Beyond LLMs: 8 AI Models and Their Impact on PLM

Artificial Intelligence (AI) is evolving far beyond Large Language Models (LLMs), introducing a rich ecosystem of specialized architectures that bring new possibilities to enterprise software. In Product Lifecycle Management (PLM), where complexity, data integration, and collaboration define success, these AI models offer game-changing capabilities. This blog explores eight categories of AI models introduced in a recent LinkedIn post by Ravit Jain, defines each, describes a PLM-specific use case, and explains how these models can be created or fine-tuned for maximum business impact.

Ravit Jain, LinkedIn post

Large Language Models (LLMs)

Definition: Large Language Models are deep learning architectures trained on billions of words to understand and generate human-like language. LLMs like GPT-4, LLaMA, and Falcon can write coherent text, answer questions, and translate languages, making them incredibly versatile.

PLM Use Case: LLMs can dramatically reduce the time spent on manual documentation by generating engineering change orders, compliance reports, audit logs, and supplier communication drafts. They can summarize intricate requirements documents and provide natural-language interfaces for PLM system queries by non-technical users.

Implementation: Organizations can fine-tune open-source LLMs using internal data such as historical product documentation, change records, and standard operating procedures. Training data is typically structured as instruction-response pairs or long-form technical writing, and models can be optimized using supervised fine-tuning or reinforcement learning methods like RLHF (Reinforcement Learning from Human Feedback).

Latent Consistency Models (LCMs)

Definition: Latent Consistency Models are a new generation of image synthesis models that generate high-quality images directly from latent representations, without the repeated refinement steps used in traditional diffusion models. This direct approach makes them significantly faster and more efficient, particularly when working with high-resolution content.

PLM Use Case: LCMs can rapidly visualize product variants, photorealistic prototypes, or virtual showroom assets. Engineers and marketers can quickly generate realistic images of new configurations without waiting for rendering teams, improving feedback cycles, and customer engagement.

Implementation: To apply LCMs, companies can use datasets of CAD renderings paired with photos or CGI imagery. Training teaches the model to map latent representations of rough design inputs into consistent, detailed images. This model can then be integrated into the PLM UI for on-demand visualization.

Large Action Models (LAMs)

Definition: LAMs are multimodal agents that combine natural language processing with the ability to execute tasks across interfaces. These models understand instructions and can interact with APIs, UIs, and data systems to complete goals autonomously.

PLM Use Case: Imagine a digital assistant embedded in PLM that takes user requests like “Start a change request for this assembly,” or “Assign this nonconformance to Quality.” LAMs interpret these commands and carry out multi-step workflows automatically, improving productivity and reducing errors.

Implementation: LAMs are created by chaining LLMs with external action libraries (e.g., via LangChain or Open Agents). To ensure traceability, they require secure access to PLM APIs, robust error-handling logic, and guardrails. Training may include simulated environments where the model learns successful action sequences.

Mixture of Experts (MoE)

Definition: MoE models distribute inferences across specialized “expert” sub-networks. Only a few experts are active per query, making the model efficient while maintaining specialization. Each expert can be fine-tuned for a specific domain or task.

PLM Use Case: PLM involves diverse domains such as engineering, procurement, manufacturing, and compliance. MoE models can smartly route queries to the right expert, e.g., one who understands stress test reports, another who estimates costs, and another who handles part classifications.

Implementation: You can develop expert sub-models by training smaller neural networks on siloed datasets (e.g., FEA analysis results, supplier quotes, compliance guidelines). Based on the input characteristics, a gating model selects which expert(s) to activate during inference.

Vision-Language Models (VLMs)

Definition: VLMs combine visual understanding with language processing, allowing them to understand diagrams, images, and textual context. This enables multimodal tasks such as visual Q&A, image-based retrieval, and annotation.

PLM Use Case: Users can upload a screenshot of a part drawing and ask, “What changes were made to this design?” VLMs interpret the image, map it to product metadata, and return relevant information. This reduces dependency on manual part ID lookups and improves accessibility.

Implementation: Pre-trained models like CLIP or BLIP can be fine-tuned using image-text pairs from PLM data (e.g., engineering drawings + metadata, markup files + descriptions). This builds semantic understanding of PLM visuals, enabling intelligent image-driven search and analysis.

Small Language Models (SLMs)

Definition: SLMs are compact language models that offer many of the capabilities of LLMs but with fewer computational requirements, enabling deployment on edge devices or embedded systems.

PLM Use Case: A technician on the factory floor can ask a handheld device, “Show me torque settings for part 9821,” and the SLM responds in real-time. This supports mobile task execution, on-the-go troubleshooting, and hands-free interaction.

Implementation: Fine-tune compact models like TinyBERT or DistilBERT on technician handbooks, FAQs, and work instructions. These models can be embedded in mobile PLM apps or IoT devices, with minimal inference latency and offline functionality.

Masked Language Models (MLMs)

Definition: MLMs, like BERT, learn to predict masked words in sentences. They are excellent at understanding context and are often used for information retrieval, classification, and auto-completion.

PLM Use Case: PLM systems often suffer from incomplete metadata or inconsistent BOMs. MLMs can fill in missing part attributes, auto-suggest category tags, or detect anomalies in structured product data.

Implementation: Fine-tune MLMs on structured PLM records where random fields are masked during training. Over time, the model learns typical patterns and can suggest corrections or completions when data is entered or audited.

Segment Anything Models (SAMs)

Definition: SAMs are advanced computer vision models that can segment any object in an image with minimal prompting, providing pixel-level masks for part recognition or inspection.

PLM Use Case: When servicing equipment in the field, technicians can take photos, and SAM can identify and segment each component. The system then links those parts to records in the PLM database for service history, replacement options, or inspection results.

Implementation: Using Meta’s SAM as a foundation, fine-tune it to product imagery labeled from internal PLM records. The model learns to associate visual segments with part numbers, supporting automated inspection, quality audits, or AR-assisted maintenance workflows.

Takeaways

These eight AI models represent a powerful toolkit for transforming PLM from a data repository into a dynamic, intelligent decision-support system. PLM vendors and enterprises can unlock next-generation performance, traceability, and user experience by aligning each model type with specific lifecycle phases and roles. The future of PLM is not just digital—it’s intelligent, multimodal, and deeply integrated with the AI frontier.

Recommendations for Moving Forward

1. Audit Your Data Readiness: Ensure your PLM data is well-structured, labeled, and accessible. AI models thrive on clean, contextualized data.
2. Start with LLM or SLM Pilots: Integrate language models to automate documentation or support front-line workers. These have the fastest time-to-value.
3. Explore Visual Intelligence: Invest in VLMs or SAMs for engineering and maintenance use cases involving images, diagrams, or CAD.
4. Use Expert Models Selectively: Apply MoEs or LAMs where decision logic or task automation is fragmented across domains or stakeholders.
5. Develop a Responsible AI Strategy: As models take on greater responsibility in lifecycle decisions, ensure traceability, governance, and human-in-the-loop oversight.
6. Collaborate with AI/ML Experts: Consider partnering with AI vendors or internal data science teams to build, fine-tune, and deploy custom models that align with your PLM strategy.

FAQs

How are multimodal AI models different from large language models (LLMs)?
Multimodal AI models process both images and text for visual analysis and query processing. LLMs focus on language, making them strong in text but limited with visual PLM data.

How do AI models beyond LLMs support engineering and manufacturing workflows?
AI models automate workflows, improve data quality, and enable image-based insights. LAMs execute tasks, while SAMs and MLMs enhance inspection and accuracy.

Why are companies looking beyond LLMs when applying AI to PLM systems?
Companies look beyond LLMs because PLM involves complex data, visuals, and workflows. Models like MoE and SLMs offer more specialized and efficient capabilities.