New method from Tsinghua and Microsoft enables DeepSeek Harness to save 87.93% in tokens.
Researchers from Tsinghua University and Microsoft propose HarnessMM, a method where models pre-generate subsequent commands with validation conditions. If tool results pass the checks, they execute directly without further LLM calls. On 98 HumanEval tasks, DeepSeek Harness success rates rose from 71.43% to 85.71%, reducing token usage by 87.93% compared to the original version; Codex maintained a 90.82% success rate and reduced token usage by 29.54%.
Continue looking at this topic
Direct useCodex repair test: when 0007 becomes 7, how do you verify the patch? ↗OpportunitiesDiscovered a skill using Claude Code to create city check-in maps, with a single post receiving 2,315 likes and 3,039 saves. ↗Hotly discussed viewsI must share this openly: the most comprehensive AI comic series workflow template I've ever created. It includes planning proposals, asset image prompts, and episode scripts.News and AnalysisCan Claude Code prevent writing outside the working directory? What boundaries still need your own verification? ↗Return to original text
A short introduction is retained here; the full text from the source website is not reposted. For specific wording and subsequent updates, please refer to the original article.
Open AIHOT original