This work revisits on-policy Reverse Kullback-Leibler distillation and decomposes its objective into a teacher-fitting term and a student-entropy term, without introducing an explicit FKL branch, and proposes Adaptive Entropy Distillation (AED), which uses the teacher's entropy to dynamically calibrate token-level imitation strength.
Abstract
Knowledge distillation (KD) is widely used to transfer the capabilities of large language models (LLMs) to smaller students, but existing objectives often struggle to balance faithful imitation and robust generation. In particular, existing methods mainly combine FKL and RKL, overlooking that RKL itself provides a mechanism for adjusting the student's imitation strength. Motivated by this, we revisit on-policy Reverse Kullback-Leibler (RKL) distillation and decompose its objective into a teacher-fitting term and a student-entropy term, without introducing an explicit FKL branch. We show theoretically that the token-level optimal student distribution corresponds to a tempered variant of the teacher distribution, where the adaptive weight controls the trade-off between mode-seeking and uncertainty preservation. Guided by this insight, we propose \textbf{Adaptive Entropy Distillation (AED)}, which uses the teacher's entropy to dynamically calibrate token-level imitation strength. Experiments on instruction-following and mathematical reasoning benchmarks demonstrate that AED achieves superior overall performance and generally improves teacher--student distributional and entropy alignment.
On-policy distillation is a practical post-training recipe for large language models, supplying dense teacher supervision on the student's own trajectories. In privileged-context self-distillation, teacher and student are the same model conditioned on the same prefix, but the teacher also sees a hint or the full solution trace. This makes supervision abundant but harder to trust: the teacher can be confident about continuations its privileged view makes obvious but the student cannot yet justify. The distillation pull is strongest where teacher and student disagree most, and over many updates it accumulates into drift that degrades out-of-distribution (OOD) reasoning. We introduce GeoSD, a geometric self-distillation objective that treats this drift as movement in the student's predictive behavior and counters it in two complementary ways. A Hellinger loss scales each teacher preference by the overlap the student already shares with it, attenuating the pull on tokens the student cannot yet support. Since these pulls still compound over training, a proximal term penalizes how far the student's predictions drift from a recent checkpoint, measured as a Fisher-Rao distance. Both are distances in the same geometry of next-token distributions, and a natural-gradient update takes its steps in that geometry rather than in parameter space. Across mathematical reasoning benchmarks and three model families, GeoSD preserves the in-distribution gains of self-distillation while improving average OOD accuracy by 5.7-8.6 points over the base model, with gains holding across model scales from 1.7B to 32B. Analyzing why standard matching fails out of distribution, we find it wins agreement with the teacher by draining mass from alternatives at high-entropy states, resulting in confident agreement on wrong answers, whereas GeoSD keeps those alternatives in reach.
Large language model (LLM) post-training is essential for improving reasoning, adaptation, and alignment. Existing methods mainly follow two paradigms: reinforcement learning (RL) and on-policy distillation (OPD). However, RL relies on coarse-grained outcome supervision, resulting in difficult credit assignment and limited capability to acquire new knowledge. OPD, meanwhile, unconditionally matches teacher logits through KL divergence, which creates a dilemma: similar teachers provide little new knowledge, while substantially different teachers often yield ineffective guidance, largely restricting OPD to within-family distillation. We propose Distilled Reinforcement Learning (Distilled RL), which integrates teacher supervision into the RL objective to provide fine-grained guidance, selectively transfer new knowledge and avoid unconditional imitation. Distilled RL contains three components: reverse importance sampling with clipping, negative sample reset, and sequence-level geometric normalization. Through a concise and interpretable case study, we demonstrate that Distilled RL can effectively transfer previously unavailable knowledge from a teacher model to a student model. Extensive experiments across both within-family and cross-family distillation settings show that Distilled RL substantially outperforms standard RL and OPD in terms of both pass@1 and pass@k. Our code is available at https://github.com/597358816/Distilled-RL.
Chen Wang, Zhaochun Li, Jionghao Bai et al.· 2 citations
Recent advances in post-training Large Language Models (LLMs) increasingly rely on Reinforcement Learning with Verifiable Rewards (RLVR) or On-Policy Self-Distillation (OPSD). While OPSD provides dense, logit-level supervision, it inherently suffers from exposure bias due to the privileged information of the self-teacher. In multi-turn agentic settings, this leads to reasoning route convergence and the loss of clear optimization directions. To tackle these challenges, we introduce Contrastive Reinforced Policy Optimization (CRPO), which reformulates agentic OPSD from a contrastive learning perspective. By leveraging predictive entropy to distinguish between positive positions (reflective exploration) and negative positions (exposure bias), CRPO conducts group-wise contrast to preserve reliable, fine-grained optimization signals. Extensive evaluations across 13 challenging reasoning and deep-search benchmarks demonstrate that CRPO consistently outperforms existing reinforcement learning and self-distillation baselines, significantly enhancing training stability and generalization in long-horizon interactions.
Xingjian Wu, Junlin Liu, Xingchen Liu et al.· 0 citations
On-policy distillation is a powerful way to transfer reasoning ability from a strong teacher to a smaller student: the student samples trajectories from its own policy, and the teacher provides dense token-level supervision on the states the student actually visits. However, this supervision is not always reliable: a teacher can assign high likelihood to plausible but incorrect solutions, or low likelihood to correct student solutions that follow different reasoning paths. Unconditionally distilling the teacher can therefore reinforce bad modes or erase useful student behavior. To address these limitations, we introduce RG-OPD: Reward-Gated On-Policy Distillation that uses verifier feedback to decide when teacher logits should be trusted. RG-OPD bridges sparse verifier rewards and dense teacher logits, preserving token-level supervision while filtering misleading teacher signals. Across reasoning and coding benchmarks, RG-OPD produces stronger distilled students, outperforming both vanilla reverse-KL distillation and the recent TSD-KD baseline. At 1K generation length, RG-OPD improves over reverse-KL by 2.9 points and over TSD-KD by 4.9 points; in the long-generation setting, it improves over the untuned student by 8.2 points. Our code is available at https://github.com/UoC-tail/RG-OPD.
Mohammad Sadegh Akhondzadeh, Vijay Lingam, Atula Tejaswi et al.· 2 citations
dOPSD derives the teacher's privilege directly from the student's own denoising trajectory, evaluating masked positions using later, more-decoded steps of that same trajectory rather than an external label, so the teacher's advantage emerges from the model's own decoding process.
Logit-based knowledge distillation (KD) is pivotal for efficient model compression and cross-architecture learning. However, conventional methods typically rely on static, single-scale logit alignment, thereby overlooking the semantic evolution trajectory embedded in cross-scale prediction transitions. To bridge this gap, we propose Scale-Difference Evolution distillation (SDE), formulated in a structure-aware manner. Unlike the traditional prediction-imitation paradigm, SDE explicitly models the difference vectors between multi-scale logits to capture dynamic logical transitions from local features to global semantics. To optimize the distillation signal-to-noise ratio, SDE decouples scale differences into two complementary modules: Category-Dominant Difference (CDD), which isolates evidence fluctuations for core categories via a Top-K attention mask, and Relational Structure Difference (RSD), which preserves the consistency of inter-class topological evolution. Extensive experiments demonstrate that SDE consistently outperforms state-of-the-art methods across CIFAR-100, Tiny-ImageNet-200, CUB-200 and Stanford Cars. Notably, SDE achieves a significant 8.10% accuracy boost on the CUB-200 fine-grained benchmark, highlighting its superior capability in resolving inter-class ambiguities through cross-scale semantic modeling.
Hejie Lu· Proceedings of the 32nd ACM...· 0 citations