Rubric-based reinforcement learning trains policies for open-ended generation by using natural-language criteria as rewards. According to the arXiv preprint, these rewards can become ineffective as training progresses, and the authors identify two failure modes behind this degradation.
The proposed solution, HarnessRL, co-evolves a just-in-time training harness alongside the policy. Rather than relying on a fixed reward scheme, the harness is updated continuously so that training signals remain informative throughout the learning process.
The abstract does not specify the two failure modes in detail, and the full method is not described in the excerpt. The preprint is available on arXiv under identifier 2610.26426.