Tailored Industrial Automation LLM

• Tailored LLM for Automation is a domain-adapted neural language model fine-tuned to generate industrial code with strict syntax and semantic checks.
• The approach leverages LoRA-based fine-tuning and iterative online refinement, boosting compile rates from 6.5% to 70% and semantic accuracy from 3% to 45%.
• The framework minimizes human labeling by using synthetic datasets and automated feedback, making it scalable for high-cost, data-scarce industrial applications.

A tailored LLM for automation refers to a domain-adapted neural LLM, often Transformer-based, that is fine-tuned and iteratively refined to generate, evaluate, and improve code or control outputs for industrial tasks. In contemporary industrial applications, such models must overcome critical limitations: public datasets rarely contain relevant domain code (e.g., IEC 61131-3 Structured Text for programmable logic controllers, PLCs), and target syntax is rigid with high penalty for incorrect outputs. The work by Wang et al. is exemplary in blending preference-based learning, automated feedback, and parameter-efficient fine-tuning to achieve state-of-the-art code generation for an industrial automation language (Haag et al., 2024).

1. Model Architecture and Data Curation

The tailored approach starts from a GPT-style Transformer with approximately 14 billion parameters (Phi-3), acting as both the baseline (ψ₀, via supervised fine-tuning, SFT) and as the backbone for iterative preference-based updates using Low-Rank Adaptation (LoRA). LoRA applies parameter-efficient fine-tuning (PEFT): the large core model is frozen, and only O(10⁶) LoRA adapter parameters are updated, optimizing computational cost and minimizing catastrophic forgetting.

ST code and natural-language intent pairs are rare in public corpora. Here, a synthetic ST dataset is created by translating APPS Python prompt-solution pairs into ST via GPT-4, followed by strict filtering for complexity and non-transferable logic. This results in ≈2 000 high-quality ST code-intent samples, forming the basis for first-phase SFT and out-of-domain benchmarking.

Sample generation is performed iteratively; every i-th iteration, a random subset of intents Iᵢ (sampled by rate β) is selected, and new candidate codes Ĉ = {Ĉ₁,…,Ĉₖ} are synthesized by the current model ψᵢ₋₁. This process dynamically surfaces new failure modes and adaptation opportunities.

2. Online Preference-Based Fine-Tuning Framework

The preference-based learning regime is novel in leveraging two automated sources of feedback: a compliant IEC 61131-3 compiler (“Rusty,” for syntax), and an LLM expert (GPT-4, for semantics).

• Syntax Evaluation: For each output, Ĉⱼ is checked by the compiler. κ(Ĉⱼ)=1 only if it compiles error-free.
• Semantic Evaluation: A secondary LLM acts as a “PLC ST expert” (not a generator), providing φ(Ĉⱼ)=1 if the code matches its associated intent.

Pairs of syntactically and semantically correct samples (κ∧φ=1) are labeled as “positive,” while others are labeled as “negative.” By pairing every P∈Ĉ_P with every N∈Ĉ_N, training tuples (I, P, N) are created: D = { (I, P, N) | κ(P)=φ(P)=1, at least one of κ(N),φ(N)=0 }. This pool supports Direct Preference Optimization (DPO), where the objective is to maximize the likelihood ratio of positive samples against negative ones, referencing the initial SFT baseline ψ_ref:

$$\theta = \arg\max_{\theta} \mathbb{E}_{(I, P, N) \sim D} \Big[ \log \sigma(\log \psi_\theta(P|I) - \log \psi_{ref}(P|I)) + \log \sigma(\log \psi_{ref}(N|I) - \log \psi_\theta(N|I)) \Big]$$

An $L_2$ regularizer $R(\theta)$ with strength $\lambda$ penalizes excessive drift from ψ_ref, tuning model specialization against generalization.

3. Iterative Online Refinement Process

Each fine-tuning iteration consists of:

1. Sampling: Fresh intent subset $I_i$ of size ⌊β·|I|⌋.
2. Generation: ψᵢ₋₁ synthesizes candidate code Ĉ.
3. Evaluation: Syntax via κ (Rusty compiler), semantics via φ (GPT-4 expert).
4. Pairing: DPO dataset $D_i = {(I,P,N)}$ created.
5. Update: ψᵢ₋₁ fine-tuned on $D_i$ (LoRA adapters only) → ψᵢ.
6. Validation: Evaluate held-out compile rate (κ), semantic rate (φ), and joint rate.

Hyperparameters include β (exploration/computational trade-off) and λ (regularization balance), both set via validation. Only those outputs passing both syntax and semantic checks are used for positive reinforcement, ensuring high specificity.

Metrics Table

Metric	Definition	Baseline (0/SFT)	GPT-3.5	Final (Iter 11)
P_compile	Fraction codes compile (κ=1)	6.5%	40%	70%
P_semantic	Fraction codes correct intent (φ=1)	3%	40%	45%
P_joint	Fraction codes compile & correct intent	0.6%	20%	39%

Compilation success rises from 6.5% to 70%; semantic rate from 3% to 45%; joint correctness to 39% (Haag et al., 2024).

4. Validation and Comparative Results

Performance was benchmarked against zero-shot and baseline GPT-3.5 Turbo models. The tailored LLM outperforms both baselines and generic models in all key metrics. Notably, continuous preference-based iteration substantially improves syntactic and semantic success even though the underlying model parameter count remains unchanged, demonstrating data-centric training advantages over simple model scaling.

The framework proves robust for industrial PLC programming, with 70% of outputs compiling and 39% simultaneously meeting intent, suitable for human-in-the-loop deployment and scalable to other industrial languages.

5. Practical Implications and Extensions

The architecture’s modularity allows extension:

• Code expert checkers and compilers can be swapped to address vendor-specific dialects or other IEC 61131-3 languages (e.g., SCL).
• The feedback loop may include unit- and integration-test frameworks as additional “semantic experts,” or multi-agent ensembles to redress single-LLM bias.
• The online DPO pipeline generalizes to other automation DSLs (ladder logic, function block diagrams).

This approach supports minimal human labeling of corpora, relying instead on automated feedback—making high-cost, data-scarce task domains tractable for LLM adoption.

6. Conclusions and Significance

By integrating compiler evaluation, LLM-generated semantic feedback, and preference-based fine-tuning within a tightly coupled online loop, this framework achieves robust, scalable, and domain-specific code generation for industrial automation. Key advances include the ability to bootstrap training with synthetic datasets, achieve substantial gains via iterative sample mining, and optimize joint syntactic and semantic accuracy. This methodology is poised to accelerate automation adoption across complex, safety-critical domains where precise code semantics and low error tolerance are necessary, setting a strong precedent for broader industrial LLM deployment (Haag et al., 2024).