AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

LMM4-IC4K: A Large Multimodal Model Powered Integrated Circuit Footprint Geometry Understanding

Yida Wang, Taiting Lu, Runze Liu, Lanqing Yang, Zhe Chen, Yuehai Wang, Yixin Liu, Kaiyuan Lin, XiaoMeng Chen, Dian Ding, Yijie Li, Yifan Yang and 3 more

ICML 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Multimodal LLM benchmarks

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces LMM4-IC4K, a large multimodal model-based approach and benchmark for understanding integrated circuit PCB footprint geometry from drawings.

Problem

Automated PCB footprint geometry labeling of integrated circuits is hard because footprint drawings are unstructured and abstractly annotated, and existing large multimodal models struggle with precise visual perception of these diagrams.

Approach

Adapts a large multimodal model to parse IC footprint drawings and label package geometry, with a dedicated IC footprint dataset (name suggests about 4K samples). Only the title and a truncated abstract were available, so the specific technique is not confirmed.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
train all three tasks in one chain-of-thought conversation rather than separate conversations (ablation)LMM4-IC4K · ICGeoQAoverall IoU_IC (%): 63.5 -> 71.6Improved
train tasks 1 then 2 and 3 in one conversation vs all three chained (ablation)LMM4-IC4K · ICGeoQAoverall IoU_IC (%): 62.6 -> 71.6Improved
train in three separate single-task rounds vs one chained conversation (ablation)LMM4-IC4K · ICGeoQAoverall IoU_IC (%): 25.4 -> 71.6Improved
add synthetic pre-training stage before real-data fine-tuning (vs real only) (ablation)LMM4-IC4K · ICGeoQAoverall IoU_IC (%): 65.1 -> 71.6Improved
two-stage synthetic then real vs mixing synthetic and real together (ablation)LMM4-IC4K · ICGeoQAoverall IoU_IC (%): 68.2 -> 71.6Improved
synthetic first then real vs real first then synthetic (ablation)LMM4-IC4K · ICGeoQAoverall IoU_IC (%): 24.7 -> 71.6Improved
add synthetic-only pre-training to base Qwen2-VL-7B (ablation)Qwen2-VL-7B · ICGeoQAoverall IoU_IC (%): 1.7 -> 42.9Improved
increase real-world data from 0% to 100% (ablation)LMM4-IC4K · ICGeoQAoverall IoU_IC (%): 42.9 -> 71.6Improved
add 4-shot domain examples to GPT-5GPT-5 · ICGeoQAoverall IoU_IC (%): 40.6 -> 43.0Improved
add 4-shot domain examples to Gemini 2.5 FlashGemini 2.5 Flash · ICGeoQAoverall IoU_IC (%): 16.5 -> 38.0Improved
add 4-shot domain examples to Qwen2-VL-7BQwen2-VL-7B · ICGeoQAoverall IoU_IC (%): 1.7 -> 3.8Improved
fine-tune with expert data instead of 4-shot prompting of GPT-5GPT-5 · ICGeoQAoverall IoU_IC (%): 43.0 -> 71.6Improved

Limitations

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy