analiza aiqbench
This study reveals that on-policy distillation (OPD) induces a distinct update geometry characterized by subspace locking, where training rapidly converges to a narrow, functionally sufficient low-dimensional channel. Unlike supervised fine-tuning, OPD updates affect fewer weights and avoid principal directions, suggesting it operates in a relaxed off-principal regime rather than as a simple intermediate between SFT and RLVR. These findings provide critical insights for practitioners aiming to optimize training efficiency and understand the specific parameter-space dynamics of modern reasoning model improvements.
On the Geometry of On-Policy Distillation
arXiv:2606.07082v2 Announce Type: replace Abstract: On-policy distillation (OPD) is increasingly used to improve large language model reasoning, but its training dynamics remain poorly understood. We characterize the trajectory of OPD updates in parameter space and compare it with supervised fine-tuning (SFT) and reinforcement learning with verifiable rewards (RLVR). A suite of parameter-space diagnostics consistently places OPD in a relaxed off-principal regime: compared with SFT, its updates affect fewer weights and avoid principal directions more strongly, while compared with RLVR, they remain less tightly constrained. Beyond this static localization, OPD exhibits subspace locking: its cumulative updates rapidly enter a narrow low-dimensional channel. Constraining training to the update subspace formed early in training preserves OPD performance but substantially degrades SFT, indicating that the locked subspace is functionally sufficient for OPD. Control experiments further show that sparsifying the update tokens and shifting rollout generation off-policy preserve the rank dynamics, whereas mixing the OPD objective with RLVR changes them. Overall, these results suggest that OPD is not merely an intermediate point between SFT and RLVR, but induces its own update geometry in parameter space.
Abstract:On-policy distillation (OPD) is increasingly used to improve large language model reasoning, but its training dynamics remain poorly understood. We characterize the trajectory of OPD updates in parameter space and compare it with supervised fine-tuning (SFT) and reinforcement learning with verifiable rewards (RLVR). A suite of parameter-space diagnostics consistently places OPD in a relaxed off-principal regime: compared with SFT, its updates affect fewer weights and avoid principal directions more strongly, while compared with RLVR, they remain less tightly constrained. Beyond this static localization, OPD exhibits subspace locking: its cumulative updates rapidly enter a narrow low-dimensional channel. Constraining training to the update subspace formed early in training preserves OPD performance but substantially degrades SFT, indicating that the locked subspace is functionally sufficient for OPD. Control experiments further show that sparsifying the update tokens and shifting rollout generation off-policy preserve the rank dynamics, whereas mixing the OPD objective with RLVR changes them. Overall, these results suggest that OPD is not merely an intermediate point between SFT and RLVR, but induces its own update geometry in parameter space.
Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
arXiv:2606.07082 [cs.LG]
(or arXiv:2606.07082v2 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2606.07082 Focus to learn more arXiv-issued DOI via DataCite
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 · 270 cuvinte
Subiecte conexe
Geometry-Aware Distillation for Prompt Tuning Biomedical Vision-Language Models
arXiv · 5 iun. 2026
TASSO: TAsk-Specific Subspace Optimization for Continual Learning of Vision-Language Models
arXiv · 25 aug. 2026
Beyond Output Matching: Preserving Internal Geometry in NVFP4 LLM Distillation
arXiv · 8 iun. 2026
FOCAL-VLA: Subtask-Guided Geometry Distillation and Implicit World Modeling for Vision-Language-Action Models
arXiv · 22 sept. 2026