Llms Achieve 206% Improvement with Codified Expert Knowledge for AI Agents

Why It Matters

The approach proves that codified expert knowledge can dramatically elevate LLM performance, democratizing access to specialised insights and accelerating decision‑making across data‑intensive industries.

Key Takeaways

• 206% quality boost over baseline LLM visualizations.
• Framework merges classifier, RAG, and 14k expert rules.
• Non‑experts create expert‑level charts using prompts.
• Evaluated in five domains with twelve expert reviewers.
• Physics‑agnostic design supports cross‑domain deployment.

Pulse Analysis

The scarcity of domain experts has long constrained the scalability of data‑driven decision making, especially in engineering fields where visualising simulation outputs demands both technical and visual design expertise. Traditional LLMs, while fluent in language, lack the nuanced, tacit knowledge that guides the selection of appropriate chart types and interpretation of complex datasets. The recent study from Siemens and Eindhoven University introduces a systematic software‑engineering framework that captures this hidden expertise and injects it into large language models, promising a new route to bridge the knowledge gap.

The architecture combines a request classifier, Retrieval‑Augmented Generation, and a massive library of codified rules—approximately 8,000 primitive and 6,000 compositional directives—organized around visualisation‑design principles. By feeding domain‑specific code snippets and rule‑based guidance to the LLM, the system autonomously produces visualisations that match expert standards. In a controlled evaluation across five engineering scenarios, twelve evaluators rated the AI‑generated outputs at a mean quality score of 2.60 on a 0‑3 scale, a 206 % improvement over the baseline, while code variance dropped markedly.

Beyond the immediate performance gains, the framework signals a shift toward democratizing specialist knowledge. Non‑technical users can now obtain high‑fidelity visual insights with simple natural‑language prompts, freeing senior engineers to focus on higher‑order tasks. The physics‑agnostic design suggests applicability to sectors ranging from automotive to biotech, though broader validation remains necessary. Future work may integrate LLM‑as‑Judge mechanisms for automated quality assurance and expand the rule base to other data‑type domains, positioning codified expert knowledge as a cornerstone of next‑generation AI agents.