SPOT combines data rectification via minimal Oracle edits with a KL-constrained reward objective to improve Qwen3-8B reasoning accuracy by 6.2% using only 4k math pairs while preserving old knowledge.
Here is the corrected version
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
-
Surgical Post-Training: Proximal On-Policy Distillation for Reasoning with Knowledge Retention
SPOT combines data rectification via minimal Oracle edits with a KL-constrained reward objective to improve Qwen3-8B reasoning accuracy by 6.2% using only 4k math pairs while preserving old knowledge.