Teacher-Student Learning

On-policy distillation (OPD) has become increasingly popular for transferring teacher capabilities to student models. In this work, we ask a critical research question: Is on-policy sampling always beneficial for distilling arbitrary teacher-student pairs? We show that a simple alternative, Semi-OPD, which distills from offline rollouts generated by the initial student, can often outperform OPD in both accuracy and training efficiency. Across 17 teacher-student pairs ranging from 1.5B to 235B parameters, Semi-OPD outperforms OPD in 14 cases, with up to +13.6% accuracy and 11.4x training speedup. We further find that the choice between OPD and Semi-OPD depends on the alignment between the initial teacher and student, quantified by an output-token overlap ratio: OPD is beneficial only when the two are highly aligned with high overlap ratios. Our deeper investigation suggests that effective distillation requires on-policyness w.r.t. both the student and the teacher. For misaligned pairs, student rollouts can become increasingly off-policy w.r.t. the teacher as context length grows, weakening the distillation signal. In contrast, Semi-OPD is often more stable, as it distills on shorter contexts while covering full trajectories and exposing the student to more teacher-preferred tokens. Beyond proposing Semi-OPD as an efficient alternative, our work motivates the community to rethink when to use OPD and to study stronger OPD variants with meaningful teacher-student pairs.

On-policy distillation (OPD) efficiently transfers capabilities from a stronger teacher to a student language model through dense token-level supervision on student-generated rollouts and has shown promise on complex tasks such as mathematical reasoning. However, in multi-turn agentic tasks, student decisions shape subsequent observations, causing early errors to accumulate across turns. The resulting trajectories can drift away from the teacher's rollout distribution, making the teacher's token-level supervision less reliable or even counterproductive for OPD training. To address this issue, we introduce STI-OPD, a stochastic teacher intervention framework for multi-turn agentic OPD. During multi-turn interaction, STI-OPD uses teacher intervention guided by teacher-student policy discrepancy to replace the student's proposed action with a teacher-generated one to maximize the acquisition of reliable supervision. We further develop a stochastic intervention strategy, addressing the limitations of previous threshold-based or fixed-schedule approaches, that estimates policy discrepancy using KL divergence and maps it to an intervention probability. By sampling whether to intervene from this probability, STI-OPD adaptively balances teacher control with student exploration. To learn from the resulting mixed-policy trajectories, we introduce an Importance-Weighted Reverse KL objective that corrects the token sampling mismatch between teacher-generated responses and the student policy to preserve the original OPD objective. Across tool-integrated reasoning and long-horizon interaction, STI-OPD outperforms the strongest prior OPD baseline on every evaluated benchmark and student size. Ablations further show that both discrepancy-guided intervention and importance weighting contribute to these gains.

Semi-supervised learning (SSL) relies on two core mechanisms: self-training under the Teacher-Student (T-S) framework and joint optimization of labeled and unlabeled losses. Despite their effectiveness, we find both mechanisms introduce distinct optimization pathologies. First, parameter coupling enforces strict synchronization between teacher and student, where strong regularization on the student degrades the teacher's fitting ability, thereby limiting the permissible generalization intensity. Second, the imbalance in gradient update consistency between labeled and unlabeled losses drives the shared parameters to prematurely converge to labeled-dominated local minima, creating a bottleneck for global optimization. To address both issues, we propose the Pioneer Student (PiS), an auxiliary branch that operates in an independent parameter space and periodically transfers accumulated knowledge back to the T-S model. Extensive experiments show that PiS is a universal plug-and-play module that consistently improves mainstream SSL methods.

Understanding loss landscapes is central to explaining neural-network training, yet their structure remains only partially understood even in simple models. We study the Gaussian population loss of shallow, bias-free ReLU networks in the teacher--student setting. This provides a simple model for studying essential aspects such as feature learning and overparameterization. For teacher networks with positive output weights and planar features, we show that including a learned linear skip removes all spurious local minima with non-negative student output weights once the student network is at least as wide as the teacher network. In contrast, without the skip, we construct a fixed teacher network with positive output weights and only three hidden neurons in input dimension two whose spurious local minima persist at every student width at least three. Thus, a learned linear skip can remove spurious minima that persist under arbitrary overparameterization. Furthermore, we show that a positive output weight student network always learns the subspace spanned by the teacher features: student features at local minima with non-negative student output weights lie in the span of the teacher features. For ReLU networks in two dimensions, even heavily overparameterized student networks have effective width controlled by the teacher width: every critical point with positive student output weights has at most twice as many distinct student feature directions as teacher neurons. Finally, we transfer the benignity result to empirical minima over parameter balls of any prescribed radius, with the required sampling accuracy depending on that radius.

Unsupervised domain adaptation (UDA) reduces the annotation burden in panoptic segmentation by leveraging a cost-effectively labeled source domain (e.g., synthetic) and an unlabeled target domain to bridge the distribution gap. Existing panoptic UDA methods rely on teacher-student consistency learning built upon suboptimal per-pixel segmentation architectures. In contrast, state-of-the-art mask transformers are rarely adopted due to their pronounced vulnerability to confirmation bias in consistency learning, where erroneous teacher predictions are reinforced during training. Our earlier approach, MC-PanDA, mitigates this issue through fine-grained confidence estimation, which suppresses gradients from unreliable masks while sampling informative yet reliable locations for loss computation. However, this method entails a complex multi-stage training and requires careful hyperparameter tuning. This work presents MC-PanDA++, which addresses these limitations by introducing: (i) self-supervised vision encoders that provide a stronger and more robust initialization, further reducing the reliance on human annotations, (ii) per-class, self-adapting mask-wide loss scaling that stabilizes training and enables the usage of a single set of hyperparameters across domains, and (iii) a single-stage training pipeline that decreases overall conceptual complexity. Together, these improvements result in a conceptually simpler, better-performing, and more robust method for domain-adaptive panoptics. Source code: https://github.com/martinovicivan/MC-PanDA

When a student is stuck, a tutor faces the assistance dilemma: help given too early can hinder productive struggle, while help withheld too long leaves the student in a frustrating, persistent impasse (i.e., wheel spinning). Generative AI tutors increasingly use guardrails restricting answer-giving, yet little is known about how such tutors behave once an impasse persists. We analyze 20,462 student turns from 1,260 authentic sessions with a guided LLM chemistry tutor, identifying 6,630 impasse turns of three major types: conceptual errors, expressed uncertainty, or help-seeking. We then used these impasses to simulate three tutoring conditions to study variation in AI tutor guidance through impasses: baseline, no-direct-answer, and guided tutor. For a sample of 150 impasses, prompt specificity changed pedagogy: a baseline tutor provided the answer directly in 50.7% of responses, a no-direct-answer tutor asked a follow-up question every time, and the guided tutor responded in a wide variety of ways depending on the context. We then analyzed impasse trajectories in authentic interactions, finding that each additional impasse turn lowered the odds of next-turn recovery by 12.7% (AOR = 0.873, p < .001), and early dropouts were caught in recursive concept elicitation before reaching execution. The benefit of questioning decayed as impasses persisted (scripted question x depth AOR = 0.78; follow-up x depth AOR = 0.83), whereas addressing the student's error grew more beneficial (AOR = 1.14); after a failed scripted question, repeating it was followed by recovery in 28.1% of cases, compared with 39.8% when the tutor addressed the error instead. For learning analytics, these findings identify impasse depth and type as observable, turn-level dialogue signals that analytics can use to trigger graduated, state-sensitive assistance in real time.

Reasoning distillation from powerful teacher models to smaller students faces the Gap Curse: as teachers grow more sophisticated, their complex distributions increasingly diverge from what students can approximate, causing performance degradation. Existing mitigation strategies either filter out challenging examples through data selection or introduce weaker intermediate assistant models, inherently compromising supervision coverage or quality. We propose Teacher Alignment, which directly adapts the teacher toward the student's distribution without discarding data or degrading reasoning quality. However, naive alignment through standard knowledge distillation triggers catastrophic collapse of the teacher's reasoning capabilities. To address this, we reformulate teacher alignment as reinforcement learning and introduce TeacherGRPO, built on Group Relative Policy Optimization with two key innovations: (i) Curriculum Selective Alignment applies dual token- and distribution-level curricula to focus rewards on high-signal reasoning gaps while filtering noise from trivial tokens and uncertain tail distributions, and (ii) Importance-Adaptive Length Regularization selectively penalizes verbose redundancy while preserving pedagogically critical reasoning steps. The aligned teacher then distills knowledge to students via standard pipelines. Extensive experiments show TeacherGRPO significantly outperforms baselines across diverse reasoning benchmarks and distillation methods. Our code is available at https://github.com/LzyFischer/TeacherGRPO.

Multi-organ segmentation using deep learning requires large amounts of annotated patient data; however, institutions often lack sufficiently large and diverse annotated datasets. Privacy constraints further prevent institutions from sharing patient data to overcome this limitation. Moreover, due to the labor-intensive nature of annotation and the scarcity of diverse expertise, institutions typically have labels for only a small portion of their local data, leaving the larger unlabeled portion unused. In this work, we propose a flexible semi-supervised federated multi-task student-teacher framework that leverages federated learning (FL) to improve multi-organ segmentation using both labeled and unlabeled data across participating sites. At each communication round, the proposed framework initiates local training, where clients with labels for the same task form a federation to produce an aggregated teacher model. The resulting teachers generate task-specific features for all data at each client. Subsequently, all clients form a second federation to train a multi-task student model with a shared encoder and task-specific decoders that replicate the teacher-generated features across all segmentation tasks. The aggregated student model is then used to update the local teachers and initiate the next training round. Extensive experiments demonstrated the effectiveness of the proposed method compared with local and federated single-organ models, yielding an average performance gain of 13 percent across clients. The experiments also demonstrated the impact of multi-task learning and unlabeled data and the applicability of the framework in relaxing labeled-data requirements for client participation. The code is available at https://github.com/AshknMrd/FedMust.

Interaction paradigms used in robot-assisted autism intervention have historically employed robots as teachers, clinical assistants, or more-abled peers to promote a variety of social skills. These modalities often leverage the expertise of trained practitioners to ensure that child-robot interactions are productive or clinically grounded to yield positive therapeutic benefits for children across the autism spectrum. Yet, despite the fact that the majority of children with autism spectrum disorder (ASD) attend mainstream schools and spend 80% or more of their time in the general classroom [27], there is a paucity of research incorporating validated classroom teaching pedagogies into robot-assisted autism interventions. In this work, we introduce a novel teaching methodology for advancing social skills in school-aged children with ASD. We evaluate the effectiveness of a novel robot-assisted autism intervention which incorporates the learning-by-teaching pedagogy and explores the comparative benefits of employing a robot versus a human confederate for improved performance on a set of social skills tasks. Results show that 80% of study participants performed better in the robot condition (mean performance in the robot condition=63%, mean performance in the confederate condition=37%), irrespective of the scenario order. Further, 90% of all participants were significantly more engaged in the robot condition (mean engagement: robot=61%, confederate=32%) and, while the effect did not result in the confederate condition, analyses indicate that overall engagement in the robot condition contributed to improved performance. These results suggest that robots employed in a learning-by-teaching context may help enhance engagement and improve performance on a simple social skills task for children with ASD.

Teacher-student systems, in which a teacher neural network generates training labels so that a student neural network can learn to implement the same function, are widely used as an abstract setting to study learning. However, the structure of the teachers is often overlooked by assuming randomly-generated, normally-distributed parameters. This hides substantial variation in how learnable different teachers are. We formalize learnability as the success rate of converging to the global minimum, as a function of overparameterization, learning algorithm, student initialization distribution, and teacher geometry. We both identify an easy distribution that maximizes node dissimilarity and a hard distribution that minimizes it, and show that these two distributions induce markedly different success rates across a large range of settings and for different activation functions. To explain the gap, we study the loss landscape of small neural networks that contain two distinct kinds of suboptimal local minima, out-of-bounds (OOB) minima at the edge of the data distribution and interior minima within. Assuming infinite data and a fast readout layer, we analytically reduce the loss landscape of small networks to two dimensions, showing that the region of attraction of interior minima changes as a function of teacher structure. In larger networks, maximally dissimilar teachers induce more interior minima, while minimally dissimilar teachers induce more OOB minima. Motivated by these analyses, we show that differentially increasing the learning rate of the readout layer and decreasing the learning rate of the inner biases increases success rates. These findings provide an important step in narrowing the gap between the study of teacher-student networks and more structured functions that arise in practice.

Test-Time Adaptation (TTA) has recently emerged as a promising strategy that allows the adaptation of pre-trained models to changing data distributions at deployment time, without access to any labels. To mitigate error accumulation, researchers have widely adopted the teacher-student framework, though its long-term stability is often taken for granted. In this work, we challenge the common strategy of setting the teacher weights to an exponential moving average of the student by showing that error accumulation still occurs, although it is mostly apparent on longer sequences compared to those commonly utilized. We analyze the stability-plasticity trade-off within the teacher-student framework and propose to use an intransigent teacher that does not update its weights. Surprisingly, we show that this simple change allows TTA methods to significantly improve their performance on multiple datasets with longer scenarios and result in increased robustness to changes in hyperparameters. Finally, we show that those changes can be seamlessly and effectively applied to various architectures and experimental setups, including semantic segmentation. The code is available at https://github.com/dmn-sjk/intransigent_teacher.

Learning reliable correspondences between images and point clouds is fundamental for 2D-3D matching. Despite recent progress in detection-free methods, existing approaches primarily optimize matching within a single model and often struggle to maintain reliable correspondences under challenging conditions such as noisy inputs, low overlap, and ambiguous structures. In this work, we propose TeaMatch, a novel framework that introduces teachability as a criterion for cross-modal representation learning. We define teachability as the ability of a representation to be effectively recovered by weak learners under degraded inputs, reflecting its structural consistency and robustness. To this end, we construct a set of task-specific weak students that simulate common failure modes and train them to imitate the teacher on a training split while evaluating their recoverability on a disjoint meta split. The teacher is then optimized to improve the students' ability to recover reliable correspondences, guided by correspondence-level and geometry-aware constraints. Our framework can be seamlessly integrated into existing coarse-to-fine matching pipelines without additional inference cost. Extensive experiments demonstrate that TeaMatch improves matching robustness and achieves state-of-the-art performance on challenging 2D-3D matching benchmarks.

In many real-world scenarios, encountering continual shifts in domain during inference is very common. Consequently, continual test-time adaptation (CTTA) techniques leveraging a teacher-student framework have gained prominence, allowing models to adapt continuously even after deployment. In such a framework, a weight-averaged mean teacher is used to produce pseudo-labels from test data for self-training. The mean teacher gets updated as an exponential moving average of the student parameters using a high value of momentum that is kept fixed even if different distributions of test data are encountered. To combat the resulting drift of the model, we propose a novel controlled teacher adaptation methodology that dynamically sets a proper momentum value depending on the quality of the incoming data. Additionally, we estimate class prototypes from the source pretrained model to help align the target data as they come in. Importantly, our method does not require access to source data or its statistics at any stage of the pipeline, making it truly source-free. We perform extensive experiments on benchmark datasets to demonstrate that our approach outperforms different state-of-the-art adaptation frameworks, many of which require access to source data.

On-policy self-distillation achieves strong pass@1 accuracy by using a single model as both teacher and student, with the teacher conditioned on a correct demonstration to provide dense token-level feedback. We show that this could come at a hidden cost: rollout diversity decreases and pass@k curves flatten (i.e., generating more rollouts fails to improve accuracy). We trace this to compounding biases in the design of self-distillation with sampled demonstrations. The teacher scores each student rollout while conditioned on a sampled correct rollout, channeling its feedback through the model's own biases. We theoretically analyze the optimal self-distillation policy and show that it tilts the base distribution by a pointwise conditional mutual information score between the student's rollout and the correct rollout used as context. Unlike the ideal optimal on-policy reinforcement learning (RL), which preserves probability ratios among equally correct rollouts, self-distillation can amplify existing probability gaps, concentrating mass on already-dominant modes. On a controlled graph path-finding task and science question-answering benchmarks, self-distilled models match or exceed RL on average performance but exhibit substantially lower functional and semantic diversity, failing on out-of-distribution settings that require diverse strategies.

Learning-based automated coaching systems for complex motor tasks such as high-performance driving remain limited in the ability to be adaptive by their reliance only on local, context-dependent reasoning, failing to account for the long-term temporal nature of student learning and the cumulative impact of repeated teacher-student interactions. In this paper, we propose an imitation learning based computational model for adaptive teaching with a dedicated temporal reasoning module that can reason over the interaction history under low-data regimes. To compensate for limited amounts of interactive training data, and based on the repetitive nature of the teaching process, the model relies on a nearest neighbor retrieval and cross attention prior, reasoning only on a narrowed-down set of semantically similar past interactions with an encoder-decoder based concurrent teaching model. We validate our approach with (i) a novel semi-synthetic closed-loop longitudinal student-teacher interaction dataset based on Waymo Open Motion Dataset and (ii) a small-scale real-world naturalistic simulator race coaching dataset. Our results reveal the consistent advantage of our adaptive teaching model with the nearest neighbor retrieval and cross-attention prior over a non-adaptive baseline as well as a suite of adaptive models that differ in their choice of priors and temporal fusion mechanisms.

On-policy distillation (OPD) trains a student on its own rollouts guided by teacher feedback and is becoming increasingly important for large language model (LLM) post-training. Like reinforcement learning (RL), however, OPD faces an on-policy systems bottleneck, as rollouts can dominate training time for reasoning workloads. Asynchronous training pipelines can alleviate this bottleneck by decoupling rollout generation from learner updates, but doing so introduces stale-policy data. While prior work has studied stale data in asynchronous RL, its effects in OPD remain underexplored. We present the first systematic study of staleness in asynchronous OPD, focusing on a practical setting where teacher feedback is implemented through local KL losses and full-vocabulary teacher logits are too expensive to store or transfer, necessitating finite teacher-score caches. We first show that KL direction changes the stale-data problem: teacher-weighted forward KL is more robust to stale rollouts, whereas student-weighted reverse KL is vulnerable. Second, for this vulnerable reverse-KL case, we study whether methods designed to stabilize asynchronous RL can mitigate OPD staleness. In our experiments, they do not improve over a simpler OPD-specific surrogate: recomputing the reverse-KL signal under the current student at learner time. Third, we analyze how finite teacher-score caches create a bias-variance tradeoff for sparse and sampled reverse-KL OPD estimators. This motivates multi-sample Monte Carlo (MC), which preserves MC correctability while reducing one-sample variance. Finally, we present and open-source AsyncOPD, a fully asynchronous OPD training pipeline built from these estimator choices. Experiments show that AsyncOPD improves training throughput by $1.6\times$ to $3.8\times$ over strict synchronous training while reaching comparable accuracy.

On-policy distillation (OPD) trains a student policy using teacher signals computed on trajectories sampled by the student itself. Recent work shows that sampled-token OPD can be fragile on long-horizon reasoning tasks and that local teacher-support matching is a simple and effective repair. This paper introduces blockwise policy-drift gating, a lightweight student-only old-current drift controller for OPD under rollout reuse. The method computes log-probability shifts between the behavior student and the current student on the sampled token path, aggregates these shifts over fixed blocks or spans, and uses the resulting detached, mean-normalized gates to reweight OPD position losses. It does not change teacher targets, teacher top-K supports, or the rollout policy. In a six-variant Qwen3 math reasoning benchmark with a uniform 200-step training budget for all trained variants, we use pass@8 as the primary problem-level solve-rate metric. Fixed 64-token block gating improves sampled-token OPD mean pass@8 from 0.4978 to 0.5160 across AIME24, AIME25, MATH500, and AMC23. On Teacher-TopK/LSM, Block64 gives the best four-benchmark mean pass@8 among trained students. The results identify local old-current policy drift as a practical control signal for reused OPD rollouts and motivate block-level gating as a simple default for improving solve-rate robustness.

Distilling conditional diffusion models aims to transfer the behavior of a large teacher to a smaller student while preserving alignment across conditioning inputs. Unlike recognition tasks, knowledge distillation in conditional diffusion often struggles to transfer knowledge beyond the training distribution, since the predicted noise strongly depends on the conditioning signal. As a result, effective distillation requires exploring a large conditioning space. In practical settings, this creates a major bottleneck. Paired image-condition data may be limited, and generating synthetic images for every available condition is often computationally infeasible, while the pool of conditions, such as text prompts, can be extremely large. Recent work addresses this issue by switching conditions during training, exposing the student to a broader conditioning space without changing the distillation objective. Yet this raises a complementary question: once a large conditioning corpus is available, how should the training effort be allocated? In this work, we introduce ARIA, a framework that adaptively allocates training effort across coarse regions of the conditioning space. By maintaining online estimates of teacher-student discrepancy at the region level, ARIA focuses updates where misalignment persists while preserving the original distillation objective. Empirically, ARIA improves over RC across most architectures and settings, with the clearest gains observed in unseen and underrepresented regimes. We also provide a theoretical analysis showing that the proposed tracking mechanism follows the evolving discrepancy during training under bounded variance and drift assumptions.

Prompt distillation compresses large vision-language models (VLMs) such as CLIP into lightweight student models by matching teacher predictions on unlabeled domain images. PromptKD (CVPR 2024) established this paradigm with a single PromptSRC-finetuned ViT-L/14 teacher and a ViT-B/16 student. We propose TheProfessor, a multi-teacher extension that distills from a fixed two-teacher ensemble: a domain-finetuned PromptSRC ViT-L/14 teacher and a zero-shot EVA-CLIP-L/14 teacher whose logits are pre-computed per dataset. We evaluate single-teacher PromptKD, equal-probability ensembling, and confidence-weighted ensembling on four base-to-novel datasets: Caltech-101, DTD, UCF101, and EuroSAT. In a 12-run single-seed sweep, confidence-weighted ensembling improves average HM from 87.52 to 89.28 (+1.77 points), while equal averaging improves average HM to 88.88 (+1.37 points). Gains are dataset dependent: they are negligible on Caltech-101 (+0.16 HM for confidence weighting), modest on UCF101 (+0.62), and largest on domain-shifted EuroSAT (+5.78). These results update our earlier Caltech-only analysis and show that multi-teacher prompt distillation is most useful when the second teacher contributes complementary supervision under domain shift.

On-policy distillation transfers reasoning ability through dense token-level supervision, yet the nature of the transferable signal remains unclear. We discover that reasoning chains contain two types of knowledge that require different discovery mechanisms: decisions (where to branch), which surface through student uncertainty, and evidence (intermediate steps that justify decisions), which hides in positions where the student is confident yet wrong. Current methods capture only decisions; the substantive knowledge in evidence tokens remains untransferred. We propose DEAR(Decision-Evidence Aware Reasoning Distillation), which first identifies decisions via student entropy, then discovers their supporting evidence through hidden-state cosine similarity to decision anchors, boosted by teacher-student divergence to prioritize the largest knowledge gaps. Across three student-teacher configurations on math and code benchmarks, DEAR consistently outperforms standard OPD, with up to +2.5pp on competition math and +5.7pp on code generation.

On-policy distillation (OPD) trains an LLM on states induced by the current or recent student policy: the student generates complete or partial rollouts, a teacher or self-teacher scores the resulting tokens under their generated contexts, and dense log-probability, logit, or distributional signals are converted into post-training updates. This survey studies OPD as a feedback-to-update problem rather than a single loss family. We develop a formula-driven taxonomy from two routes -- direct distributional losses and policy-gradient-style log-ratio updates -- and use it to organize core methods, verifier- or outcome-guided hybrids, industrial reports, framework implementations, failure modes, and stabilization recipes under explicit evidence boundaries. The taxonomy shows that OPD effectiveness depends not only on KL direction or teacher access, but also on state compatibility, support construction, temporal credit, vocabulary-level probability routing, gates and weights, and regularization. We further separate two mechanisms often conflated in sampled-token OPD stability discussions. Temporal credit asks how teacher-student log-ratio returns should weight sampled actions across a rollout; vocabulary routing asks where probability mass should move when negative feedback suppresses a sampled token. This distinction yields bias boundaries for immediate, return-to-go, discounted, and baseline-corrected estimators, motivates GAE-OPD as a value-based hypothesis for log-ratio returns, and motivates Counterfactual Routed OPD (CR-OPD) for routing probability mass toward teacher-supported, student-reachable alternatives. We close by mapping actionability diagnostics, failure mechanisms, case studies, open problems, and a reporting checklist onto the same feedback-to-update variables.

Lip-syncing deepfakes are among the most challenging forms of manipulated media because their artifacts are localized almost exclusively to the mouth region and evolve dynamically over time. Detecting such deepfakes requires precise temporal and spatial modeling of lip motion. In this paper, we propose LoCC, a novel detection framework that performs fine-grained detection and localization of lip-syncing deepfakes at both segment and frame levels. Unlike prior approaches that analyze videos holistically, our method evaluates whether each frame aligns with a counterfactual estimate generated from its temporal neighbors. Real videos exhibit strong and stable consistency, whereas lip-sync deepfakes introduce localized inconsistencies. Following a teacher-student learning paradigm, our model effectively captures these frame-level discrepancies and achieves superior performance over state-of-the-art methods on multiple benchmark lip-syncing deepfake datasets, including LAV-DF, AVDF1M, FakeAVCeleb, and KODF, and generalizes well across compression levels and datasets.

On-Policy Distillation (OPD) improves the learning efficiency of standard reinforcement learning through dense, token-level supervision from teachers. In the standard KL objective of OPD, token-level losses are uniformly averaged, implying equal weights for all tokens. However, we discover that not all tokens are created equal: as student rollouts grow longer, they deviate further from the teacher's distribution, leading to degraded supervision quality at later positions. As a result, OPD using only the first 30% of tokens can perform comparably to using all tokens, whereas OPD using only the last 30% of tokens barely learns anything. In this work, we provide a principled understanding of this issue through the lens of constrained optimization. Based on these insights, we derive Importance-Weighted On-Policy Distillation (IW-OPD), in which the weight assigned to each token depends on the accumulated discrepancy between the student's and teacher's distributions, naturally upweighting earlier tokens and downweighting later ones with larger deviations. We show that IW-OPD converges significantly faster than OPD, with better learning efficiency, and achieves better final performance than standard OPD in both same-size and cross-scale settings, improving performance up to 6.9 points on AIME-2025.

Subliminal learning lets a student inherit a teacher's hidden trait from distillation data that never names it. We ask when such transfer can be audited before training. The answer is not model identity or scale alone, but channel location: the carrier through which the trait reaches the student. We find three regimes. In a controlled initialization-dependent body channel, a pre-training screen works. Coverage, the cosine between the student's initial distillation update and the teacher's fine-tuning displacement, predicts held-out transfer (Spearman $ρ\approx 0.95$; AUROC 0.997). In pretrained language models, masked single-token traits instead ride convergent vocabulary geometry. This channel is initialization-independent, so initialization-alignment screens, including coverage, are not mechanistic; the useful handles are post-hoc detection and targeted mitigation. Even when a single-token named entity is removed from the loss, the student's held-out probability for that entity rises to 0.40 on average ($\sim 2500\times$), and a related semantic class transfers. In an untied-head model, orthogonalizing the trait's output row against entangled neighbours collapses leakage, while equal-size random-subspace edits do not. Thus removing a target string from distillation labels does not remove the corresponding preference: neighbouring tokens can carry it. Finally, conditional behaviours can route through the network body. For sycophancy, with agreement and correction markers masked from the loss, transfer reaches about 0.63 of the teacher's effect, localizes to body computation, and evades four audits across two model families. We scope this as masked transfer of a condition-present policy. Channel location is necessary for deciding which audits can be sound. It is not a deployment-ready screen: an audit used outside its carrier regime can give false assurance.

On-policy distillation (OPD) improves reasoning models by applying dense teacher supervision on student-sampled trajectories. However, scaling OPD to long-horizon mathematical reasoning exposes a reliability and efficiency problem: standard OPD assigns every sampled candidate the same long rollout budget, even though some trajectories may quickly become weakly aligned with the teacher and provide less useful supervision. Prior analyses suggest that successful OPD depends on local teacher-student compatibility, which can be measured by top-k overlap on student-visited prefixes. When this overlap is low, continuing to generate or train on long suffixes may waste computation and introduce noisy learning signal. To address this, we introduce Prefix-Guided On-Policy Distillation (PG-OPD), a simple rollout-allocation framework that uses fixed-length prefixes to estimate trajectory value before expensive long-horizon generation. PG-OPD first decodes every sampled candidate to the same prefix length, computes teacher-student top-k overlap within an early probe window of that prefix, and selectively continues high-overlap candidates to a fixed long length. Low-overlap candidates stop at the fixed prefix, avoiding unnecessary suffix generation. Across diverse teacher-student combinations on AMC, AIME, and HMMT benchmarks, PG-OPD improves average accuracy by up to 4.80 points while reducing training time by up to 2.46x. These results suggest that prefix-level compatibility provides a practical signal for directing OPD computation toward trajectories that remain learnable from the teacher.

Chain-of-Thought (CoT) distillation transfers multi-step reasoning from large reasoning models to smaller students, but verbose teacher traces inflate both training and inference cost. Existing CoT compression methods fall into two families, selective pruning and generative rewriting, yet prior studies have left key factors entangled: granularity is confounded with importance criteria in pruning, restructuring level is rarely isolated in rewriting, and compression budgets are not systematically evaluated across domains or regimes. We recast CoT compression along three dimensions: importance criterion, restructuring level, and compression budget. Sweeping these across two model families, Math and General domains, and Long-/Short-CoT regimes, we find that (i) importance criterion utility is strictly governed by granularity: step-level criteria converge on a shared reasoning backbone, while token-level pruning requires symbol-aware signals to preserve the logical core; (ii) restructuring level inverts across domains: Math degrades monotonically with structural disruption, while aggressive rewriting acts as a denoiser on General tasks; (iii) training-time compression does not necessarily translate to inference-time savings: Long-CoT students retain verbose habits despite concise supervision, making the training ratio an optimistic lower bound on deployment cost. These findings yield condition-aware guidelines for matching compression to deployment context.

Model distillation -- training on outputs from stronger third-party models -- is widely used to boost performance, but raises concerns about unfair advantages and policy violations. This motivates a fundamental question: can we detect whether a model was distilled from another? We show that, while identifying a teacher model from a student in isolation is highly challenging, it becomes tractable in a reference-based setting: given a model and an earlier-generation checkpoint from the same lineage, we can identify the teacher model used to train the later checkpoint. We introduce a distillation detection method based on reference-based membership inference. By comparing how strongly a student model preferentially aligns with outputs from different candidate teachers relative to a reference checkpoint, our method identifies the most likely teacher and detects evidence of distillation. To handle unknown distillation pipelines such as hidden prompts, we infer proxy prompt templates directly from model outputs. We additionally identify a distinctive glyph-level signal specific to o1/o3 models. Evaluating distillation detection is challenging because modern model lineages are already heavily entangled. To address this, we develop a hybrid evaluation spanning both controlled distillation experiments and real-world models. Across both settings, our approach recovers the true teacher with near-perfect accuracy in single-teacher distillation scenarios, even when the underlying distillation pipeline is largely unknown. We further introduce statistical tests for both teacher attribution and distillation detection, and extend our framework to open-world settings where no teacher is guaranteed to be present among the candidates. Applying our method to contemporary models yields new evidence regarding potential distillation relationships involving QwQ, DeepSeek-R1, and GPT-OSS.

Knowledge distillation transfers a teacher's competence to a small student but is brittle in the small-student regime: forcing the student to imitate logits from a much larger teacher hurts generalization on benchmark families beyond the training corpus. Reinforcement learning (RL) avoids logit imitation by training on the student's own rollouts. However, on questions where every rollout fails - yielding zero advantage and being silently discarded - injecting a stronger teacher's response into the policy gradient breaks the on-policy assumption and induces drift. We introduce Zone of Proximal Policy Optimization (ZPPO), inspired by Vygotsky's zone of proximal development. ZPPO keeps the teacher inside the prompt rather than the policy gradient. On hard questions, where the student's mean rollout accuracy is below half, ZPPO constructs two reformulated prompts. A Binary Candidate-included Question (BCQ) pairs one correct teacher response with one incorrect student response as anonymized candidates the student uses as references. A Negative Candidate-included Question (NCQ) aggregates the student's wrong rollouts into a single prompt to surface their shared failure modes. A prompt replay buffer recirculates each hard question until it either graduates - the student's mean rollout accuracy on it reaches half or more - or is FIFO-evicted under finite capacity, amplifying BCQ and NCQ inside the student's current zone of proximal development. We post-train Qwen3.5 students at four scales (0.8B-9B) as vision-language models with a 27B teacher and evaluate them on a 31-benchmark suite (16 VLM, 10 LLM, 5 Video); ZPPO outperforms off/on-policy distillation and GRPO, with the largest gains at the smallest scale.

Standard on-policy distillation (OPD) for large language models estimates the reverse-KL objective using student-sampled tokens, yielding an unbiased single-sample Monte Carlo estimator that avoids vocabulary-wide computation. However, we show that this estimator suffers from severe training pathologies in practice: sample inefficiency, unstable generation dynamics, and a substantial performance gap compared to exact full-vocabulary OPD. Reward-level diagnosis traces these pathologies to the log-ratio reward, which is unbounded by construction, producing extremely high-variance gradients concentrated at early positions and persisting throughout training; standard post-hoc scaling fail as they operate only after this distortion occurs. To solve this problem, we propose PowerOPD: a family of natively bounded, sign-consistent rewards from the Box-Cox power transformation, parameterized by alpha > 0, of which the log-ratio is the degenerate alpha -> 0 limit. Across six mathematical reasoning benchmarks and four Qwen3 teacher-student pairs, PowerOPD achieves benchmark-averaged Avg@8/Pass@8 gains of up to +6.37/+5.71 over vanilla OPD, +3.01/+3.54 over post-hoc stabilization, and +2.59/+8.90 over full-vocabulary OPD, while reducing wall-clock time by 59.2% and peak GPU memory by 23.1%. Larger alpha generally improves accuracy, consistently shortens responses, and keeps gradient norms more than 3,000x smaller than vanilla OPD.

Hybrid linear attention models offer an appealing path to faster long-context inference: they reduce the quadratic cost and KV-cache burden of full softmax attention while retaining much of the quality of Transformer models. A practical way to obtain such models is to convert a pretrained Transformer instead of pretraining a new architecture from scratch, but this conversion is still brittle. Simply copying the teacher attention projections into a Gated DeltaNet (GDN) student does not specify the new recurrent decay, write, and output-gating dynamics. As a result, the converted model often starts in a poor dynamical regime and must spend many distillation tokens repairing initialization rather than learning the remaining teacher behavior. We propose Taylor-Calibrate, a lightweight initialization method for hybrid GDN students. The method uses Taylor-guided teacher attention statistics to set the value projection, memory timescale, write gates, and output gate, then applies a short per-layer alignment step to match each converted layer to the teacher output. Across four teacher settings and three retained-layer policies, Taylor-Calibrate gives substantially stronger zero-shot students, with up to an 88x improvement in a representative ablation, and reaches matched recovery targets with 4.9x--9.2x fewer training tokens than naive conversion.