for multiple-choice and open-ended forms of Tasks A-D on held-out layouts.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Parameter	Value / Default	Meaning / Notes
A. Base model and task formulation
base_model	internlm/internlm- xcomposer2-4khd-7b	Pretrained VLM used as initialization (loaded with custom model code).
task_type	causal language modeling	Supervised instruction tuning in a chat-style transcript format.
data_format	single-turn, single-image	Each sample contains one image and one question-answer pair (serialized with role and boundary tokens).
B. Optimization and training schedule
num_train_epochs	$1$	Number of training epochs.

Appendix

Table 9 : Fine-tuning and inference parameters.

Figure 7 : Dataset structure.

Complexity	Task	Train (#)	Validation (#)	Test (#)	Total (#)	Circuit Type Distribution
Easy	A	$17\,997$	$999$	$1\,001$	$19\,997$	NMOS: $5\,000$ ( $25.0\%$ )
						PMOS: $5\,000$ ( $25.0\%$ )
						RES: $5\,000$ ( $25.0\%$ )
						CAP: $4\,99$ 7 ( $25.0\%$ )
Medium	B	$5\,302$	$292$	$300$	$5\,894$	Gate Driver: $1\,000$ ( $17.0\%$ )
						Ahuja OTA: $995$ ( $16.9\%$ )

Appendix

Table 10 : Subsets splits cardinality and balance analysis.

	Layout area ( $\mu m^{2}$ )			Polygon count (#)			Layers (#)			Aspect ratio (-)			Routing density (-)
Topology	p25	med	p75	p25	med	p75	p25	med	p75	p25	med	p75	p25	med	p75
Dataset groups
Single Devices	$94$	$125$	$663$	$43$	$51$	$228$	$8$	$9$	$10$	$0.6$	$0.6$	$2.1$	$5.5$	$5.9$	$6.5$
Base Circuits	$2\,965$	$6\,149$	$9\,828$	$4\,073$	$4\,674$	$5\,471$	$13$	$14$	$14$	$1.0$	$1.5$	$2.1$	$3.7$	$4.0$	$4.6$
Mixed	$6\,209$	$10\,545$	$14\,887$	$9\,407$	$13\,109$	$17\,365$	$14$	$14$	$15$	$0.7$	$1.3$	$2.0$	$3.5$	$3.8$	$4.3$
Single Devices

Appendix

Table 11 : Dataset statistics across groups and circuit topologies.

Task	Layout	Question and choices	Target
A	single_cap_131	What type of device is this? A. NMOS; B. PMOS; C. resistor; D. capacitor. Pick one answer only.	D (capacitor)
B	ahuja_ota_481	What circuit is shown? A. gate driver; B. LDO; C. Miller OTA; D. HPF; E. Ahuja OTA; F. LPF. Select exactly one option.	E (Ahuja OTA)
C	ahuja_ota_308	Can you count the transistors in this circuit? A. 12; B. 14; C. 11; D. 15. Select exactly one option.	B (14)
D	mixed_8977	Can you count the PMOS devices? A. 10; B. 9; C. 8. Select exactly one option.	C (8)
E	mixed_661	What base circuits are combined? A–F enumerate topology multisets; the correct option is F: one HPF, one LDO, and one Miller OTA. Select exactly one option.	F

Appendix

Table 12 : Representative held-out Q&As retrieved from the released THEIA task files. The answer letter is the assistant target in the corresponding JSON record.

Task	3 options	4 options	5 options	6 options
C ( $1\,399$ )	$370$	$359$	$349$	$321$
D ( $1\,003$ )	$272$	$258$	$223$	$250$

Appendix

Table 13 : Number of answer options for Task-C/D held-out questions after distractor construction.

Topology	Train ( $5\,302$ )	Validation ( $292$ )	Test ( $300$ )
Ahuja OTA	$895$ ( $16.9\%$ )	$49$ ( $16.8\%$ )	$51$ ( $17.0\%$ )
Gate Driver	$900$ ( $17.0\%$ )	$50$ ( $17.1\%$ )	$50$ ( $16.7\%$ )
HPF	$865$ ( $16.3\%$ )	$48$ ( $16.4\%$ )	$49$ ( $16.3\%$ )
LDO	$890$ ( $16.8\%$ )	$49$ ( $16.8\%$ )	$50$ ( $16.7\%$ )
LPF	$873$ ( $16.5\%$ )	$48$ ( $16.4\%$ )	$50$ ( $16.7\%$ )
Miller OTA	$879$ ( $16.6\%$ )	$48$ ( $16.4\%$ )	$50$ ( $16.7\%$ )

Appendix

Table 14 : Task-B class counts (percentage within split) and majority-class baseline.

Figure 8 : Example of a base circuit layout, an HPF. Devices are highlighted in coloured boxes.

Panel of Figure 9 — Figure 9 : Examples of Miller OTA layouts.

Panel of Figure 10 — Figure 10 : Examples of mixed topologies.

Explore similar work

Sep 28, 2026cs.LG

Inspector: Conversational and Lightweight Analyzer of Analog Circuit Layouts Using LLM and CNNs

The integration of artificial intelligence into computer-aided design frameworks has sparked a shift in the design of analog integrated circuits (ICs), transitioning the field from using manual and algorithmic-based solutions to adopting automated and intelligent paradigms. In this scenario, the GDSII file represents the industry-standard database containing the ultimate and most accurate source of information of the analog circuit, encapsulating the complex physical geometries and parasitic realities that define tape out performance. This paper proposes a novel framework that combines fine-tuned LLMs and CNNs to analyze GDSII files of analog circuits, enabling a conversational interface between the tool and the designers. Experimental results using thousands of analog designs across four realistic tasks demonstrate that the proposed solution outperforms state-of-the-art general-purpose massive VLMs by a significant margin (up to 81%), thus providing a lightweight solution to the problem of GDSII analysis.

Abril Cano Castro, Giuseppe Chiari, Michele Piccoli +2

DEIB Politecnico di Milano Milano, Italy

Analog Design Analog Convolutional Neural Networks Large Language Model Fine-Tuning Conversational Context

Ask PDF

Jan 12, 2026cs.MA

VLM-CAD: VLM-Optimized Collaborative Agent Design Workflow for Analog Circuit Sizing

Vision Language Models (VLMs) have demonstrated remarkable potential in multimodal reasoning. However, they can have spatial blindness and logical hallucinations when interpreting densely structured engineering content, such as analog circuit schematics. To address these challenges, we propose a Vision Language Model-Optimized Collaborative Agent Design Workflow for Analog Circuit Sizing (VLM-CAD) designed to support step-by-step reasoning over multimodal evidence. VLM-CAD bridges the modality gap by integrating a neuro-symbolic structural parsing module, Image2Net, which transforms raw pixels into explicit topological graphs and structured JSON representations to anchor VLM interpretation in deterministic facts. To ensure the reliability required for engineering decisions, we further propose ExTuRBO, an Explainable Trust Region Bayesian Optimization method. ExTuRBO employs agent-generated semantic seeds to warm-start local searches and uses Automatic Relevance Determination to provide sensitivity evidence for the final design report. Experimental results on 12 sizing tasks covering six circuits and four technology platforms show that VLM-CAD achieves a pooled Strict Pass@1 of 23.3% and a Relaxed Pass@1 of 91.7%, while providing sensitivity evidence for final design reports.

Guanyuan Pan, Shuai Wang, Yugui Lin +4

Hangzhou Dianzi University, China · SCBC, Guangdong University of Foreign Studies, China · University of Cambridge, United Kingdom

Analog Design Multimodal Reasoning Agentic Workflows Large Language Model Agents

Ask PDF

Jun 13, 2026cs.AR

PANDA: An LLM-Enhanced Performance-Driven Analog Design Framework Bridging Design Intent and Layout Generation

Traditional design of analog circuits heavily relies on manual interventions across topology, sizing, and layout, with prior automation addressing stages in isolation. In this work, we propose PANDA, an LLM-enhanced framework that bridges high-level design intent to final layout by actively managing cross-stage dependencies through guided topology synthesis, substructure-aware sizing, and constraint-driven layout generation. This shifts automation from algorithm-centric execution to intent-centric co-design, reducing turnaround time from days or weeks to hours while improving design performance.

Haoyi Zhang, Weijian Fan, Xiaohan Gao +3

School of Integrated Circuits, Peking University, Beijing, China · Beijing Advanced Innovation Center for Integrated Circuits, Beijing, China · Institute of Electronic Design Automation, Peking University, Wuxi, China

Analog Design Layouts Co-Design Model Size

Ask PDF