InCoder-32B-Thinking: Industrial Code Generation via Error-Driven Reasoning Traces

Introducing the InCoder-32B-Thinking model and its learning system developed for industrial code intelligence. This model uses an error-driven thought chain (ECoT) technique that mimics the thinking process of engineers in areas with complex hardware constraints, such as semiconductor design, GPU optimization, and embedded systems. In particular, by building an **Industrial Code World Model (ICWM)** that learns the physical feedback of the actual tool system, it is possible to accurately predict and self-verify the behavior changes of hardware due to code modification without physical execution. This has strengthened the ability of logical reasoning beyond simple code generation, and has proven the highest level of performance that surpasses existing open source models in various benchmarks. As a result, this technology contributes to dramatically increasing the efficiency and accuracy of software development in complex industrial processes.
https://arxiv.org/pdf/2604.03144

Видео InCoder-32B-Thinking: Industrial Code Generation via Error-Driven Reasoning Traces канала Research Paper Review