Halv used 51% fewer tokens per correct answer on SWE-rebench (halv.ai)

Halv has achieved a significant improvement in the efficiency of coding-agent responses by utilizing 51% fewer tokens per correct answer in a benchmark of 20 software engineering tasks, compared to a vanilla version of the same model. The Halv toolkit not only reduced the overall token usage by 30.2% and cut costs by 24.1%, but it also solved 10 tasks, outperforming the vanilla model’s 7 correct solutions. This optimization means that Halv generated approximately 2.05 times more correct answers per token spent, highlighting its ability to convert a finite budget into more verified results. The technical innovations in Halv stem from incorporating three key components: Crux, which provides a prebuilt code map; RTK, which minimizes command output prior to model input; and Halv Engine Headroom, which preserves useful context throughout processing. These enhancements lower search uncertainty and improve the model's navigation of necessary information, leading to more accurate outcomes. As a result, Halv demonstrates not just a reduction in resources but a paradigm shift in how coding agents can maximize their performance and reliability with less expenditure. This advancement underlines the potential for more economical and effective AI systems in programming tasks, setting a precedent for future developments in the AI/ML community.