Knowledge Distillation

Text-to-SQL is a fundamental task in natural language processing that enables users to interact with structured databases using natural language. While large language models (LLMs) have demonstrated remarkable performance on this task, their substantial computational requirements hinder deployment in resource-constrained settings. In this paper, we introduce SQuaD-SQL (Small-Qualified and Distilled for SQL), a novel approach that empowers small language models (SLMs) to approach the performance of LLMs on the Text-to-SQL task while significantly improving efficiency through knowledge distillation and synthetic data generation. Our method comprises three key components: (1) LLM-based synthetic data generation, where structured knowledge is extracted from LLMs via carefully designed prompting strategies; (2) parameter-efficient fine-tuning, enabling full model training on a single consumer-grade GPU; and (3) domain-adaptive fine-tuning, where domain-specific synthetic data further enhances performance in targeted domains. Experiments on the WikiSQL dataset demonstrate that SQuaD-SQL achieves an execution accuracy of 86.9% on the test set, approaching the performance of LLMs while offering faster inference and lower memory usage. These results suggest that, with proper training strategies, SLMs can serve as practical and efficient alternatives for Text-to-SQL applications in resource-limited environments.

Reliable confidence estimation is essential for deploying large language models (LLMs) in confidence-aware systems, where downstream decisions such as retrieval, tool use, and adaptive computation depend on accurately estimating answer reliability. Existing approaches, however, largely treat confidence as a property of completed responses, overlooking how confidence-related information evolves throughout the answering process. In this work, we investigate confidence from a temporal perspective by comparing pre-solution Feeling-of-Knowing (FOK) and post-solution Judgement-of-Learning (JOL) confidence estimates across frontier and open-source LLMs. We show that post-solution confidence is consistently better calibrated and more discriminative than pre-solution confidence, while linear probes trained on hidden representations recover substantially richer confidence-related information than models explicitly verbalise. Building on this observation, we introduce future confidence distillation, which trains predictors operating on pre-solution hidden representations using teacher confidence estimates produced by post-solution correctness probes. Despite requiring only pre-solution representations for inference, distilled predictors recover much of the calibration improvement achieved by post-solution confidence, remain highly sample efficient, and transfer across datasets within the same domain. Together, our findings demonstrate that confidence-related information evolves throughout the answering process and can be anticipated before answer generation is complete, enabling significantly more reliable yet low-cost confidence estimation.

Deep learning has achieved remarkable success in various domains including time series analysis, computer vision and natural language processing. However, high computational and memory demands of state-of-the-art architectures pose challenges for deployment in resource-limited environments. Knowledge Distillation (KD) addresses this by transferring knowledge from a large teacher model to a smaller, more efficient student model while maintaining competitive performance. In this work, we investigate the effectiveness of KD for Time Series Classification (TSC) across three architectures: the classical Fully Convolutional Network (FCN), the convolutional Inception model and the transformer-based ConvTran model. We evaluate our approach on UCR Archive, the largest benchmark repository of time series datasets, by modifying architectural components such as convolutional filters, Inception modules and attention heads across the three architectures. Our results consistently show that KD most effectively benefits student models of intermediate complexity across all three architectures, with the distilled FCN student reducing parameters by a factor of 38, the distilled Inception student achieving nearly the same performance as the teacher with 42% fewer parameters and the distilled ConvTran student with 2 attention heads showing the most significant improvement through distillation. To encourage further research and reproducibility, we provide our implementation at https://github.com/MSD-IRIMAS/KD-4-TSC.

Coreset selection aims to identify a small and highly representative subset of a massive dataset for efficient model training. The problem remains challenging even in the few-shot knowledge distillation (KD) setup, where a full-scale pre-trained teacher informs the student network. Typical sample selection strategies often struggle to surpass the random selection baseline. In this paper, we showcase few-medoids, an embarrassingly simple coreset selection strategy that chooses the samples closest to the centroid (average image) of each class. We present extensive KD experiments on four datasets, covering a wide range of image classification problems, and three teacher-student model pairs, comprising both convolutional and transformer networks. Although the proposed method is embarrassingly simple, our empirical results indicate that few-medoids is able to consistently surpass the random selection baseline, as well as the other coreset selection strategies. We therefore consider that few-medoids can be used as a drop-in replacement for commonly-used baselines (e.g. herding or k-center Greedy), in future research on coreset selection. To reproduce the reported results, we publicly release our code at https://github.com/CemilAndreiDilmac/Few-Shot-KD-Coreset.

Computer-aided design (CAD) for industrial components requires long-horizon procedural modeling, robust feature dependencies, editable parametric geometry, and production-grade B-Rep execution. Existing text-to-CAD methods have made promising progress in generating CAD programs from natural-language descriptions, but they still struggle when user prompts are ambiguous, underspecified, or only describe high-level design intent. They also rarely exploit expert procedural knowledge naturally available in industrial workflows, such as CATIA operation recordings, macro logs, drawing notes, and engineering descriptions. We present ArtisanCAD, a skill-guided industrial CAD agent with expert-grounded knowledge distillation. The core of ArtisanCAD is CAD intermediate representation (CAD-IR), an executable procedural representation that encodes parameters, ordered operations, MCP tool bindings, dependencies, generated entities, and verification rules. CAD-IR plays two key roles: it first serves as the carrier for distilling expert CAD procedures into reusable parameterized skills; then it provides a procedural scaffold that turns vague or intermediate-level prompts into complete executable CAD operations. ArtisanCAD retrieves expert-derived skills, instantiates and revises CAD-IR, executes the resulting procedure through a dedicated CATIA-MCP backend, and uses multi-view visual feedback for iterative refinement, and finally generates production-ready B-Rep models. On the Text2CAD benchmark, CAD-IR improves generation from intermediate prompts by reducing mean Chamfer Distance from $14.83$ to $9.88$, showing its ability to bridge ambiguous textual intent and executable CAD construction. On four complex automotive components, CAD-IR enables expert CATIA recordings to be distilled into reusable skills, allowing ArtisanCAD to generate editable CATIA-native B-Rep models for new variant requests.

Computational pathology foundation models (PFMs) have advanced whole-slide image analysis. However, their size and inference cost hinder local deployment in pathology departments. We propose MuCoDi, a pretraining framework that distills frozen tile embeddings from multiple PFMs into compact edge-oriented encoders. Instead of regressing individual teacher features, MuCoDi trains lightweight MobileOne and RepViT students with a contrastive distillation objective adapted from MoCo v3, where cached Virchow2, UNI2, and H-Optimus-1 embeddings replace momentum-encoder keys. We pretrain students on 14.3M TCGA tiles from only 11.8K WSIs and evaluate frozen encoders on 23 clinically curated downstream classification tasks. RepViT-based MuCoEdge students retain near-teacher performance while reducing model size by orders of magnitude: MuCoEdge-R2.3 and MuCoEdge-R1.5 reach 71.0% external AUROC, within 0.8 percentage points of the best teacher (Virchow2, 71.8%), while MuCoEdge-R2.3 obtains the best external F1 and the second-best AUPRC (51.8% and 53.3%). MuCoEdge-R1.0 reaches 70.9% AUROC with only 6.4M parameters and 1.12 GFLOPs. On a Raspberry Pi 5, sub-million-parameter MobileOne students achieve up to 605-fold single-tile speedup over Virchow2 while retaining 66.5% to 66.9% external AUROC, demonstrating that PFM-quality pathology representations can be moved toward practical edge deployment. Code is available at https://anonymous.4open.science/r/mucodi-6243.

Tabular Foundation Models (TFMs) have demonstrated strong empirical performance as black-box inference engines through in-context learning. However, their use in transfer learning is limited by two obstacles: strict context-size constraints and sensitivity to distribution shifts between source and target tasks. Directly pooling heterogeneous source data can therefore lead to negative transfer. To address these challenges, we propose Context-Constrained Transfer Learning via ANchoring and DIstillation (TL-ANDI), a posterior-aware distillation framework for TFMs. TL-ANDI constructs a compact source context by solving a budget-constrained optimal transport problem whose cost jointly measures target covariate coverage and posterior compatibility. The selected anchor samples are then equipped with locally distilled labels and combined with a residual calibration step using target data.

Distilling a fine-tuned teacher into a LoRA-adapted student is a standard recipe for parameter-efficient compression, but output-level KD does not explicitly control which rank-$r$ weight subspace the adapter occupies. We propose \textbf{SAD-LoRA} (\textbf{S}pectral \textbf{A}lignment \textbf{D}istillation), which selects this subspace from the data-weighted student-space reference update $\DWT\Sigx^{1/2}$ and maintains it during training via a differentiable principal-angle loss on $\colspan(B)$. We show that the data-weighted distillation error decomposes exactly into subspace misalignment, within-subspace coefficient mismatch, and irreducible rank residual; standard KD can affect the first term only indirectly through output gradients. On controlled synthetic problems with a flat teacher spectrum, SAD-LoRA reduces the subspace-misalignment term from $51\%$ to nearly zero and lifts final subspace alignment from $0.49$ to $1.00$. On RoBERTa-large to RoBERTa-base distillation across six GLUE tasks, SAD-LoRA improves rank efficiency: at $r{=}4$, it matches or beats the strongest included spectral baseline on five of six tasks, and at $r{=}8$ it gives the best result on SST-2 and CoLA. Ablations identify subspace alignment as the load-bearing component, while coefficient matching is auxiliary.

Plasma disruption is a critical threat to tokamak safety. Existing data-driven predictors mainly rely on time-series diagnostic signals, while visible images provide complementary spatial cues including plasma deformation, local brightening, and radiation-structure evolution. Although the image modality improves the model's discriminative capability, it also substantially increases the computational cost during inference. To address this issue, we propose a hierarchical multi-to-single-modal knowledge distillation framework for disruption prediction on a synchronized EAST multimodal dataset. During training, visible images and time-series signals are used to train a multimodal teacher, which learns disruption precursor representations through Transformer-based encoders and a prototype-guided spatiotemporal hypergraph module. During inference, only the time-series student is retained, with multimodal knowledge transferred through graph-structure-level, representation-level, and decision-level distillation. On the 640-discharge EAST dataset, the results demonstrate that the proposed framework can preserve the discriminative advantages of multimodal learning while substantially reducing inference cost, and providing an effective route for efficient disruption prediction in EAST. The source code of this paper will be released on https://github.com/Event-AHU/OpenFusion.

End-to-end Vision-Language Models (VLMs) show immense potential in autonomous driving. However, standard Supervised Fine-Tuning (SFT) often suffers from reasoning hallucinations and conservative biases. While traditional tool-augmented frameworks and Chain-of-Thought (CoT) approaches mitigate these issues, they incur exorbitant token consumption and unacceptable latency, rendering real-time deployment impractical. To resolve this reliability-efficiency trade-off, we propose CritiqueDriveVLM, a novel unified three-stage framework internalizing reasoning directly into the VLM. First, we introduce Critique-Driven Multi-Turn Reinforcement Learning (RL) guided by a multi-dimensional verifier. By providing granular scalar feedback and a multi-turn penalty, we force the policy to internalize logical deduction, cultivating a robust System-2 Teacher that achieves high accuracy without fragile external tools. Subsequently, we propose Latent Thought Distillation to overcome the latency bottleneck. By aligning the Student's latent representations with the Teacher's fully converged reasoning states, we compress deep logical capabilities into a fast, CoT-free System-1 Student. Extensive experiments on the widely-used DriveLMM-01 benchmark demonstrate remarkable improvements. Compared to the base model, our tool-free Teacher significantly boosts Multiple Choice Quality (MCQ) from 55.54% to a state-of-the-art 76.54%. Crucially, our distilled Student preserves competitive reasoning depth while drastically minimizing generation length to an average of merely 28 tokens. This slashes inference latency by 88% (from 3482 ms to 416 ms), paving a highly robust pathway for low-latency autonomous driving.Our source code is available at https://github.com/MICLAB-BUPT/CritiqueDriveVLM.

Falls among the elderly represent a significant public health concern due to their prevalence, consequences, and societal burden. While deep learning has improved fall detection, accurately identifying impact moments (when an individual hits the ground) remains challenging. Additionally, current algorithms often rely on complex models with high computational demands, limiting real-time deployment feasibility. In this work, we propose DistillH-Mamba, a novel architecture for impact fall detection that addresses these challenges through three key innovations: First, we introduce a hypergraph-based approach that captures higher-order relationships between multiple joints simultaneously, enabling more accurate modeling of complex interactions during impact falls. Second, we integrate the Mamba architecture with hypergraphs for impact detection, significantly accelerating processing speed while efficiently capturing both long-term dependencies and sudden skeletal motion changes. Third, we employ relational knowledge distillation that preserves crucial spatial-temporal relationships while reducing computational demands for real-time impact fall detection. Evaluated on the 3D Skeletons UP-Fall and UMAFall datasets, our DistillH-Mamba model achieves 97.38% accuracy in detecting impact within fall events and 73.8% reduction in inference time compared to its teacher model, outperforming state-of-the-art methods in both precision and efficiency.

Vision-language-action (VLA) models open a new path toward intuitive robot control by directly linking perception, language, and action in a single end-to-end framework. Yet for UAVs, practical adoption remains difficult because existing solutions are either computationally heavy or insufficiently capable in complex environments. In this work, we propose a practical expert-distillation pipeline (Exp2VLA) for language-conditioned drone navigation. The core idea is to distill expert behavior, obtained from reinforcement learning, teleoperation, or other controllers, into training data that can be used to fine-tune compact VLA models. This allows existing control strategies to be transferred into a unified language-guided navigation model, reducing manual system integration and lowering the barrier for deploying new robot behaviors. Experiments in both sim-to-sim and simulation-in-the-loop settings across multi-object scenes show that the fine-tuned models can handle varied semantic commands and generalize to unseen target compositions. The proposed framework demonstrates how expert-policy distillation can help mechatronic systems move from specialized control modules toward more flexible and reusable robot intelligence.

Most existing knowledge distillation methods focus on homogeneous models (e.g., CNN-to-CNN), thereby overlooking the flexibility and potential of knowledge transfer across heterogeneous models. Due to intrinsic inductive bias discrepancies between heterogeneous models that cause spatial distribution inconsistencies, prior heterogeneous distillation methods often weaken or discard spatial information in heterogeneous representations. However, the spatial information in representations often encodes transferable global structural semantics as well as architecture-specific local details, and therefore should not be directly ignored. To better leverage the spatial information encoded in heterogeneous representations, we propose a Spatial-Frequency Joint-Aware Heterogeneous Knowledge Distillation framework (SFKD). By leveraging the complementary properties of wavelet transform spatial locality and Fourier representations in characterizing global energy distributions, we first apply multi-level discrete wavelet transform to explicitly decouple spatial information. The resulting wavelet sub-bands are further refined by a dual-stream dual-stage refinement module, and finally combined with a Gaussian-filtered frequency loss to selectively capture informative global information. Extensive experiments on multiple benchmark datasets under both homogeneous and heterogeneous models demonstrate the superiority of our method.

Large 3D foundation models such as MASt3R achieve state-of-the-art stereo reconstruction but are computationally demanding for deployment under strict hardware constraints -- a critical limitation in domains such as planetary exploration, where onboard computing is severely restricted. We study how far such models can be compressed through knowledge distillation, using lunar stereo reconstruction as a challenging and practically relevant case study. Starting from a 688M-parameter MASt3R teacher fine-tuned on lunar imagery, we distill its dense geometric predictions into a family of lightweight students spanning different encoder types (CNN vs ViT), decoder widths and depths, and training strategies. To bridge the dimensional mismatch between teacher and student, we propose a structured SVD-based initialization that projects the teacher's decoder weights into the student's smaller latent space, yielding a warm start that significantly improves convergence and final performance. Based on our results on lunar data, we can obtain a distilled student that retains most of teacher's reconstruction accuracy while reducing the model size up to 7 times, and even outperforms a baseline trained directly with sparse ground-truth annotations. Beyond compression, our study highlights both principles and practical insights for distilling geometric foundation models: a convolutional encoder underperforms transformer-based alternatives (though pretraining availability remains a confounding factor), preserving encoder capacity is more critical than maintaining a large decoder, feature-level distillation consistently outperforms output-only supervision, and SVD-based initialization improves optimisation stability. These findings provide practical guidelines for deploying 3D reconstruction models in resource-constrained environments.

LiDAR-based 3D object detection is essential for autonomous driving systems. However, traditional Ego-only Perception (Eo-Perception) suffers from limited perspective and occlusions in a complex outdoor environment, leading to performance bottlenecks. Recently, research on multi-agent Collaborative Perception (Co-Perception) has demonstrated excellent performance, but high communication costs and accumulated pose error hinder its application. To address this, we explore a novel C2E (Co-Perception to Eo-Perception) paradigm through the Multi-to-Single (M2S) agent contrastive knowledge distillation framework. Our M2S framework first designs Multi-Level Feature Enhancement module to provide more stable features, and introduces Auxiliary Point Cloud Reconstruction and Multi-Teacher Contrastive Distillation mechanisms to mitigate domain gaps in point cloud and feature distributions within the C2E paradigm. Benefiting from this, our M2S can retain the excellent performance of collaborative perception while effectively avoiding the drawbacks, such as communication delays and positioning errors. Extensive experiments on the V2XSet, V2V4Real and DAIR-V2X datasets show the effectiveness and generalizability of our M2S framework when combined with the state-of-the-art CoSDH model and other excellent 3D detectors. Our M2S framework can deliver up to a 8.64% improvement in 3D mAP performance without introducing any communication costs.

Language models deployed in high-stakes roles can potentially favor certain entities, brands, or viewpoints, steering user decisions at scale. Such preferential biases can be introduced by any actor in the model's supply chain and are most dangerous when the model reveals its preference only on the relevant topic while behaving identically to its unmodified base on all other inputs. Recent work has shown that these biases can transfer through context distillation on semantically unrelated data, with the signal residing entirely in the soft logit distribution and remaining invisible to text-based inspection. However, the defender faces a fundamental asymmetry: without knowing the bias topic, no detection method can reliably surface a stealth preferential bias, regardless of whether it examines generated text, internal representations, or model weights. Here we introduce Distill to Detect (D2D), a method that surfaces hidden biases by distilling the distributional shift between a suspected model and its base into a cartridge (a KV-cache prefix adapter), concentrating the dominant divergence and amplifying the bias signal into generated text. We show that D2D successfully amplifies the hidden biases of stealth models to the extent that they can be reliably detected across multiple bias types. We also propose a theoretical framework that explains the efficacy of D2D through the lens of Fisher-weighted projection of the logit distribution shift, supported by empirical observations. By turning the capacity bottleneck of prefix-tuning adapters into a detection tool, D2D provides a practical building block for auditing hidden behaviors in deployed language models.

While knowledge distillation (KD) is widely adopted for training lightweight models by leveraging supervision from larger teacher models, relying solely on output token distributions has proven insufficient for compressing Multimodal Large Language Models (MLLMs). Since output tokens are a byproduct of the model attending to visual inputs, prior works have explored explicitly distilling attention to provide a direct supervisory signal. While promising, the precise utility of which attention signals to distill remains under-explored. In this work, we challenge the conventional reliance on prompt-to-vision attention by revealing that downstream performance correlates strongly with response-to-vision attention similarity to the teacher, but negligibly with that of prompt-conditioned attention. Furthermore, we observe that attention distributions exhibit significant variance across individual tokens, indicating that a uniform distillation objective is suboptimal. To this end, we introduce Token-level Response-visual Attention Guidance (TRAG), a distillation objective that 1) shifts the focus to response-to-vision signals and 2) employs token-specific objectives by adaptively weighting the Kullback-Leibler divergence based on attention entropy, effectively guiding the student to mirror the teacher's precise visual focus. Extensive experimental results on multiple benchmarks demonstrate that TRAG significantly outperforms prior distillation baselines.

Temporal Action Localization (TAL) typically relies on segment annotations or offline access to full videos, limiting scalability and online use. We introduce Point-Supervised Online TAL (POTAL), which localizes actions in streaming videos using only one temporal point per instance. To solve POTAL, we propose OnPoint, an offline-to-online multi-level distillation framework that transfers knowledge from a point-supervised offline teacher to an online student via (i) pseudo-segment instance distillation, (ii) class-activation sequence distillation, and (iii) anticipatory window-level distillation. We further improve robustness by incorporating the original point labels into student training and by refining anchor decoding with actionness-guided attention calibration. Experiments on five datasets show OnPoint consistently outperforms strong baselines, establishing a solid foundation for POTAL.

Federated learning is bandwidth-bound on two orthogonal axes: model size, which limits how often parameter-averaging methods can afford to merge, and class count, which makes per-probe soft-label distillation prohibitive at large vocabularies. Both ceilings tighten as modern systems scale. We collapse the class-count axis to $\lceil \log_2 C \rceil$ bits per probe by transmitting only each peer's $\arg\max$ class index, where $C$ is the number of output classes. The resulting protocol, TallyTrain, is not merely compressed: under non-IID training it can be preferable to soft-label distillation, because under-trained peers are confidently wrong and majority voting filters this noise where soft-label averaging amplifies it. Across standard benchmarks, TallyTrain matches or beats soft-label distillation at up to three orders of magnitude less communication. We also relax the model-size axis: we compose the cheap hard-label consensus with sparse parameter merges to obtain a bandwidth-bridge variant, which Pareto-dominates every tested operating point of the standard FedAvg, FedProx and FedDF baselines.

Deploying 3D point cloud analysis in privacy-sensitive, resource-constrained settings faces two barriers: data cannot be centralized, and models must run on limited edge hardware. We present a multi-seed benchmark jointly evaluating federated learning (FL) and knowledge distillation (KD) for 3D point cloud classification. It spans 13 FL algorithms and 10 KD objectives (a 130-pair cross-product) across 504 training runs, evaluated on ModelNet40 and a clinical craniosynostosis dataset. We report three findings. First, under extreme non-IID label skew, standalone FL degrades sharply: on ModelNet40, the strongest method reaches 76.32% against a 92.26% centralized reference; on clinical data, the best reaches 75.83% against 100%. Second, distillation successfully compresses the teacher into a student 74.51% smaller and roughly twice as fast at inference, often matching or surpassing the teacher. Third, the combined pipeline exposes an evaluation pitfall: when distillation keeps a hard-label cross-entropy term on a labeled proxy split, a collapsed federated teacher (8.50%) paired with Logit-MSE still yields a 92.94% student. This 84.4-point gap reflects the proxy labels rather than the federated model, reusing the very labels whose privacy motivated federation. Objectives without hard labels instead track teacher quality ($r \approx 0.99$) and collapse when the teacher does. We therefore recommend evaluating FL-KD pipelines with label-free distillation so reported accuracy reflects the federated teacher, not the proxy.

Certified training aims to produce models whose predictions can be formally verified against adversarial perturbations, typically by optimising upper bounds on the worst-case loss over an allowed perturbation set. For neural networks, certified training methods based purely on tight relaxation bounds produce networks that are amenable to certification, but sacrifice standard accuracy. Conversely, adversarial training often yields stronger empirical robustness and standard accuracy, but the resulting models are generally difficult to certify with neural network verifiers. Recently, the literature has shown that better standard-certified accuracy trade-offs can be achieved by combining adversarial training objectives with loose over-approximations based on Interval Bound Propagation (IBP), effectively interpolating between lower and upper bounds of the worst-case loss. Building on this, we introduce AD-CERT, a certified training objective that combines adversarial distillation with an IBP upper bound. We show that distilling adversarial information over the logit space from an empirically robust teacher provides an effective lower bound surrogate for certified training, with AD-CERT achieving state-of-the-art certified performance on several robustness benchmarks. Furthermore, in a unified setup, distilling adversarial information at the logit-level is shown to improve certified accuracy over a robust feature-space distillation objective by up to 5.40 percentage points.

The rapid expansion of sensor-based networks introduces major challenges in scalability, adaptability, and knowledge transfer, especially in open environments where new subsystems can dynamically join or leave. In this work, we propose a Holonic Active Distillation architecture within a Holonic Multi-Agent System (HMAS) to address these issues. Our approach integrates Clustered Stream-Based Active Distillation (CSBAD), a framework in which specialized student models collect local data, query pseudo-labels from teacher models, and cluster into groups of similar sensors. Results show that the holonic organization balances local specialization with global generalization, while efficiently adapting to sensor departures and re-integrations. We also analyzed trade-offs among incremental model updates, system reorganization, and scalability limits. Our findings highlight the advantages of holonic learning for multi-sensor systems while identifying key challenges related to model drift and long-term adaptation.

This paper investigates knowledge distillation from a large reasoning model (DeepSeek-R1) to a compact student model (Qwen2.5-7B). Using historical problems from the John O'Bryan Mathematics Competition at Northern Kentucky University (2011-2025), we build a Chain-of-Thought (CoT) training corpus through a dual-agent framework. The dataset is used to fine-tune the student model with Low-Rank Adaptation (LoRA) on Apple Silicon hardware using the MLX framework. The base Qwen2.5-7B model achieves 64.67% accuracy on competition problems, while the DeepSeek-R1 teacher achieves 91.40%. An initial 1,000-iteration training run revealed severe overfitting, with validation loss reaching a minimum at iteration 200 before rising steadily. Based on this finding, we ran five independent training runs each limited to 200 iterations with varied random seeds to assess result stability. Across these five runs, the fine-tuned student model achieves a mean accuracy of 69.43% (std dev 0.17%) on the competition dataset, a 4.76 percentage-point improvement over the base model, and generalizes to 73.1% (std dev 0.18%) on the MATH-500 benchmark. We further study how response length affects answer quality across six reasoning levels (R1-R6): accuracy declines consistently from 69.43% at R1 (mean 220 words) to 41.9% at R6 (mean 31.2 words), with the two-person speed section most sensitive to token reduction. These results demonstrate that CoT distillation improves compact student models and that response length is a critical factor in mathematical reasoning quality.

Visual environments are a demanding setting for quantum reinforcement learning (QRL): high-dimensional observations, unstable RL optimisation, and constrained variational quantum circuits (VQCs) are difficult to train jointly. This paper studies knowledge distillation (KD) as a staged hybridisation strategy for visual QRL. Instead of training a hybrid visual agent end-to-end from pixels, we first train a classical visual teacher, freeze its encoder as a feature interface, and distil the teacher's policy behaviour into compact downstream heads. These heads can be classical or VQC-based, enabling small quantum-compatible students to be evaluated under the same frozen representation as compact classical controls. We evaluate the pipeline on CartPole Pixels and Acrobot Pixels. The results show that staged KD enables shallow VQC heads to acquire non-trivial visual-control behaviour in settings where direct pixel-based training would be substantially more difficult. Angle-encoded VQC heads retain near-teacher performance, while amplitude-encoded heads push compactness to an extreme regime, at the cost of greater fragility, stronger budget sensitivity, and higher simulation time. Overall, staged KD reframes visual QRL as a compact-head learning problem, opening a practical route for training small quantum-compatible policies outside the standard end-to-end RL loop.

Distillation and Reinforcement Learning (RL) fine-tuning are the primary pillars of diffusion post-training. While traditionally studied in isolation, the interaction between these phases remains poorly understood, and in particular how fine-tuning impacts the generative quality of distilled models. We introduce Rewarded Moment Matching Distillation (RMMD), a novel framework that simultaneously distills diffusion models and maximizes a reward function. RMMD preserves the high-fidelity ``naturalness'' characteristic of advanced distillation (such as 8-step Moment Matching) by adapting the sampling loop for on-policy training and repurposing the distillation loss as a proxy for integral KL regularization. By evaluating the FID-Reward Pareto fronts on ImageNet, we demonstrate that RMMD achieves superior trade-offs compared to single-step baselines (DI++) and multi-step competitors (DRaFT, HyperNoise). Finally, we apply RMMD to GenCast, a state-of-the-art weather forecasting model, to distill it while optimizing the Continuous Ranked Probability Score (CRPS) metric. The resulting distilled model achieves a 7.5x speedup while outperforming the teacher model on 93% of target weather variables, and being better calibrated. This proves that RMMD scales to complex, high-dimensional scientific domains.

A key step toward artificial general intelligence is to train models that can perform multiple tasks. In this paper, we study how to build such models by first training separate RL experts for individual tasks and then consolidating them via distillation, as an alternative to directly training a single model on mixed tasks. We show that off-policy distillation degrades in multi-task settings due to the mode-covering nature of forward KL: aggregating data from multiple tasks introduces a large number of behavioral modes that can exceed the student's capacity, forcing it to average across behaviors and leading to degraded performance. In contrast, on-policy distillation is mode-seeking but requires strong initialization. Inspired by these observations, we propose a two-phase approach: off-policy distillation followed by on-policy refinement. Evaluation across conversational agents and text-based games confirms that this two-phase approach matches single-task RL expert performance for each individual task, whereas off-policy or on-policy distillation alone fails to match this performance.

Score distillation turns a pretrained 2D diffusion model into a 3D generator, but the per-step gradient is estimated from a single randomly chosen view: it is high-variance and blind to global shape consistency. Prior work addresses this by retraining the diffusion prior on multi-view data; this improves consistency but makes the sampling contribution inseparable from prior quality. We instead isolate the sampling axis. The per-step gradient is one noisy sample of an expectation over views; aggregating K samples per step at a fixed total UNet budget reduces variance without touching the prior. We introduce Multi-View Aggregated Score Distillation (MV-SDI), which aggregates gradients from K views per step via gradient accumulation, keeping peak memory unchanged and the 2D prior frozen, and draws views as antithetic antipodal pairs, a prior-independent geometric property, for balanced angular coverage. At a fixed 10,000-UNet-call budget, K=2 raises CLIP R-Precision from 74.8% to 83.8% and CLIP score from 0.297 to 0.312, with consistent gains on HPSv2 and ImageReward and a 0.0% divergence rate on the 43-prompt benchmark; optimization steps halve as a consequence. K=4 gives a fourfold step reduction at R-Precision 86.9% and CLIP 0.307, still well above the single-view baseline on every alignment metric. MV-SDI is compatible with gradient-based score-distillation pipelines, including Score Distillation via Inversion, and requires no retraining and no multi-view data.

Knowledge distillation (KD) is a key technique for compressing Large Language Models (LLMs), yet methods relying on a single KL objective often fail to balance primary distribution fitting with long-tail probability modeling, limiting both generation quality and generalization. To address this, we analyze the complementary roles of forward and reverse KL divergence (FKL/RKL) in distribution alignment from theoretical and empirical perspectives. We then propose a reinforcement-learning-based adaptive KL-weighted distillation framework, in which a policy network dynamically assigns weights to FKL and RKL based on teacher-student distributional characteristics, guided by immediate reward signals to achieve dual alignment on principal and long-tail modes. Extensive experiments demonstrate consistent improvements across Rouge-L and BertScore metrics, surpassing greedy heuristics by 0.4-0.6 points and outperforming other baseline methods on diverse benchmarks.

Agentic search equips large language models with dynamic retrieval abilities, but existing reinforcement learning methods remain limited by reward sparsity in knowledge boundary calibration -- deciding when to trust parametric memory, when to rely on retrieved evidence, and when to abstain. Binary rewards can penalize undesirable outcomes, but provide little guidance on the reasoning process required to make calibrated decisions across different knowledge states. To address this, we propose KbSD (Knowledge boundary Self-Distillation), a framework that tackles this limitation through dense token-level supervision, outcome-level sparse rewards, and quadrant-adaptive optimization. KbSD constructs a hint-augmented teacher, architecturally identical to the student, that receives explicit knowledge boundary signals -- including parametric certainty, retrieval quality, and ground-truth answers -- to generate calibrated reasoning demonstrations. This information-asymmetric self-distillation enables dense supervision without requiring a larger external model. To further account for the heterogeneous reasoning distributions across knowledge states, we introduce a quadrant-adaptive distillation objective: reverse KL for concentrated integration, forward KL for diverse refusal, and Pareto-optimal bidirectional KL for asymmetric quadrants requiring both precision and coverage. Experiments on multiple benchmarks show that KbSD consistently improves both task accuracy and hallucination mitigation over strong baselines, with the largest gains appearing in the challenging quadrants where sparse rewards are least informative.

Diffusion and continuous-flow generative models achieve high-quality generation, and their deterministic sampling can be formulated as solving learned ODE dynamics. However, accurate ODE discretization often requires many steps, making efficient few-step generation a key challenge. Among acceleration strategies, reflow-based distillation simplifies teacher ODE trajectories so that a student model can approximate the teacher transport with fewer steps. We identify a theoretical limitation of this paradigm, namely that trajectory matching can under-determine the distribution induced by the student model. In particular, two student models can attain the same trajectory-matching loss while inducing different endpoint marginal distributions, which may lead to different generation quality. To address this limitation, we introduce a marginal-alignment regularizer that penalizes the discrepancy between the student-induced marginal and the corresponding teacher marginal at the endpoint of each distillation interval. The regularizer is computed by tracking log-density changes along the ODE induced by the student model and evaluating scores from the frozen teacher model, without requiring auxiliary trainable networks or adversarial optimization. The resulting framework applies uniformly to the reflow family, including vanilla reflow and piecewise reflow. We further prove a telescoping total-variation bound showing that local marginal alignment controls the final-time discrepancy between the student-induced and teacher-induced distributions. Experiments on benchmark backbones demonstrate the effectiveness of the proposed method for few-step generation.

We present DistilledGemma, an efficient and accurate system for the HIPE-2026 shared task on person-place relation extraction from multilingual historical newspaper articles in English, German, and French. Our approach adopts a three-stage knowledge distillation pipeline designed to balance classification accuracy with computational efficiency. In the first stage, we systematically explored prompt engineering strategies across eight large language models to identify the most effective reasoning architecture for this challenging task. In the second stage, we applied supervised fine-tuning (SFT) via QLoRA to a Gemma 4 26B A4B teacher model, leveraging its strong multilingual capabilities to generate silver-standard chain-of-thought traces across the training corpus. In the final stage, we performed response-level distillation to transfer these learned reasoning patterns into a compact Gemma 4 E2B student model. In the official evaluation, our team WHEREAMI ranked 3rd on the standard test set with an accuracy profile mean score of 0.688, and 2nd on the binary test set with a mean score of 0.8156. Notably, by distilling knowledge from the 26B teacher to the 2.3B student, we preserved strong reasoning capabilities while reducing the deployed model size to approximately 2.3B effective parameters; the LoRA adapters used during training were merged into the student for inference. This configuration ranked 2nd in the balanced efficiency-accuracy profile across both the standard and binary test sets. These results demonstrate that knowledge distillation provides a practical and scalable solution for historical document processing, achieving competitive performance without excessive computational cost.

Recent large language models (LLMs) achieve strong performance on entity matching without requiring task-specific training data. However, applying these models to large sets of candidate pairs remains slow and costly. In contrast, entity matchers using traditional machine learning methods or small language models (SLMs), such as RoBERTa, offer much faster inference but require task-specific training data. This paper investigates whether the need to provide task-specific training data can be avoided by using knowledge-distillation workflows, in which an LLM serves as a teacher model to label training pairs that are subsequently used to train a smaller student model. We investigate knowledge distillation for entity matching along the following dimensions: pair-selection strategy, teacher model, label post-processing method, and student model. We evaluate the workflows using the Abt-Buy, Walmart-Amazon, WDC Products, DBLP-ACM, and DBLP-Scholar benchmarks, and compare the performance of student models trained with machine-labeled data to the performance of the same models trained using the benchmark training sets. Our experiments show that student models trained using the machine-labeled sets perform approximately on par with models trained on the benchmark training sets, with the remaining differences in both directions staying below two F1 points. Using GPT-5.2 to label the training sets for all five benchmarks costs US$28.31 to US$40.88, whereas manually labeling the same training sets is estimated to require 470 hours of work. At inference time, Ditto is 41.5 to 534 times faster than directly using an LLM to perform the matching tasks. These results indicate that current LLMs, when combined with a suitable pair-selection method, can substantially reduce or even eliminate the manual effort required to label use case-specific training data for entity matching.

Training small language-model agents for long-horizon interactive tasks requires both fast imitation and reward-driven improvement. On-policy distillation (OPD) provides dense teacher guidance and typically improves rapidly in the early stage, but its gains saturate once the student approaches the teacher, limiting the final performance ceiling. Reinforcement learning (RL) directly optimizes environment rewards and encourages exploratory improvement toward a higher reward-defined ceiling, but sparse and delayed feedback makes early-stage learning much less efficient than OPD. In this paper, we propose ATOD (Annealed Turn-aware On-policy Distillation), a hybrid online distillation algorithm that explicitly exploits this complementarity. (1) ATOD uses an annealed OPD-RL schedule: OPD dominates early training to approach teacher-level behavior, while RL is gradually strengthened to drive reward-based exploration. (2) ATOD introduces Turn-level Disagreement-Uncertainty Reweighting (T-DUR), which softly amplifies high-utility turns and improves dense supervision in long trajectories. Experiments on ALFWorld, WebShop, and Search-QA show that ATOD consistently outperforms competing post-training baselines: across the three student sizes, ATOD improves average success rate by 3.03 points over OPD and 23.62 points over GRPO, while surpassing the corresponding teacher models by 2.16 points.

Knowledge Distillation (KD) enables training smaller student models under the guidance of larger teacher models, and the widely adopted TRL library implements it. Yet, TRL treats both models symmetrically, missing opportunities to exploit their pronounced asymmetry in memory footprint, and communication requirements. This paper presents an HPC-aware methodology for KD that decouples teacher and student partitioning efficiently. Our approach achieves up to 67% higher samples-per-second than TRL by avoiding unnecessary teacher-model data structures and selecting the best split strategy. We combine vertical and horizontal partitioning of models, deriving an analytical expression that identifies the existence of inflection points between splitting regimes. These results showed that exploiting teacher--student asymmetry through topology-aware parallelism notably accelerated GKD training on production HPC clusters at our company

Cross-domain diagnosis remains a major challenge in cervical cell pathology due to pronounced domain shifts across institutions and the subtle visual differences among disease stages, which jointly impair model generalization. To address these issues, this paper proposes a two-stage framework for cross-domain cervical cell detection. In the first stage, we propose the Spatially-Continuous Unpaired Neural Schrödinger Bridge (SC-UNSB), which constructs a synthetic intermediate domain to mitigate cross-domain distribution shifts by modeling image translation as an entropy-regularized optimal transport process. In the second stage, we propose a dual-level feature alignment strategy within a knowledge distillation, which progressively aligns shallow structural features and deep semantic representations to facilitate the transfer of domain-invariant knowledge from the source to the target model. Experimental results demonstrate that the proposed method effectively mitigates domain shift and category ambiguity, improving the cross-domain detection performance.

We argue that safety classifiers should model user intent as an explicit signal between the prompt and the final label. To study this, we introduce AIMS, a human-annotated dataset of 1,724 difficult safety prompts, each paired with an intent description and harm label. We use AIMS to evaluate intent-aware training across supervised fine-tuning, preference learning, reasoning distillation, and reinforcement learning. Despite its size, AIMS enables competitive safety classifiers across training regimes: DPO from model-generated intent errors improves over SFT, and intent-conditioned distillation outperforms reasoning-only distillation in most teacher-student pairs. Most notably, directly rewarding intent faithfulness with GRPO yields the strongest average performance across five external safety benchmarks, while our intent-aware models form the inference latency-F1 Pareto frontier. These results show that faithful intent modeling is a compact, high-quality supervision signal for more robust safety classifiers.

Outcome-based reinforcement learning provides a stable optimization backbone for language agents, but its sparse trajectory-level rewards provide little guidance on which intermediate decisions should be reinforced or suppressed. On-policy self-distillation offers dense token-level supervision, yet existing skill-conditioned variants often rely on external skill memories or retrieved privileged context, which are costly to maintain and can be mismatched with the state distribution induced by the current policy in multi-turn interaction. We propose \textbf{OPID} (\textbf{O}n-\textbf{P}olicy Sk\textbf{i}ll \textbf{D}istillation), a framework that extracts skill supervision directly from completed on-policy trajectories. OPID represents trajectory hindsight as hierarchical skills: episode-level skills capture global workflows or failure-avoidance rules, while step-level skills capture local decision knowledge at critical timesteps. A critical-first routing mechanism uses step-level skills when critical decisions are identified and falls back to episode-level skills as default guidance otherwise. The selected skill is injected into the interaction history, allowing the old policy to re-score the same sampled response under both original and skill-augmented contexts. The resulting log-probability shift yields a token-level self-distillation advantage, which is combined with the outcome advantage for policy optimization. OPID thus preserves RL as the primary training objective while introducing dense, distribution-matched hindsight supervision. Experiments on ALFWorld, WebShop and Search-based QA demonstrate that OPID generally improves agent performance, sample efficiency, and robustness over outcome-only RL and existing skill-distillation baselines. Our code is available at https://github.com/jinyangwu/OPID/tree/main.

Image priors can synthesize target conditions for 3D Gaussian street scenes, but independently edited views do not define a coherent 3D target. Direct fitting can propagate view-specific noise, while existing pipelines do not jointly handle imperfect sparse anchors and standard-rasterizer deployment. To address this gap, teacher-relative appearance residual distillation is introduced for appearance baking. A structured space for frequency decomposition, confidence estimation, and primitive-level lifting is formed by residuals between teacher anchors and original renders. The direct optimization signal is supplied by renderer-space matching, while primitive assignment is regularized by support-aware Gaussian-space aggregation. Supported detail is admitted and unsupported noise is suppressed through confidence-gated coarse-to-fine optimization, after which all residuals are baked into fixed-geometry spherical-harmonic coefficients. The teacher and auxiliary training modules are discarded at inference. Evaluation across Waymo street assets, Tanks and Temples scenes, and multiple target conditions shows a favorable overall balance of target alignment, content preservation, artifact suppression, and cross-view consistency over editing-based baselines. Ablations confirm the effectiveness of the main components. Code will be released at https://github.com/Cagares/Baking-for-3D-Gaussian.

As machine learning models and datasets continue to grow, developing complex models has become increasingly computationally demanding. Knowledge distillation reduces deployment cost by compressing a large, well-trained teacher model into a compact student model, but it does not address settings where constructing the teacher itself is the bottleneck. Motivated by this challenge, we introduce Knowledge Cascade (KCas), a reverse knowledge distillation framework that uses information from a small, inexpensive student model to guide the development of a more complex teacher model. Although this direction is counterintuitive because the teacher typically has greater representational capacity, we show that student-to-teacher transfer can be principled when supported by statistical scaling relationships. We first develop KCas for nonparametric multivariate functional estimation in reproducing kernel Hilbert spaces via smoothing splines, where selecting multiple smoothing parameters is a major computational bottleneck. KCas transfers student-selected smoothing parameters to the full-sample regime through asymptotic scaling laws, substantially reducing computational cost for high-dimensional and large-scale datasets while retaining theoretical guarantees. Beyond smoothing splines, we illustrate the same principle through kernel density estimation and deep learning hyperparameter transfer. Simulations and real-data experiments show that KCas achieves substantial computational savings while maintaining strong statistical performance, and can sometimes outperform the corresponding full-sample procedure.

LLM-based text embedders have substantially improved retrieval and semantic representation quality, but their deployment remains costly: large backbone models slow down embedding inference, while high-dimensional full-precision embeddings impose substantial storage and bandwidth overhead on large-scale indexes. In this paper, we present BITEMBED, an extreme low-bit framework for LLM-based text embedding that jointly targets encoding efficiency and vector storage. BITEMBED converts pretrained LLM backbones into BitNet-style embedding encoders with ternary weights, quantized activations, and lightweight normalization refinement. The converted model is adapted to representation learning through continual contrastive pre-training, followed by supervised contrastive fine-tuning with both similarity-distribution distillation and attention-relation distillation from a full-precision teacher. Beyond quantizing the backbone, BITEMBED further trains output embeddings to support multiple storage precisions meeting different storage needs in various scenarios. Experiments on MMTEB (eng, v2) with Qwen3-0.6B and Gemma3-270M show that BITEMBED is largely comparable to full precision teacher embedders. Moreover, BITEMBED flexibly obtains text embeddings of various precisions, achieving a trade-off between performance and storage cost.

Knowledge Distillation (KD) is widely used to obtain compact models for efficient inference in resource-constrained environments. Yet the computational overhead of the distillation process itself is often overlooked, raising the question of whether a better student model can be obtained with less data and less compute via data pruning. However, existing data pruning methods are not designed for KD: some introduce substantial overhead, such as obtaining training dynamics through retraining, while others rely on heuristic selection rules that fail to capture what KD actually requires, often resulting in suboptimal subsets. To address these issues, we propose IF-Beta, an efficient data pruning framework that combines influence functions with a learnable sampling policy. Empirically, we first demonstrate that influence functions can serve as an effective and efficient estimator of sample impact in KD settings, where only a pretrained teacher is available. Building on this, our sampling policy is specifically parameterized by a Beta distribution, whose highly flexible two-parameter family allows the policy to adapt to diverse pruning regimes rather than being tied to fixed heuristic forms. Next, we formulate KD pruning as optimizing this policy through a bilevel objective, where the inner loop operates in the teacher feature space with a KD-aligned objective, enabling fast proxy training, while the outer loop updates the policy parameters to maximize distillation performance. This design ensures that IF-Beta is both computationally efficient and inherently aligned with the goals of KD. Extensive experiments on CIFAR-10/100 and ImageNet show that IF-Beta consistently outperforms other baselines across a wide range of pruning ratios. Remarkably, IF-Beta enables students trained on less data and less compute to surpass the performance of students distilled on the full dataset.

The computational complexity of Transformers scales quadratically with the number of tokens, which significantly constrains the efficiency of vision models, particularly recent ViT-based foundation models in dense prediction tasks. Instance segmentation, a typical dense visual prediction task in the remote sensing field, faces similar challenges. In this paper, inspired by the recent advances of knowledge distillation in large language models, we introduce RS4D - a new remote sensing instance segmentation method with linear computational complexity, which addresses the inefficiency of long sequence modeling through distilled state space modeling (SSM). We propose an adaptive noise and masking knowledge distillation training method for pre-training lightweight SSM backbones, which effectively compresses knowledge from the vast self-attention space into a compact, dense linear state space. We also design a remote sensing image instance segmentation architecture based on this lightweight visual encoder, where we explore variants of three different backbones and two segmentation heads. Extensive experiments are conducted on multiple benchmark datasets, including SSDD, WHU, and NWPU. Compared to ViT-based approaches, our proposed SSM backbone achieves an 8x reduction in parameters and a 9x reduction in FLOPs while maintaining comparable or superior accuracy to both ViT- and CNN-based instance segmentation methods. The implementation codes have been publicly available at https://github.com/QinzheYang/RS4D.

Black-box LLMs (accessible only via API) are vulnerable to distillation attacks, in which an attacker queries the model and trains a student on its outputs. A recent line of work proposes output perturbation defenses that modify the teacher's output to reduce student performance while preserving utility for legitimate users. As a relatively new family of approaches, output perturbation defenses lack a shared threat model, making it difficult to compare them, reason about composing them with other attacks, or evaluate their robustness against realistic adversaries. This underspecification matters beyond technical evaluation: when defenses are deployed to protect intellectual property or justify regulatory compliance, an imprecise threat model can create a false sense of security. We propose a threat model framework that describes attackers along three dimensions: a query budget, a data budget, and an interface profile that captures how attackers interact with the API. Using antidistillation sampling as a case study, we show that whether the defense is considered effective depends on the assumed threat model. We argue that future work on distillation defenses, along with any governance or policy frameworks built around them, should explicitly specify and stress-test attacker capabilities along our three dimensions.

Heterogeneous Knowledge Distillation (HKD) aims to transfer knowledge across varying architectures (e.g., from Transformer to CNN) but inherently suffers from severe training instability. We reveal that this instability stems from two highly coupled challenges: massive feature norm discrepancies that cause optimization drag, and severe gradient conflicts between the primary and distillation objectives arising from distinct inductive biases. To achieve stable distillation, we propose SPOFA, a framework built upon a novel Feature and Gradient Dual Stabilization mechanism. Specifically, at the feature level, we introduce a LayerNorm-based decoupling projector that explicitly decouples feature magnitude from direction, creating a bounded and stable space for semantic alignment. At the gradient level, we propose a momentum-driven Exponential Moving Average (MEMA) dynamic scaler. By establishing a robust historical baseline of the optimization trajectory, MEMA actively evaluates instantaneous gradient conflicts and adaptively penalizes harmful distillation signals, guaranteeing stable convergence. Importantly, SPOFA achieves this dual stabilization with an extremely lightweight parameter footprint. Extensive experiments on two mainstream benchmarks demonstrate that SPOFA achieves state-of-the-art accuracy, significantly outperforming computationally expensive methods while introducing only minimal computational overhead compared to standard baselines.

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.

When two or more parameters or labels produce similar data, they are degenerate, or hard to distinguish. Degeneracies render both label prediction and inverse problems difficult, since both machine learning algorithms and probabilistic samplers rely on the distinguishability of data and its gradients with respect to parameters. However, identifying degeneracies in physical models or real-world datasets can be elucidating about the choice of model or the underlying process that produces the data. We present the degeneracy distillery, a method that (1) detects and (2) resolves degenerate parameter combinations (a) automatically and (b) symbolically, from parameter-data (or parameter-simulation) pairs alone, through estimation and flattening of the Fisher information matrix. By exploring the information geometry of the likelihood, we characterize degeneracies as an intrinsic property of the physical model, requiring no realised data observation. We demonstrate our approach on a range of synthetic and real-world problems, discovering symbolic coordinate transformations that identify the combinations of parameters of a model which yield independent effects on the data. The resulting coordinates flatten the Fisher information in expectation globally, in contrast to posterior-based methods that flatten only at a single point, and substantially reduce the simulation budget required for downstream neural posterior estimation. In test cases we require up to $10\times$ fewer simulations for posterior estimation at matched validation calibration whilst simultaneously gaining physical insight on the system.

Large language models have demonstrated significant capabilities in generating diverse and context-aware responses for empathetic dialogue. However, their computational demands severely limit their deployment in resource-constrained environments. While knowledge distillation offers a promising compression solution, it often fails to transfer the nuanced understanding essential for empathy, as it overlooks the implicit contextual cues that guide human connection. To bridge this gap, we propose a \textbf{pr}ivileged \textbf{i}nformation-enhanced knowledge \textbf{d}istillation method for \textbf{e}mpathetic dialogue generation (PRIDE). Our method leverages privileged information, such as expert psychological annotations or future event summaries, which is available exclusively during training but unavailable at inference time. This allows us to transfer the teacher model's empathetic reasoning to smaller models without relying on extra inputs during deployment. Specifically, PRIDE has three key components: (1) An empathy-reasoning prompt that guides the teacher to explicitly decompose the empathetic process into understanding feelings and analyzing situations step-by-step; (2) A multi-source attention mechanism that directs the student to effectively integrate privileged information; (3) A dual-alignment loss that combines reversed Kullback-Leibler divergence and maximum mean discrepancy to ensure robust knowledge transfer at both logit and feature levels. Experiments on multi-modal and text-only datasets demonstrate that our method achieves competitive performance, and in some cases matches or even surpasses larger teacher models in terms of accuracy and semantic relevance.

Text-attributed graphs (TAGs) are widely used in many real-world domains, and learning on TAGs requires jointly modeling text semantics and graph structure. A standard approach for modeling TAGs is to combine a language model (LM) and a graph neural network (GNN), but joint training is computationally expensive and difficult to scale. Dataset distillation is a promising way to reduce training costs, but existing methods are not well suited to TAGs because they are typically designed for a single modality or still require repeatedly training expensive LM-GNN models on the full dataset during distillation. To address this, we propose TaLK, an effective dataset distillation method for TAGs that couples an LM with a graph-aware neural tangent kernel. This design enables efficient dataset distillation, avoiding repeated joint training on the full dataset while reflecting both textual and structural information for effective TAG learning. Experiments on multiple TAG benchmarks show that TaLK consistently outperforms existing baselines and achieves up to 97% of full-dataset performance with only 1% synthetic data.

Large language models (LLMs) achieve strong performance across many tasks, but their high computational cost limits deployment in resource-constrained environments. Knowledge Distillation (KD) offers a practical solution by transferring knowledge from a teacher model of a larger size to a smaller student model. While prior work has mainly examined task-specific or small-scale settings, the post-training stage for building general instruction-following models has received limited attention. In this paper, we conduct a systematic study of KD in post-training using the large-scale Tulu 3 dataset. We find that KD outperforms supervised fine-tuning (SFT) in low-data regimes, but its advantage diminishes as more training data is added. Distilling from a stronger instruction-tuned teacher restores substantial gains even with abundant data, indicating that KD remains effective when the teacher provides knowledge that the student cannot easily acquire from the training data alone. We further study domain-specific, low-resource scenarios and propose a two-stage KD strategy that leverages synthetic teacher-labeled data followed by refinement on human annotations. This method consistently improves student performance, providing practical guidance for building compact models in data-scarce environments.

We extend our prior work on Text2DSL automatic generation of domain-specific language (DSL) code from natural language descriptions along two complementary axes. First, we replace prompt-only synthetic generation with context-aware distillation, in which a teacher large language model (DeepSeek-V4-Flash) operates under an explicitly defined structured context comprising a BNF grammar, an API specification, and a closed identifier vocabulary; the resulting corpus is verified by a two-tier pipeline combining AST validation through esprima and runtime acceptance through the production polkitd daemon and the pkcheck client. This scales the verified PolkitBench corpus from 4,204 to 10,073 natural-language-to-Polkit-rule pairs at 100.0% AST validity and 99.7% runtime pass rate. Second, we conduct the per-component factorial ablation of structured context that was identified as future work in the precursor study: eight conditions C0-C7 are evaluated on GigaChat-10B-A1.8B with the new corpus. Three findings emerge. (i) The new harder corpus collapses the baseline mode (Syntax Valid 97.6% -> 58.5%, Combined Score 0.482 -> 0.252), whereas the context-enhanced mode degrades only marginally (Syntax 98.6% -> 97.4%, Combined 0.801 -> 0.750), confirming that structured context is not a cosmetic improvement but a load-bearing mechanism. (ii) The best absolute condition is the full context C7 across all metrics, while the strongest partial conditions (C5 = BNF + Vocabulary, C6 = API + Vocabulary) both contain the vocabulary. (iii) A Shapley-style decomposition assigns the largest semantic-quality effect to the vocabulary (Combined +0.198), the largest structural-validity effects to API (+24.7 pp) and BNF (+22.3 pp).

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.

Streaming text-to-speech synthesis in cascaded LLM-TTS systems still faces latency challenges as most TTS models require full context before initiating generation. We present S5-TTS, a streaming variant of T5-TTS that enables low-latency, word-by-word incremental speech synthesis through encoder-decoder language modeling and monotonic alignment learning. S5-TTS begins generating speech immediately after receiving the first few words, substantially reducing end-to-end response latency. To maintain quality under limited lookahead, we introduce a lookahead-causal masking mechanism with Conv-based auxiliary attention that preserves intelligibility and speaker similarity, and employ interleaved multi-source distillation to further restore naturalness. Experiments show that S5-TTS achieves comparable quality to full-context T5-TTS, supports zero-shot synthesis with high speaker similarity, and significantly reduces end-to-end latency for practical conversational AI systems.

Knowledge Distillation (KD) has established itself as a pivotal technique for compressing large pre-trained language models. However, existing methods that force a student to strictly mimic the teacher's sentence embeddings or internal features often incur prohibitive computational costs and yield suboptimal performance due to the inherent capacity gap. To address these challenges, we propose TALAS (Teacher-Anchored Layer Alignment with Sharpness-aware minimization), a unified framework that synergizes hierarchical (multi-layer) alignment with robust optimization. First, we introduce a Teacher-Anchored mechanism that selectively distills final sentence embeddings only into the student's upper layers, thereby reducing overhead while respecting capacity constraints. Second, we bridge the semantic gap in lower layers via Layer-Aligned Self-Distillation, which propagates knowledge top-down using internal geometric relational constraints in the embedding space. Finally, to prevent the student from memorizing point-wise teacher noise, we integrate Adaptive Sharpness-Aware Minimization (ASAM) into the training objective, guiding the model towards flat minima for enhanced generalization. Empirical results on standard sentence embedding benchmarks demonstrate that TALAS consistently outperforms strong distillation baselines while achieving superior training efficiency in terms of computational cost and memory footprint.

Dataset distillation (DD) has proven to reduce training cost while preserving accuracy. While promising, the factors that make one distilled dataset more effective than another remain poorly understood. In this work, we investigate this question through the lens of discrete visual tokenizers. Whereas many prior DD efforts emphasize matching global data distributions, we suggest that the effectiveness depends on which semantic concepts are captured and how they are composed. Discrete visual tokenizers provide a finite vocabulary that enables direct statistical analysis of such compositional structure. Through quantitative analysis of token-level statistics, we introduce the structural score to measure the adequacy of token compositions. We observe that distilled datasets with balanced token composition yield higher validation performance. On the other hand, divergence from the original data does not necessarily harm performance. We further show that samples with high structural scores in the discrete token space can effectively guide diffusion-based DD. Our findings highlight the importance of token composition in dataset effectiveness, offering a principled complement to distributional similarity considerations in DD.

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.

As expressive text-to-speech (TTS) and voice conversion (VC) systems increasingly generate non-verbal vocalizations (NVVs) to enhance naturalness, reliable speaker verification (SV) becomes essential to objectively assess identity consistency across both verbal and non-verbal segments. Yet current SV systems generalize poorly to NVVs, and fine-tuning on NVV data causes catastrophic forgetting of speech performance. We present the first systematic study across 10 NVV types and propose a framework combining frozen Data2Vec self-supervised features with ECAPA-TDNN, enhanced by a Mixture of Experts (MoE) module with learned domain-aware routing. A conditional distillation loss on speech inputs via a pretrained teacher retains speech-to-speech accuracy, while a contrastive loss bridges the speech-NVV domain gap. Our method reduces speech-NVV EER from 38.93% to 22.66% over a pretrained baseline, and improves speech EER from 13.17% to 9.24% via distillation.

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.

Egocentric video understanding is inherently limited by the narrow perspective of wearable cameras: a single viewpoint, a single modality, a single model cannot capture the full richness of human action. We argue that a truly expressive egocentric representation must subsume complementary knowledge across viewpoints, modalities, and foundation model representations, yet remain deployable from egocentric video alone. To this end, we introduce a hierarchical multi-teacher distillation framework that produces UNIEGO, a unified egocentric encoder trained with nine teachers spanning ego-exo viewpoints, RGB, depth, and skeleton modalities, and four foundation models. Rather than distilling directly from heterogeneous teachers whose incompatible architectures and feature geometries induce conflicting gradients, our framework interposes a layer of representation-specific Proxy models that translate diverse teacher knowledge into a homogeneous egocentric space. A second distillation stage, Selective Proxy Distillation (SPD), then adaptively selects, for each training sample, the subset of proxies that are both correct and confident, distilling exclusively from reliable supervision and suppressing erroneous signals. SPD is further stabilized by initializing UNIEGO as a learned convex combination of proxy parameters, placing the unified model in a well-conditioned region of the loss landscape before distillation begins. UNIEGO achieves state-of-the-art performance across three egocentric video understanding tasks - action recognition, video retrieval, and action segmentation on three challenging ego-exo benchmarks, outperforming naive multi-teacher distillation baselines and demonstrating that structured, proxy-mediated knowledge transfer yields richer and more discriminative egocentric representations.

In batch-style trace distillation, the same memory operation may receive contradictory feedback across different batches. Existing methods lack a cross-batch, operation-level evidence accumulation mechanism, making it impossible to distinguish stably effective operations from accidental hits. This paper formalizes the requirement as two structural conditions, alignability and comparability, and proposes Marginal Advantage Accumulation (MAA). MAA constructs differential signals to make them comparable across batches, accumulates signed evidence per operation via EMA, and ensures cross-batch traceability through semantic identity merging. As a post-processing architecture, MAA achieves the best results in 14 out of 16 settings across 4 benchmarks and 4 target models, consistently outperforming existing batch-level distillation baselines and matching or surpassing online alternatives in most settings, while reducing optimization-phase token consumption by approximately 75%.