The paper studies how different configurations of privileged information (PI) affect on‑policy self‑distillation (OPSD) for large language models. By evaluating 16 PI setups across three agentic benchmarks and multiple student models, it finds that PI effectiveness varies with content, generation source, and model scale, with information‑dense trajectories and stronger generators yielding larger gains. The authors also propose KL‑based metrics, particularly Δd−_err, to predict PI utility before training, showing that selecting PI with this metric improves post‑training performance and remains beneficial when combined with GRPO.
arXiv:2608. 04794v1 Announce Type: new Abstract: Self-distillation (SD) has emerged as a compute-efficient alternative to reinforcement learning with verifiable rewards: a self-teacher, conditioned on privileged information (PI) about the answer such as a reference solution, supplies dense per-token supervision to a student that never sees it.
The paper introduces Privileged Self-Practice (PSP), a method that retains privileged information (PI) in the prompt rather than the loss during on‑policy self‑distillation for multi‑turn agents. PSP injects short per‑task instructions from an analyzer model when rollouts fail, sampling again with the instruction in context and training with the unchanged GRPO objective. Experiments on AppWorld and SWE‑bench Verified show PSP consistently outperforms plain GRPO, boosting task‑goal completion by up to 65% and resolved rate by up to 61% across three student models.
The paper examines on‑policy self‑distillation (OPSD) for multi‑turn agents, showing that using privileged information (PI) in the loss can make agents appear confident yet underperform plain RL, sometimes worse than the untrained base model. To address this, the authors propose Privileged Self‑Practice (PSP), which keeps PI in the prompt and uses it only during sampling, not in the loss. PSP consistently outperforms plain GRPO across AppWorld and SWE‑bench Verified, improving task‑goal completion by up to 65% and resolved rate by up to 61%.
arXiv:2611.06542v1 Announce Type: new Abstract: Proprietary language models exhibit strong reasoning and coding capabilities, but their output distributions are inaccessible, making fine-grained dist...
arXiv:2607. 02502v1 Announce Type: cross Abstract: On-policy self-distillation (OPSD) has emerged as a practical method for training large language models (LLMs) to reason, where a single model acts as both the teacher and the student with different levels of information access.
arXiv:2605.28791v2 Announce Type: replace-cross Abstract: On-policy self-distillation (SD) improves LLM reasoning by using teacher-side privileged information (PI) to turn sparse verifier outcomes in...
arXiv:2607. 04428v1 Announce Type: cross Abstract: Diffusion large language models (dLLMs) generate text by iteratively denoising a masked sequence, offering a parallel alternative to autoregressive models, but eliciting strong reasoning through post-training remains difficult: supervised fine-tuning is off-policy and suffers from exposure bias, while reinforcement learning gives only sparse, sequence-level rewards and is hard to apply without tractable sequence likelihoods.
SOD: Step-wise On-policy Distillation for Small Language Model Agents proposes a new framework that adaptively reweights distillation strength at each reasoning step based on step-level divergence. This approach mitigates cascading errors in tool-integrated reasoning by attenuating misleading teacher signals in high-divergence regions while preserving dense guidance where student and teacher align. Experiments on math, science, and code benchmarks show up to 20.86% improvement over the second-best baseline, with a 0.6B student scoring 26.13% on AIME 2025.
arXiv:2605.10194v2 Announce Type: replace Abstract: On-policy self-distillation (OPSD) uses a model as its own teacher under privileged context, providing token-level supervision on the model's own r...
Activation-Conditioned Self-Distillation (ACSD) is a new on‑policy self‑distillation method that uses a frozen copy of the base model to extract a steering vector by contrasting activations from self‑generated trajectories that reach verified correct answers with all other trajectories. The student learns from next‑token distributions on its own prefixes, without needing reference text or teacher parameter updates, and is used alone at inference. Across five models, ACSD achieves the highest mean accuracy on four mathematical benchmarks, with notable gains on DeepSeek‑R1‑0528‑Qwen3‑8B and LiveCodeBench v6 compared to the OPSD baseline.
Enabling agents to learn from experience and internalize it into their policy has become a central problem in self-evolving AI. 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; however, existing methods still rely heavily on designer-specified privileged artifacts (e.
VASPO introduces a new integration of self-distillation and PPO by leveraging privileged value estimation to guide language model training. The method uses a self-teacher and a critic that both evaluate the same privileged information on student-generated prefixes, allowing the critic to compute token-level PPO advantages from predicted continuation returns while the teacher modulates positive advantages via bounded probability ratios. Experiments on six mathematical reasoning benchmarks show that VASPO improves mean accuracy over standard PPO by 2.59 and 1.81 percentage points on Qwen3-4B and OLMo3-7B, respectively, and achieves the highest mean accuracy among evaluated methods on Knights and Knaves.