Papers
arxiv:2604.03144

InCoder-32B-Thinking: Industrial Code World Model for Thinking

Published on Apr 3
ยท Submitted by
taesiri
on Apr 6
#3 Paper of the day
Authors:
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,

Abstract

Industrial software development lacks expert reasoning traces for hardware constraints, so a model was trained on error-driven reasoning chains and domain-specific execution traces to generate high-quality code reasoning and performance.

AI-generated summary

Industrial software development across chip design, GPU optimization, and embedded systems lacks expert reasoning traces showing how engineers reason about hardware constraints and timing semantics. In this work, we propose InCoder-32B-Thinking, trained on the data from the Error-driven Chain-of-Thought (ECoT) synthesis framework with an industrial code world model (ICWM) to generate reasoning traces. Specifically, ECoT generates reasoning chains by synthesizing the thinking content from multi-turn dialogue with environmental error feedback, explicitly modeling the error-correction process. ICWM is trained on domain-specific execution traces from Verilog simulation, GPU profiling, etc., learns the causal dynamics of how code affects hardware behavior, and enables self-verification by predicting execution outcomes before actual compilation. All synthesized reasoning traces are validated through domain toolchains, creating training data matching the natural reasoning depth distribution of industrial tasks. Evaluation on 14 general (81.3% on LiveCodeBench v5) and 9 industrial benchmarks (84.0% in CAD-Coder and 38.0% on KernelBench) shows InCoder-32B-Thinking achieves top-tier open-source results across all domains.GPU Optimization

Community

This is an automated message from the Librarian Bot. I found the following papers similar to this paper.

The following papers were recommended by the Semantic Scholar API

Please give a thumbs up to this comment if you found it helpful!

If you want recommendations for any Paper on Hugging Face checkout this Space

You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: @librarian-bot recommend

๐Ÿš€ Introducing InCoder-32B-Thinking: The First Industrial Code Thinking Model

We are excited to release InCoder-32B-Thinking, a 32B-parameter thinking-augmented code model that bridges general programming intelligence with the rigorous demands of real-world industrial software development โ€” spanning chip design, GPU kernel optimization, embedded systems, and 3D modeling.

๐Ÿ”‘ Key Highlights

  • 81.3% on LiveCodeBench V5 โ€” the highest among all open-weight models, outperforming models with 10ร— more parameters.
  • 84.0% compile pass rate on CAD-Coder and 38.0% on KernelBench L2, establishing the strongest open-source results across all evaluated industrial domains.
  • 70.4% on SWE-bench Verified, demonstrating competitive agentic coding capability.
  • Top-tier performance across 14 general and 9 industrial code benchmarks.

๐Ÿ’ก What Makes InCoder-32B-Thinking Different?

Unlike conventional code models, InCoder-32B-Thinking is trained through two synergistic innovations:

  1. Error-driven Chain-of-Thought (ECoT) Synthesis โ€” Generates reasoning traces by explicitly modeling iterative error-correction processes, mimicking how real engineers diagnose and fix hardware-level bugs through multi-turn interaction with execution environments.

  2. Industrial Code World Model (ICWM) โ€” The first world model for industrial code environments, trained on domain-specific execution traces (Verilog simulation logs, GPU profiling data, compiler diagnostics, embedded system outputs). ICWM learns causal dynamics between code and hardware behavior, achieving 96.7% outcome prediction accuracy across five industrial domains โ€” enabling large-scale data synthesis without expensive real toolchain execution.

๐Ÿ“Š Benchmark Results at a Glance

Domain Benchmark InCoder-32B-Thinking Best Competitor
General LiveCodeBench V5 81.3% 80.2% (Qwen3-Coder-480B)
General SWE-bench Verified 70.4% 74.8% (InCoder-32B)
Chip Design RealBench Module Syn@1 63.1% 23.1% (Kimi-K2.5)
GPU Optimization KernelBench L2 38.0% 28.0% (Claude-Sonnet-4.6)
3D Modeling CAD-Coder Compile 84.0% 77.0% (Claude-Sonnet-4.6)
Embedded Systems SuperCoder Acc 93.0% 88.0% (Claude-Sonnet-4.6)
Code Optimization EmbedCGen Main 47.9% 79.0% (Claude-Sonnet-4.6)

๐Ÿ”— Resources

Sign up or log in to comment

Get this paper in your agent:

hf papers read 2604.03144
Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash

Models citing this paper 0

No model linking this paper

Cite arxiv.org/abs/2604.03144 in a model README.md to link it from this page.

Datasets citing this paper 0

No dataset linking this paper

Cite arxiv.org/abs/2604.03144 in a dataset README.md to link it from this page.

Spaces citing this paper 0

No Space linking this paper

Cite arxiv.org/abs/2604.03144 in a Space README.md to link it from this page.

Collections including this paper 1