When Verilog and CUDA Fail: How Industrial Code Models Are Learning to Think Before They Write

Problem

Industrial code must be validated against real hardware, toolchains, and constraints. A Verilog module may be syntactically correct yet fail during simulation or synthesis; a CUDA kernel may appear logical but encounter out‑of‑bounds errors due to grid configuration, index mapping, or memory limits; an embedded program may not run because of register ordering or interrupt logic.

InCoder‑32B Thinking

Developed by Beihang University and partners, the model introduces two mechanisms tailored to industrial code.

Error‑driven Chain‑of‑Thought (ECoT)

The training loop follows generate → execute → error → fix . The model learns the full trajectory of locating a problem, fixing it, and re‑validating, rather than only the final answer. This enables handling of errors such as GPU kernel out‑of‑bounds caused by mismatched shapes and index mapping, or RTL compilation failures due to illegal port declarations.

Industrial Code World Model (ICWM)

ICWM acts as a world simulator for industrial code: given a task environment and candidate code, it predicts whether the code will pass, fail to compile, raise runtime errors, or under‑perform, and it generates diagnostic information. The paper reports a prediction accuracy of 96.7 % and multi‑turn trajectory consistency of 94.4 % , effectively substituting a real execution environment for large‑scale data generation and training.

Adaptive Thinking Depth

The model dynamically adjusts its reasoning depth based on task complexity and environmental feedback. For GPU kernel optimization the median thinking length reaches 19 015 characters , while agentic coding steps average only 91 characters —a difference of over 200× . This shows that some problems require long‑chain reasoning (performance tuning, hardware constraints) while others need short, decisive actions.

Benchmark Results

InCoder‑32B Thinking was evaluated on 14 general‑code benchmarks and 9 industrial‑code benchmarks . It remains competitive on generic tasks and achieves significant gains on industrial tasks, e.g., CAD Coder 84.0 % and KernelBench L2 38.0 % . The improvements span chip design, GPU optimization, embedded systems, compilers, and 3D modeling, indicating that the model has learned a low‑level capability to understand execution feedback, organize reasoning, and perform repairs.

Open‑Source Release

Resources:

Hugging Face: https://huggingface.co/Multilingual-Multimodal-NLP/IndustrialCoder

GitHub: https://github.com/CSJianYang/Industrial-Coder

Key Insight

When code models begin to predict the consequences of code in real industrial environments, the threshold for industrial code intelligence shifts from merely “being able to write programs” to “being able to understand systems”.