What Does Multi-Harness RL Learn? Credit Assignment and Portability in Coding Agents
2026-09-07 12:00Science🔥 42.2 heat score
1sources
1days unfolding
42.2heat score
5mentions
SummaryAI generated
A study on multi-Harness reinforcement learning for code agents revealed the dominant role of evaluation in Harness design. By comparing two grouping rules under the “intra-group relative advantage” mechanism, the study found that Cross grouping rules could significantly predict the generation of high-quality Harnesses (4.48 percentage points higher than the random baseline), whereas Within rules failed to do so. In 24,000 sealed evaluations, this mechanism increased the solution rate from 2.14% to 9.27%, a 3.7-fold improvement over the training recipe. Experiments further confirmed that the improvement effects of the two grouping rules on individual Harnesses were very similar, but Cross rules offered stronger adaptability; although their portability was not superior to Within rules, and re-collecting half of the training data did not change this conclusion.
The research team isolated the “within-group relative advantage” mechanism in multi-Harness reinforcement learning in the field of code agents. By comparing the “Within” and “Cross” grouping rules, they found that the evaluation of Harness was the dominant variable: in 24,000 sealing evaluations, this mechanism increased the solution rate from 2.14% to 9.27%, a 3.7-fold increase compared to the training recipe. Experiments showed that the improvement effects of the two grouping rules on individual Harnesss were very similar (the confidence intervals included zero), and the estimated ranges of their respective seeds were greater than the difference between them; however, the inter-group advantage of the Cross rule could significantly predict the generated Harnesses (4.48 percentage points higher than the random baseline), whereas the Within rule could not do so. Additionally, re-collecting half of the training data did not change this conclusion, indicating that the Cross rule provided stronger adaptability to configurations, but its portability was not superior to that of the Within rule.