HIT researchers propose using structured data formats instead of natural language for LLM reasoning

On November 13, 2026, a Harbin Institute of Technology team unveiled a new approach that has LLMs carry out CoT reasoning through structured data formats, moving away from the lengthy natural-language reasoning style used in the past. The findings were circulated via a TMTPost write-up, which framed the work as a change in how models produce chain-of-thought. According to the team, conventional CoT output looks like essay writing, with reasoning steps buried inside verbose natural language, so they proposed encoding that reasoning as structured data instead. The method is described as cutting token consumption by 5x while also improving the model's reasoning accuracy, pairing a more compact representation with better task performance rather than trading one off for the other. No information has been released yet on implementation details, code or model releases, benchmark breakdowns, or deployment plans. At this point the work remains at the proposal stage, with the core idea and the reported token savings and accuracy gains as the only points of detail available from the public coverage so far.