GC-OPD Uses Teacher-Verifier Disagreement as the Training Signal, Lifting Qwen3-4B From 29.08 to 40.47

arXiv 2608.19181·medium signalXBlueskyLinkedInCopy link

On-policy distillation gives dense token-level guidance but, per arXiv 2608.19181 (2026-08-19), its trajectory-level scores drift progressively out of alignment with verifier rewards as input length grows, favouring locally plausible answers that skip evidence spread across a long context. GC-OPD normalizes verifier rewards and OPD scores separately within each rollout group and treats their difference as a signed disagreement residual, distributing it across tokens by relative OPD advantage rather than uniformly. Across five long-context benchmarks it raises the official Qwen3-4B average from 29.08 to 40.47 and Qwen3-8B from 35.12 to 44.65, against vanilla OPD at 39.31 and 43.56; code is at github.com/SolereZhang/GC-OPD, confirmed live.