analiza aiqbench
This paper introduces CoRD, a novel on-policy distillation technique that addresses the supervision gap in states reachable by a teacher but not yet by the student. By leveraging maximal coupling of speculative decoding to geometrically interpolate between student and teacher distributions, CoRD significantly improves performance on math benchmarks with minimal token substitution interventions. The method demonstrates superior convergence speed compared to standard approaches across various model sizes within the Qwen family.
CoRD: Coupled Rollouts Fill the Blind Spot of On-Policy Distillation
De la arXiv
Text integralarXiv:2611.07959v1 Announce Type: new Abstract: On-policy distillation trains a student on its own rollouts, so states the teacher would reach but the student does not are seldom supervised. We introduce CoRD (coupled-rollout distillation), which keeps the on-policy loss and moves only the distribution of training states toward the teacher, drawing rollouts from a geometric interpolation of student and teacher through the maximal coupling of speculative decoding. The coupling keeps the student's token wherever the two distributions agree and otherwise substitutes a token from the part of the target the student underweights, so interventions are as rare as any exact sampler allows and recede as the student learns.
Across six base students from two model families (Qwen3 and Qwen3.5, 0.6B to 4B), CoRD raises average pass@1 over on-policy distillation by 0.9 to 2.0 points and pass@8 by 1.2 to 3.4 points on five public math benchmarks at equal steps and data. On frozen teacher-reached states, a 4B student trained with CoRD reaches in 16 steps a gap closure that on-policy distillation does not reach in 80, and substitution rules intervening at least as often in other directions fall short of CoRD.
Abstract:On-policy distillation trains a student on its own rollouts, so states the teacher would reach but the student does not are seldom supervised. We introduce CoRD (coupled-rollout distillation), which keeps the on-policy loss and moves only the distribution of training states toward the teacher, drawing rollouts from a geometric interpolation of student and teacher through the maximal coupling of speculative decoding. The coupling keeps the student's token wherever the two distributions agree and otherwise substitutes a token from the part of the target the student underweights, so interventions are as rare as any exact sampler allows and recede as the student learns.
Across six base students from two model families (Qwen3 and Qwen3.5, 0.6B to 4B), CoRD raises average pass@1 over on-policy distillation by 0.9 to 2.0 points and pass@8 by 1.2 to 3.4 points on five public math benchmarks at equal steps and data. On frozen teacher-reached states, a 4B student trained with CoRD reaches in 16 steps a gap closure that on-policy distillation does not reach in 80, and substitution rules intervening at least as often in other directions fall short of CoRD.
Machine Learning (cs.LG); Computation and Language (cs.CL)
arXiv:2611.07959 [cs.LG]
(or arXiv:2611.07959v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2611.07959 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 283 cuvinte
Subiecte conexe
RoCo-ACE: Rollout-Conditioned Online Distillation for Retention-Aware Knowledge Injection
arXiv · 29 iul. 2026
RELAY: Selective Self-Distillation with Adaptive Hand-off for Rubric-Based Reinforcement Learning
arXiv · 14 oct. 2026
When Do We Need On-Policy Distillation? Distilling on Offline Student Rollouts Is Often Better
arXiv · 9 oct. 2026
Rollout-Marginal Distillation for Long-Horizon Autoregressive Video Generation
arXiv · 30 sept. 2026