SD-Zero converts binary rewards into dense self-supervision by having a model revise its own outputs and distill the improvements back into generation, yielding at least 10% gains on math and code benchmarks.
The 10K initial responses are roughly split into 5K correct and 5K incorrect responses
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Self-Distillation Zero: Self-Revision Turns Binary Rewards into Dense Supervision
SD-Zero converts binary rewards into dense self-supervision by having a model revise its own outputs and distill the improvements back into generation, yielding at least 10% gains on math and code benchmarks.