Denser $\neq$ Better: Limits of On-Policy Self-Distillation for Continual Post-Training

• URL: http://arxiv.org/abs/2607.01763v1
• Date: Thu, 02 Jul 2026 06:24:30 GMT
• Title: Denser $\neq$ Better: Limits of On-Policy Self-Distillation for Continual Post-Training
• Abstract summary: Continual post-training enables foundation models to acquire new knowledge while preserving existing capabilities.<n>Recent work suggests that on-policy learning can mitigate forgetting, with on-policy self-distillation emerging as a particularly attractive approach.<n>Our experiments show that SDPO can accelerate in-domain specialization when teacher signals are stable and well aligned, but it struggles to generalize to out-of-distribution scenarios.
• Score: 44.59915053122529
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Continual post-training enables foundation models to acquire new knowledge while preserving existing capabilities. Recent work suggests that on-policy learning can mitigate forgetting, with on-policy self-distillation emerging as a particularly attractive approach. In this work, we revisit this optimistic view through self-distillation policy optimization (SDPO). Our experiments show that SDPO can accelerate in-domain specialization when teacher signals are stable and well aligned, but it struggles to generalize to out-of-distribution scenarios. In continual post-training, SDPO exhibits stronger forgetting and can even collapse, whereas on-policy reinforcement learning methods such as GRPO adapt more conservatively and better preserve prior capabilities. Further analyses reveal that denser self-distillation induces larger drift in both parameter space and response space, and can amplify high-frequency formatting artifacts through a self-reinforcing teacher--student loop. These findings suggest that on-policy data alone is insufficient for continual learning. Dense self-distillation can accelerate specialization when teacher targets are stable and token-level supervision is reliable, but it should not be treated as a default stabilizer for continual post-training. Our code is available at https://github.com/Moenupa/SDPO-CL.

Related papers

• Latent On-Policy Self-Distillation [55.57800223145407]
  On-policy self-distillation (OPSD) offers an effective pathway by using a privileged self-teacher to provide dense supervision on the student's own trajectories.<n>Existing methods still rely heavily on designer-specified privileged artifacts.<n>We introduce Latent On-Policy Self-Distillation (LOPD), which makes the teacher's privileged context itself learnable end-to-end from experience.
  arXiv  Detail & Related papers  (2026-08-13T10:05:51Z)
• RL Forgets! Towards Continual Policy Optimization [57.4336338996653]
  Continual post-training is becoming a central paradigm for adapting vision-language models to evolving tasks.<n>Recent work has increasingly favored reinforcement learning over supervised fine-tuning, driven by the belief that reinforcement learning is inherently less prone to forgetting.<n>We propose Continual Policy Optimization (CPO), a replay-free framework grounded in a prior-task behavioral KL objective.
  arXiv  Detail & Related papers  (2026-07-05T15:46:11Z)
• Physics-Guided Policy Optimization with Self-Distillation [13.212487988108592]
  Self-distilled policy optimization (SDPO) has become a popular paradigm for LLM post-training.<n>We propose Physics-Guided Policy Optimization (PGPO), which introduces an information-modulated step-size multiplier.<n>We show that this modulation preserves the order-1 weak-approximation guarantees of vanilla SGD, and incurs negligible overhead per iteration.
  arXiv  Detail & Related papers  (2026-06-02T13:20:39Z)
• Self-Distilled Policy Gradient [51.95967256628261]
  On-policy self-distillation is a promising source of dense supervision for sparse-reward reinforcement learning.<n>We propose SDPG, a self-distilled policy-gradient framework that combines group-relative verifier advantages with normalized standard deviation.
  arXiv  Detail & Related papers  (2026-06-02T02:31:13Z)
• Preference-Based Self-Distillation: Beyond KL Matching via Reward Regularization [18.027254451537342]
  Existing self-distillation methods largely reduce learning to KL matching toward the context-augmented teacher model.<n>We propose textbfPreference-textbfBased textbfSelf-textbfDistillation (textbfPBSD), which revisits on-policy self-distillation through a reward-regularized perspective.
  arXiv  Detail & Related papers  (2026-05-06T15:31:50Z)
• TCOD: Exploring Temporal Curriculum in On-Policy Distillation for Multi-turn Autonomous Agents [55.27396165691312]
  We identify a key limitation of vanilla OPD in multi-turn agent settings, which we term Trajectory-Level KL Instability.<n>We propose TCOD, a framework that controls the trajectory depth exposed to the student and progressively expands it with a curriculum schedule.<n> Experimental results across four student-teacher pairs show that TCOD mitigates KL escalation and enhances KL stability throughout training, improving agent performance by up to 18 points over vanilla OPD.
  arXiv  Detail & Related papers  (2026-04-27T03:38:27Z)
• Self-Distilled RLVR [57.37526213765131]
  We show that learning signals solely derived from the privileged teacher result in severe information leakage and unstable long-term training.<n>We propose textbfRLSD (textbfRLVR with textbfSelf-textbfDistillation)<n>This enables RLSD to simultaneously harness the strengths of both RLVR and OPSD, achieving a higher convergence ceiling and superior training stability.
  arXiv  Detail & Related papers  (2026-04-03T15:50:07Z)
• Hindsight-Anchored Policy Optimization: Turning Failure into Feedback in Sparse Reward Settings [7.554296968064274]
  We present Hindsight-Anchored Policy optimization (HAPO) for Reinforcement Learning with Verifiable Rewards (RLVR)<n>HAPO employs a hindsight mechanism that selectively anchors optimization to teacher demonstrations during failure.<n>We show that HAPO achieves textitasymptotic consistency: by naturally annealing the teacher signal as the policy improves, HAPO recovers the unbiased on-policy gradient.
  arXiv  Detail & Related papers  (2026-03-11T21:33:41Z)
• Simplifying Deep Reinforcement Learning via Self-Supervision [51.2400839966489]
  Self-Supervised Reinforcement Learning (SSRL) is a simple algorithm that optimize policies with purely supervised losses. We show that SSRL is surprisingly competitive to contemporary algorithms with more stable performance and less running time.
  arXiv  Detail & Related papers  (2021-06-10T06:29:59Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.