A new preprint by researchers from Tsinghua and Microsoft Research Asia presents HarnessMM, a method that enables coding agents to avoid certain predictable calls to language models. On DeepSeek Harness, token use per task fell from 96.28k to 11.62k, an approximately 88% reduction, while the success rate increased from 71.43% to 85.71%. On MBPP-sanitized, tokens per task dropped from 53.98k to 10.61k, and the success rate edged up from 90.27% to 91.83%. Applied to Codex on HumanEval, the method reduced token usage by 29.54% while maintaining a 90.82% success rate. The gains come from optimizing call orchestration and do not require retraining the model.
Source: Read the original article
Disclaimer: this content is for information purposes only and is not financial advice. Cryptocurrencies are highly volatile: you may lose all of your capital. Always do your own research. Legal notice
