2 ms·
Just speculating but proximity to a reference answer is a much denser reward signal. In contrast, parsing out a final answer into a pass/fail only provides a sp
by markisus 1y ago
Just speculating but proximity to a reference answer is a much denser reward signal. In contrast, parsing out a final answer into a pass/fail only provides a sparse reward signal.
- krackers 1y agoYup, RLVR as implemented by Deepseek et al. use only outcome supervision instead of process supervision. There have been attempts to do process supervision though.