Acryl's AI Inference Control Technology Paper Accepted at Major International Conference

Acryl, a company specializing in AI transformation (AX) infrastructure, announced on the 29th that its proprietary paper on AI inference control technology—which enhances inference efficiency for large language models (LLM)—has been accepted for the main conference of EMNLP 2026, the international conference in the field of natural language processing.

The core of the accepted technology is that it increases AI inference accuracy without the need for additional graphics processing unit (GPU) resources or expanding the token budget. This technology has already been commercialized and applied to Acryl’s AI infrastructure platform, GPUBASE.

Acryl's paper, titled "Bidirectional Control of AI Inference Progress Under a Restricted Token Budget (ParentingLLM)," proposes a method for flexibly adjusting the AI’s reasoning process depending on the difficulty of the problem and the current state of inference progress. If the model attempts to reach a conclusion too quickly, additional review is encouraged; conversely, if inference becomes unnecessarily prolonged, the reasoning process is summarized to allow the model to conclude its response at the appropriate time.

The research team tested a total of four models, ranging from small to large, including Google’s Gemma and Alibaba’s Qwen families, on challenging math and science problems. As a result, all models saw improved answer rates under the same token budget conditions.

Under the 1,024-token condition, the average accuracy of the four models rose from 54.1% to 58.7%, marking a 4.6 percentage point increase. In particular, the small model "Gemma-4-E4B" improved from 39% to 49.5%, a 10.5 percentage point jump. The “budget overrun” rate—cases where the model used all tokens without reaching a conclusion—also decreased from an average of 20.7% to 13.5% for the four models, and for Gemma-4-E4B, it dropped from 26.1% to 14.1%.

Acryl has also confirmed improvements in inference accuracy using its proprietary foundation model, "Areum (A-LLM)," demonstrating the technology's applicability beyond specific models. This technology is currently applied and being sold in the LLM inference environment of GPUBASE.

The company plans to further expand this technology from generative AI to physical AI. Acryl is currently conducting a national research and development project: a 65 billion won initiative to create a “collaborative intelligence physical AI-based software (SW) platform R&D ecosystem.” Within this project, Acryl is developing environments for training, inference, and serving of vision-language-action models and for heterogeneous robot operation, all based on the physical AI platform "Jonathan PAI." The company aims to connect the ParentingLLM inference control technology to this development.

Acryl intends to advance the technology further so that, even in physical AI environments, the amount of inference can be dynamically adjusted, optimizing both decision accuracy and response speed within limited computational resources.

An Acryl representative stated, “By basing our efforts on inference optimization technology already validated in real product deployments, we aim to reduce the cost burden for companies in adopting and operating AI, and will continue to expand the technology’s application from generative AI to physical AI.”

This content was produced with the assistance of AI translation services.