Standard supervised fine-tuning trains a model to reproduce a single expert demonstration or gold response, effectively treating it as a one-hot target. According to the abstract, this can push down other plausible continuations the model would otherwise consider.
The authors propose Gold-Conditioned Distillation, a self-distillation method aimed at fixing this issue. The name suggests that the gold response is used as a condition to guide the model's own output distribution, rather than as a hard target. The abstract does not provide further details on the mechanism.
If successful, the approach could help fine-tuned models keep a wider range of valid outputs. But this is only a brief announcement, so the full paper is needed to assess the method.