Knowledge Distillation

Reinforcement learning (RL) is a natural fit for agentic knowledge base question answering (KBQA), where a model must issue executable actions, observe knowledge-base feedback, and eventually return an answer. However, current RL-based KBQA systems mainly optimize sparse rewards from the final answer, leaving intermediate action errors weakly supervised. This is especially limiting for logical-form annotated KBQA benchmarks: gold logical forms can be converted into executable action sequences, but existing pipelines use them mainly for warm-start data construction rather than for on-policy RL updates. We propose GAPD, a training-time Gold-Action Policy Distillation framework that adds dense token-level guidance to outcome-based RL. To align gold actions with on-policy student rollouts, GAPD uses MID-ANCHOR MATCHING: it treats the intermediate entities reached during student exploration and gold execution as state anchors, and matches student states to gold states through these explored entity sets. The current policy conditioned on this aligned gold action serves as a stop-gradient teacher, whose token distribution is distilled back to the ordinary student policy over generated action-token spans. GAPD consistently surpasses the current state of the art on WebQSP, GrailQA, and GraphQ.

Mainstream strategies for finetuning pretrained multimodal models often degrade out-of-distribution (OOD) robustness, a phenomenon known as catastrophic forgetting. In this paper, we develop a theoretical framework for multimodal contrastive finetuning, yielding closed-form solutions and a geometric decomposition for each strategy. This framework shows that self-distillation is more effective than other regularization approaches to retain the knowledge of the pretrained model. Our analysis reveals a largely overlooked limitation: standard Exponential Moving Average (EMA) teachers, widely used in robust finetuning, suffer from collapse. To solve this, we prove that a Weighted Moving Average (WMA) teacher maintains a persistent regularizing force over finite horizons and yields bias-free convergence in the task subspace while preserving orthogonal knowledge. These insights motivate TRACER (Trajectory-Robust Anchoring for Contrastive Encoder Regularization), which combines contrastive learning with WMA-guided multi-perspective distillation. Extensive experiments on CLIP finetuning demonstrate consistent OOD accuracy and calibration gains across three backbone architectures, and comprehensive ablations confirm that TRACER is both principled and robust to hyperparameter choices. Code is available at https://github.com/HesamAsad/TRACER.

Efficient Distillation (EDistill) compresses large language models (LLMs) by structured pruning parameters and tuning lightweight modules with high training efficiency. Although these EDistilled LLMs achieve state-of-the-art (SOTA) performance on general ability benchmarks relative to similarly sized LLMs, we identify a severe degradation in their multi-step reasoning ability, which we term reasoning collapse. We systematically analyze the geometric origins of reasoning collapse and show that the SOTA EDistill method based on width-reducing projection matrices suffers from eRank collapse, in which the effective rank (eRank) of hidden representations drops. We theoretically explain how singular values of randomly initialized projection matrices become unevenly distributed, leading to eRank collapse and thus token indistinguishability. To address this issue, we propose RED (Reasoning-preserved Efficient Distillation) for LLMs, which introduces activation-aware initialization to initialize projection matrices as channel-selection matrices, thus theoretically mitigating eRank collapse. Experiments on Llama and Qwen series demonstrate that RED substantially recovers reasoning while maintaining high training efficiency and SOTA general ability.

Knowledge distillation (KD) transfers a single scalar prediction from a large foundation model (FM) to compact vertical models (VMs), suffering from diminishing transfer ratio -- the fraction of FM improvement captured by the VM -- as a single scalar cannot convey the rich intermediate knowledge that larger FMs learn. To address this bottleneck, we propose LoopFM (Learning frOm HistOrical RePresentations of FM), a framework that opens a high-bandwidth transfer channel by structuring FM intermediate embeddings as input features (e.g., user history sequence) for downstream VMs, without requiring real-time FM inference at serving and architectural coupling between FM and VM. We provide a theoretical framework for LoopFM with a gain decomposition and transfer-ratio analysis. On three public benchmarks, LoopFM demonstrates strong AUC improvements (e.g., 6%+ on TaobaoAd) and complementary knowledge transfer capability with KD. On industrial-scale systems (billions of examples, trillion-parameter FMs), LoopFM approximately doubles the knowledge transfer ratio on top of KD, delivering a +0.5% conversion improvement in the first half after its initial launch, and +1.03% and +1.22% conversion improvement from two individual launches in the subsequent half.

Reasoning distillation transfers complex reasoning abilities from large language models (LLMs) to smaller ones, yet its success depends on how well the training data align with the student model. This paper introduces the Data-Model Compatibility (DMC) metric, which can be used to assess the suitability of a dataset for reasoning distillation on a student model. DMC provides an assessment by jointly considering data quality, relative difficulty, and student capability. We validated the effectiveness of DMC from two perspectives: (1) DMC exhibits a strong correlation with reasoning distillation performance; and (2) using DMC as the criterion for data selection leads to improved reasoning distillation performance. Both findings are consistently demonstrated across multiple student models and tasks. Moreover, since the DMC of each dataset dynamically changes during training, our experiments demonstrate that dynamically selecting datasets based on DMC can further enhance performance.

Understanding dynamic 3D environments is essential for safe autonomous driving, particularly when reasoning about human-centric, nonrigid agents. However, existing weakly supervised occupancy prediction frameworks predominantly assume rigid-body motion and rely on simple frame-to-frame offsets, limiting their ability to capture fine-grained deformations and maintain temporal coherence. To address this issue, we propose DeGO, a deformable Gaussian occupancy framework that unifies decoupled Gaussian deformation with factorized 4D foundation-model distillation. DeGO disentangles rigid and nonrigid motion, enabling each Gaussian primitive to evolve through both deformation and offset-based updates. In parallel, a factorized 4D distillation strategy transfers cross-camera and cross-frame knowledge from the VGGT foundation model, producing foundation-aligned features that enhance temporal consistency. Experiments on the Occ3D-NuScenes benchmark demonstrate that our method achieves state-of-the-art performance under weak supervision, delivering 13.5% gains on human-centric instances and 10.9% overall improvements. These results highlight the effectiveness of deformation-aware and foundation-guided occupancy modeling for dynamic scene understanding. The code is publicly available: https://github.com/vita-epfl/DeGO

Masked language modeling has become a standard pretraining objective for training encoder-based language models. In this approach, certain tokens in the input are masked, and the model learns to predict them using the surrounding context. This process enables the model to capture both syntactic and semantic properties of language. Conventionally, the tokens selected for masking are chosen at random, which may not always yield the most effective learning signals. In this work, we examine a token masking strategy based on entropy distribution. We use the model's entropy over token predictions to identify which tokens should be masked. This method aims to target tokens that are more informative and uncertain to improve the training efficacy. We also propose a novel self-masking approach that enhances training efficiency without relying on an external reference model. Experimental results demonstrate that our method achieves an average performance improvement of 5% in GLUE scores compared to the baseline. Further, we experiment with combining knowledge distillation with entropy masking, resulting in the best overall results.

While Knowledge Editing (KE) enables efficient updates, its dominant Static Fact Overwriting paradigm treats LLMs as discrete databases, forcibly injecting isolated facts. Fracturing pre-trained logical topologies, this triggers Epistemic Dissonance -- a pathology where un-evolved legacy priors force the model to explicitly negate the injected update. Idealized interventions reveal that this is an inherent structural flaw rather than mere algorithmic noise, with a zero-distortion proxy yielding a catastrophic 95.6% self-refutation rate. Given the causally driven nature of real-world knowledge, grounding updates in explicit causal narratives effectively collapses this conflict rate to just 6.6%, underscoring the imperative for a paradigm shift toward Causal Editing. To internalize this evolution, we propose CODE (Causal On-policy Distillation for Editing). By coupling causal bootstrapping with asymmetric on-policy distillation, CODE engraves causal transition logic directly into parametric memory. Experiments on LLaMA-3.1 and Qwen-2.5 show CODE drastically suppresses self-refutation to 1.8% while securing robust multi-hop accuracy (up to 83.5%), seamlessly transforming discrete fact injection into coherent knowledge evolution. Code is available at https://github.com/CrashBugger/CODE.

Mixture-of-Experts (MoE) is now the dominant architecture for frontier language models, yet it requires all expert parameters to be loaded in memory, making it less preferable for memory-constrained deployment. Existing compression methods reduce the number of experts but the output remains an MoE model with the same fundamental limitation. We present the first systematic framework for converting a trained MoE into a standard fully dense architecture: experts are scored, selected, and grouped, then concatenated into a dense FFN and refined by knowledge distillation from the MoE teacher. We evaluate 7 scoring, 5 grouping, and 2 magnitude scaling methods across a range of selected expert counts on Qwen3-30B-A3B, yielding 350 configurations. We find that the choice of scoring method is the most impactful, with our novel diversity-aware scoring consistently outperforming prior methods on Qwen3-30B-A3B, DeepSeek-V2-Lite, and GPT-OSS-20B. Under a controlled comparison at matched parameter count, MoE-to-dense outperforms dense-to-dense pruning by +6.3 pp in average downstream accuracy after ~4B-token distillation at 1.6x faster training wall-clock speed.

Context distillation compresses contextual information into model parameters, yet existing methods often ignore how multiple distilled latent memories should be stored, retrieved, and safely activated in non-oracle settings. We formulate context distillation as a latent memory management problem. We distill each context into an independent LoRA adapter, forming a modular memory bank that enables explicit memory selection. Given a query, our framework retrieves candidate memories, routes the query to the most suitable adapter, and uses a Self-Gating mechanism to decide whether latent memory should be activated. To improve efficiency, we further introduce cache sharing to reduce management overhead during inference. Experiments show that our method substantially outperforms baselines with retrieval, while Self-Gating improves robustness by deactivate unnecessary latent memories.

Given the real-time demands of UAV tracking, many methods simplify the backbone to reduce computation, but this often weakens feature representation and degrades performance in complex scenarios. To alleviate this issue, we propose EATrack, an efficient and asymmetric UAV tracking framework centered around a teacher-guided dual-branch distillation strategy that enhances the feature expressiveness of the lightweight student model. Specifically, EATrack investigates two complementary perspectives of knowledge transfer: spatially focused feature-level distillation that compensates for weakened representations by guiding the student to learn strong target representations, and prediction-level distillation that enhances spatial localization by learning the teacher's capability for accurate target localization. Furthermore, to enhance robustness against appearance variations, we introduce a fine-grained target-aware distillation strategy that selectively transfers the teacher's target modeling capacity to the student. A temporal adaptation module is incorporated at inference to enhance robustness over time. Experiments on five UAV benchmarks demonstrate that EATrack achieves a favorable balance between accuracy and speed. Code: https://github.com/GXNU-ZhongLab/EATrack

Knowledge distillation is a powerful method for model compression, enabling the efficient deployment of complex deep learning models (teachers), including large language models. However, its underlying statistical mechanisms remain unclear, and uncertainty evaluation is often overlooked, especially in real-world scenarios requiring diverse teacher expertise. To address these challenges, we introduce \textit{Multi-Teacher Bayesian Knowledge Distillation} (MT-BKD), where a distilled student model learns from multiple teachers within the Bayesian framework. Our approach leverages Bayesian inference to capture inherent uncertainty in the distillation process. We introduce a teacher-informed prior, integrating external knowledge from teacher models and task-specific training data, offering better generalization, robustness, and scalability. Additionally, an entropy-based weighting mechanism adaptively adjusts each teacher's influence, allowing the student to combine multiple sources of expertise effectively. MT-BKD enhances the interpretability of the student model's learning process, improves predictive accuracy, and provides uncertainty quantification. We validate MT-BKD on both synthetic and real-world tasks, including protein subcellular location prediction and image classification. Our experiments show improved performance and robust uncertainty quantification, highlighting the strengths of our MT-BKD framework.

Self-Distillation Policy Optimization (SDPO) provides dense token-level credit assignment for reinforcement learning with large language models by leveraging the model's own feedback-conditioned predictions as a self-teacher. Unlike GRPO, however, whose group-relative advantage naturally concentrates learning on a sweet spot of intermediate-difficulty questions, SDPO's KL-based advantage lacks an implicit notion of difficulty awareness. We analyze this gap through the lens of GRPO's advantage normalization. Extending the learnability framework to normalized rewards, we show that normalization absorbs the variance term $p(1-p)$, equalizing leading-order learnability across questions and leaving $\sqrt{p(1-p)}$ as the sole residual scaling factor in the per-question gradient. This analysis yields a simple prescription: weight each question's SDPO loss by $[\hat{p}(1-\hat{p})]^{1/2}$, resulting in SC-SDPO, a scale-consistent variant of SDPO. The proposed weights are obtained as a zero-cost byproduct of on-policy rollouts with batch-adaptive normalization, inducing an implicit curriculum that dynamically tracks the model's evolving competence. Experiments on scientific reasoning and tool-use benchmarks demonstrate that SC-SDPO consistently improves over SDPO, yielding gains of +3.2/+4.3 (mean@16/maj@16) on Qwen3-8B and +1.8/+3.0 on OLMo-3-7B, while preserving stable training dynamics throughout optimization.

Many organizations lack computational resources to fine-tune large language models (LLMs) on private (unshareable) data for better utility, while fine-tuning tiny language models (TinyLMs) alone performs poorly. To address this bottleneck, we propose a data-free knowledge distillation framework that generates LLM update vectors based on TinyLMs fine-tuned on private data. An update vector is a vector of parameter changes from an initial model to its fine-tuned version on a dataset, capturing the effect of cumulative gradient steps during fine-tuning. The key idea of our framework is a novel Gradient Transformer that transforms TinyLM's update vectors into LLM's update vectors. As derived from shadow datasets, Grad-Transformer captures the correlation between TinyLM and LLM update vectors, enabling third-party providers to generate LLM update vectors given the organization's TinyLM update vectors without accessing the organization's private data. The framework supports multi-organization collaboration to jointly update LLMs, improving performance and cost-efficiency. Extensive experiments across language modeling and reasoning tasks show that Grad-Transformer remarkably outperforms state-of-the-art knowledge distillation baselines, even under strict differential privacy protection.

Distilling demonstration effects into hidden-space interventions offers a lightweight alternative to full finetuning. However, existing multimodal variants are mostly evaluated on short-form tasks, where outputs end after a few tokens. Extending these methods to long-form generation exposes a fundamental yet underexamined limitation: token-level distillation implicitly treats all output tokens as equally informative, but long-form outputs are dominated by high-frequency template and grammatical tokens, while the tokens that actually determine output quality are sparsely distributed. In medical report generation (MRG), two such decisive tokens stand out: pathology-related tokens that determine diagnostic content, and the end-of-sequence (EOS) event that determines termination. Both receive insufficient supervision under uniform cross-entropy, and autoregressive decoding further compounds the problem by drifting away from teacher-forced trajectories. We propose DIVE, a frozen-backbone distillation framework that addresses long-form report generation through two complementary mechanisms matched to these failures. Decisive-token supervision restores supervision balance by upweighting the cross-entropy contribution of pathology-related tokens and the EOS event, ensuring that content fidelity and termination are learned during training rather than imposed at decoding time. State-conditioned dynamic steering replaces fixed open-loop residuals with hidden-state-dependent adapters, allowing the injected signal to adapt as decoding drifts. Experiments on MIMIC-CXR and CheXpert Plus with two medical VLM backbones show that DIVE consistently ranks among the strongest methods across lexical and clinical-proxy metrics. Our method achieves the best BLEU-4, ROUGE-L, and RadGraph F1 in all dataset--backbone settings, while remaining competitive on coarse label-level CheXbert F1.

Multimodal Large Language Models (MLLMs) have achieved remarkable progress on vision-language tasks, but they may also memorize and expose sensitive or restricted knowledge, raising concerns about privacy and broader safety risks. Machine Unlearning (MU) provides a promising way to remove targeted undesirable knowledge from trained models without retraining from scratch while preserving general model utility. Nevertheless, effective unlearning in MLLMs remains particularly challenging. Existing training-based methods often struggle to balance unlearning effectiveness and model utility. In contrast, training-free methods such as in-context unlearning preserve model utility by avoiding parameter updates, but they do not remove memorized knowledge at the parameter level and may remain vulnerable to reverse-engineering attacks. More importantly, in-context unlearning is insufficient in multimodal settings, where visual inputs can provide strong conditioning signals and induce undesirable outputs. To address these challenges, we propose Visual-Noise Guided In-Context Distillation (VGID), a distillation-based framework for MLLM unlearning. VGID dynamically constructs an unlearning-oriented teacher distribution from the frozen base model through dual-modal intervention that combines visual perturbation with textual in-context unlearning. The resulting intervention-induced distribution serves as a teacher signal for distillation, guiding the student model toward parameter-level unlearning without requiring external teacher models or explicit undesirable response annotations. Experimental results show that VGID achieves strong unlearning effectiveness while preserving competitive model utility, reducing forget set ROUGE-L by 0.371 with only a 0.055 drop in retain set ROUGE-L in a representative setting.

Unified multimodal embedding spaces have become the standard interface for cross-modal retrieval and multimodal RAG, and recent audio-video-text (AVT) encoders extend this setting to three modalities. Such encoders can produce a joint (T,V,A) embedding whenever all three modalities are available, but standard pairwise InfoNCE objectives leave this signal unused during training. We close this gap with fusion-as-teacher distillation, which treats a stop-gradient copy of the fused embedding as a teacher signal for the single-modal embeddings, paired with a Tuple-InfoNCE term that supervises the fused embedding directly. We instantiate this objective as OmniRetriever-7B. Across six zero-shot retrieval benchmarks, OmniRetriever-7B surpasses the closed-source Gemini Embedding 2 by 13.3-18.0 R@1 on Clotho and SoundDescs, and reaches the contemporary zero-shot specialist band of open video-text encoders on MSR-VTT and MSVD. To stress-test joint representations, we further release OmniRetriever-Bench, a 12-direction AVT retrieval benchmark totaling 3782 triples; on it OmniRetriever-7B attains AVG-all 34.84, improving over Gemini Embedding 2 by 1.72 and over the best prior open-source AVT method by 8.03.

Omnimodal large language models enable unified audio video understanding, but long joint token sequences make inference costly, and existing benchmarks do not fully isolate audio visual association in noisy user generated videos. We introduce UGC-AVQA, a public UGC benchmark with 1,000 videos and 4,816 QA pairs, where an audio removal test ensures that benchmark questions require both acoustic and visual evidence. To reduce inference cost, we propose OMAC, a training free plug in compression method that preserves salient visual memory and temporally grounded audio anchors. To further make compact models robust to compressed inputs, we introduce O-MARC, a compression distillation framework for learning with memory compressed multimodal contexts. On Qwen2.5-Omni-3B, O-MARC improves the average score across four benchmarks to 45.8, outperforming full token inference at 44.1 and OmniZip at 41.0. OMAC also keeps inference efficient, reducing latency by 34.6% (1.53$\times$ speedup) and memory by 34.7% compared with full token inference.

Knowledge distillation (KD) transfers knowledge from a large teacher model to a smaller student. In language modeling, the student is trained either on tokens sampled from the teacher (hard labels) or the teacher's full next-token distribution (soft labels). Despite soft labels appear strictly richer, we find that mixing hard and soft labels consistently yields better results. Crucially, we show that this gain cannot be explained by closer teacher matching during training. Instead, it comes from reduced exposure bias, the mismatch between training and inference distributions. To explain this phenomenon, we introduce the Bridge-Garden Decomposition theory, which categorizes generation steps into two types: Bridges, where the next token must be exact, and Gardens, where it can be flexible. We show that hard-only KD excels in Bridges by avoiding risky deviations, while soft-only KD preserves diversity in Gardens. A hybrid strategy handles both cases and, as a result, reduces exposure bias across the sequence. Guided by this theory, we develop a family of Bridge-Garden hybrid supervision methods that adaptively balance hard and soft labels. Across a primary suite of seven teacher-student pairs (including Qwen, Llama, Gemma, and DeepSeek) and benchmarks in reasoning and coding, our approach outperforms divergence-based and on-policy KD baselines while reducing training cost by 9.7x, enabling efficient model compression. Code is available at https://github.com/ghwang-s/bridge_garden_hybrid_kd_release.

We release Llamion, a family of 14B-parameter open-weight language models obtained by transforming Orion-14B into the standardized Llama-family architecture. The transformation is performed by Efficient Knowledge Preservation for Transformation (KEPT), a recipe that combines (i) Normal Parameter Mapping (NPM) for unchanged modules, (ii) Optimized Parameter Mapping (OPM), a training-free LayerNorm-to-RMSNorm initialization we prove optimal under the near-zero-mean activation regime induced by weight decay, and (iii) Cross-architecture Knowledge Distillation (XKD), an equal-size frozen-teacher distillation that aligns the converted model's outputs with the source model's on any reasonable input distribution. Llamion recovers Orion's behaviour on H6, MT-Bench, and KoMMLU with only ~123M tokens on a single A100 in four days; Llamion-Base reaches 66.87% on KoMMLU, exceeding the next-best entry of the Open Ko LLM Leaderboard by >7.0 absolute points at submission time. Capabilities entirely absent from the transfer corpus (Python programming and 200K-token context handling) survive the architectural transition intact. We release three checkpoints (Base, Chat, LongChat) that load with trust_remote_code=False in the Hugging Face Transformers library.

Knowledge distillation generally assumes a strong-to-weak relationship where stronger teachers yield better students. In this work, we examine this assumption about distillation in large language model pretraining. By varying architecture sizes and training token budgets, we create strong-to-weak, same-level, and weak-to-strong teacher-student relationships, and study distillation's effectiveness under each. We find that the teacher need not be strong: with proper mixing of the language modeling and knowledge distillation losses, even small and undertrained teachers improve larger students. At the same time, a stronger teacher is not always better: pushing the teacher further, through more parameters or more training tokens, can saturate or even reverse the distillation gains. We further observe that distillation improves generalization (out-of-distribution and downstream performance) more readily than in-domain fitting. Together, these results challenge the common belief that distillation pretraining always requires a strong teacher.

Metagenomic taxonomic annotation aims to identify the microbial origins of DNA fragments in environmental samples. Traditional methods that rely on sequence similarity are often constrained by the high microbial diversity and the incompleteness of reference databases, which has motivated the development of learning approaches such as Taxometer that perform post hoc correction to learn more informative metagenomic sequence representations. However, these methods typically rely on labels derived from similarity search tools during training, which inevitably introduces noise that can impair representation learning and degrade classification performance. To address this issue, we propose TaxDistill, a knowledge distillation framework for metagenomic classification. We introduce GenomeOcean, a 500M parameter genomic foundation model, as the teacher network to extract deep semantic features and generate soft labels based on confidence. By distilling this soft label information into a lightweight student network, TaxDistill effectively reduces the label noise introduced by initial retrieval tools. Comprehensive experiments on seven diverse CAMI2 datasets demonstrate that TaxDistill outperforms existing baselines in most scenarios. For instance, on the Gastrointestinal dataset, it improves the F1 score of MMseqs2 from 0.763 to 0.941, outperforming the Taxometer baseline. Overall, TaxDistill provides a reliable method for label correction in complex metagenomic analysis.

Large language model (LLM) agents still struggle with long-term memory question answering, where answer-supporting evidence is often scattered across long conversational histories and buried in substantial irrelevant content. Existing memory systems typically process memory before future queries are known, then retrieve the resulting units based on similarity rather than their utility for answering the query. This workflow leaves downstream answerers to denoise retrieved candidates and reconstruct query-specific evidence. We present DeferMem, a long-term memory framework that decouples this problem into high-recall candidate retrieval and query-conditioned evidence distillation. DeferMem uses a lightweight segment-link structure to organize raw history and retrieve broad candidates at query time. It then applies a memory distiller trained with DistillPO, our reinforcement learning algorithm for distilling the high-recall but highly noisy candidates into a set of faithful, self-contained, and query-conditioned evidence. DistillPO formulates post-retrieval evidence distillation as a structured action comprising message selection and evidence rewriting. It optimizes this action with a decomposed-and-gated reward pipeline and structure-aligned advantage assignment, gating reward components from validity to quality checks while exposing task-level correctness feedback early and assigning each reward to its responsible output span. On LoCoMo and LongMemEval-S, DeferMem surpasses strong baselines in QA accuracy and memory-system efficiency, achieving the highest QA accuracy with the fastest runtime and zero commercial-API token cost for memory operations.

The platonic representation hypothesis suggests that sufficiently large models converge to a shared representation geometry, even across modalities. Motivated by this, we ask: Can the semantic knowledge of a language model efficiently improve a vision model? As an answer, we introduce TextTeacher, a simple auxiliary objective that injects text embeddings as additional information into image classification training. TextTeacher uses readily available image captions, a pre-trained and frozen text encoder, and a lightweight projection to produce semantic anchors that efficiently guide representations during training while leaving the inference-time model unchanged. On ImageNet with standard ViT backbones, TextTeacher improves accuracy by up to +2.7 percentage points (p.p.) and yields consistent transfer gains (on average +1.0 p.p.) under the same recipe and compute. It outperforms vision knowledge distillation, yielding more accuracy at a constant compute budget or similar accuracy, but 33% faster. Our analysis indicates that TextTeacher acts as a feature-space preconditioner, shaping deeper layers in the first stages of training, and aiding generalization by supplying complementary semantic cues. TextTeacher adds negligible overhead, requires no costly multimodal training of the target model and preserves the simplicity and latency of pure vision models. Project page with code and captions: https://nauen-it.de/publications/text-teacher

Closed-weight generative services are increasingly deployed through query-based APIs, where users can obtain generated outputs while model parameters remain inaccessible. However, such deployment does not prevent model stealing: an attacker can repeatedly query the service, collect large volumes of released synthetic images, and use them as training data for a private substitute model. This query-output-driven process enables unauthorized knowledge distillation and capability replication without direct access to the original weights. To mitigate this threat, a practical defense should preserve the visual fidelity of released images, provide explicit control over perturbation magnitude, and scale efficiently to large-volume output release. We present WaveGuard, a single-pass, generator-based protection framework that safeguards released synthetic images under a user-specified perturbation budget. WaveGuard employs a frequency-aware perturbation generator to inject structured, imperceptible perturbations that maintain perceptual utility for benign viewers while reducing the usefulness of protected images as training data for unauthorized student models. Extensive experiments under WikiArt-related synthetic-output distillation settings show that WaveGuard achieves a favorable efficacy--fidelity--efficiency trade-off, with explicit imperceptibility control and substantial gains in protection efficiency.

Adversarial Distillation aims to enhance student robustness by guiding the student with a robust teacher's soft labels within the min-max adversarial training framework, yet its success is notoriously inconsistent: a more robust teacher often fails to improve, or even harms, the student's robust generalization. In this paper, we identify a key mechanism of this teacher dependency: the misalignment between the teacher's supervisory confidence and the student's representational limitations on a consistent subset of training data -- the Robustly Unlearnable Set. We present a theoretical framework analyzing the feature learning dynamics of a two-layer neural network, demonstrating that this mismatch creates a dichotomy in distillation outcomes. We prove that when a teacher provides confident supervision on unlearnable samples, it compels the student to memorize spurious noise patterns that eventually overpower the learned robust signal, thereby driving robust overfitting. Conversely, a teacher that exhibits high uncertainty on these samples effectively suppresses noise memorization, allowing the student to rely solely on the learnable signal for robust generalization. We empirically validate our theory across both synthetic simulations and real-image classification datasets, confirming that robust overfitting is driven by the teacher's interaction with unlearnable samples. Finally, we demonstrate that a teacher's predictive entropy on unlearnable samples serves as a strong indicator of student robustness, validating our theoretical framework and offering a principled guideline for robust teacher selection.

The practical adoption of sampling-based inference (SAI) in Bayesian neural networks (BNNs) remains limited, partly due to persistent misconceptions about the feasibility and efficiency of sampling. This position paper argues that SAI has achieved computational parity with optimization-based methods and is at the verge of superseding such methods for effective and efficient inference in BNNs. This development should be in the interest of the whole community, promoting BNNs as a principled paradigm with its long-standing yet unfulfilled promise of providing principled uncertainty quantification for neural networks. SAI can even do more -- yielding superior prediction performance through model averaging, serving as the foundation for a plethora of possible downstream tasks, and providing crucial insights into the landscape of BNNs. In order to make such a change happen and unfold the potential of sampling, overcoming current misconceptions is a necessary first step. The next step is to realign research efforts toward addressing remaining challenges in SAI. In particular, the community must focus on two core problems: sufficient exploration of the posterior landscape and high-fidelity distillation of posterior samples for efficient downstream inference. By addressing conceptual and practical obstacles, we can unlock the full potential of SAI and establish it as a central tool in Bayesian deep learning.

Cross-tokenizer knowledge distillation allows a student model to learn from teachers with incompatible vocabularies. Prior work operates on hidden states or logits; the latter is preferred as a drop-in replacement requiring no auxiliary components. Logit-based methods either use only the correct-token probability, missing the full 'dark knowledge' in the teacher's distribution, or operate on the full output distribution, relying on strict token partitioning and/or unprincipled heuristic ranking. We identify two key shortcomings of full-distribution, logit-based methods: (i) an uncommon-token failure, where critical tokens fall into the unmatched subset (e.g., Llama's 1100 multi-digit numerals under digit-splitting Qwen supervision) and are suppressed during training, reducing GSM8k from 12.89 to 2.56 compared to same-tokenizer KD from a weaker teacher; and (ii) over-conservative matching, where strict 1-to-1 matching excludes near-equivalent tokens across surface forms. These failures require distinct remedies: eliminating the partition when critical tokens are misaligned, and refining it when alignment is reliable. We propose X-Token, an approach with two complementary loss formulations targeting these issues. P-KL removes partitioning and aligns the student's distribution with the teacher's via a sparse projection matrix W (initialized from tokenizer-level string rules) to address the uncommon-token failure. H-KL retains the hybrid form while relaxing matching to align each student token with its top-ranked teacher mapping under W. Both objectives share W and extend naturally to multiple teachers. Empirically, on Llama-3.2-1B, X-Token outperforms the current state of the art GOLD by +3.82 average points with a Qwen3-4B teacher and by +0.5 with a Phi-4-Mini teacher. Further, a two-teacher setup (Phi-4-mini + Llama-3B) improves over single-teacher distillation by +1.3 points.

Open-ended image generation is no longer a simple prompt-to-image problem. High-quality generation often requires an agent to combine a model's internal generative ability with external resources. As requests become more diverse and demanding, we aim to develop a general image-generation agent that can self-evolve through trajectories and use tools more effectively across varied generation challenges. To this end, we propose GenEvolve, a self-evolving framework based on Tool-Orchestrated Visual Experience Distillation. In GenEvolve, each generation attempt is modeled as a tool-orchestrated trajectory, where the agent gathers evidence, selects references, invokes generation skills, and composes them into a prompt-reference program. Unlike existing agentic generation methods that mainly rely on image-level scalar rewards, GenEvolve compares multiple trajectories for the same request and abstracts best-worst differences into structured visual experience, provided only to a privileged teacher branch. Inspired by on-policy self-distillation, Visual Experience Distillation provides dense token-level supervision, helping the student internalize better search, knowledge activation, reference selection, and prompt construction. We further construct GenEvolve-Data and GenEvolve-Bench. Experiments on public benchmarks and GenEvolve-Bench show substantial gains over strong baselines, achieving state-of-the-art performance among current image-generation frameworks. Our website is as follows: https://ephemeral182.github.io/GenEvolve/

Federated Learning (FL) enables collaborative model training without centralizing data. However, real-world deployments must simultaneously address statistical heterogeneity across client data (non-IID), system heterogeneity in device capabilities, and communication efficiency. Existing FL approaches mitigate these challenges through improved aggregation, personalization, or knowledge distillation, but they almost universally assume a fixed client architecture, limiting adaptability to heterogeneous data complexity and hardware constraints. This architectural constraint often leads to suboptimal trade-offs between accuracy and efficiency in real-world FL systems. This work introduces FedKDNAS, a distillation-driven FL framework that combines client-side neural architecture selection with distillation of server-coordinated knowledge. Each client autonomously selects a lightweight model under accuracy-resource constraints. It then trains it locally using a hybrid objective combining supervised learning and knowledge distillation and shares only predictions on a public reference set. The server then aggregates and smooths these predictions, optionally combining them with a teacher model, to produce stable distillation targets for the next round. Extensive evaluation on six datasets against six representative FL baselines (FedAvg, Ditto, FedMD, FedDF, FedDistill, Local-KD) demonstrates that FedKDNAS consistently achieves superior Pareto efficiency, improving accuracy by up to 15% under non-IID conditions, reducing client CPU usage by approximately 28%, and decreasing communication overhead by up to 44 times while maintaining lightweight logit-based communication.

X-ray security inspection requires accurate real-time detection of prohibited items, but existing models often struggle to balance the challenges of severe occlusion, complex clutter, and strict speed requirements. To overcome these challenges, this paper proposes GSA-YOLO, a novel lightweight framework built upon the YOLOv8n architecture, specifically engineered to enhance detection robustness and inference efficiency. GSA-YOLO strategically integrates structured sparsity and adaptive knowledge transfer through three core components: Group Lasso (GL) applied to the network neck for robust feature extraction; Sparse Structure Selection (SSS) applied to the detection head for significant model slimming; and an Adaptive Knowledge Distillation (Ada-KD) mechanism for comprehensive accuracy recovery. This integrated approach synergistically enhances feature representation while pruning redundant channels, maximizing model efficiency without sacrificing performance. Rigorous evaluations on the HiXray and PIDray datasets confirm GSA-YOLO's comprehensive capability, achieving a leading inference speed of 189.62 FPS, accompanied by a reduction in computational cost from 8.7G to 8.0G. Crucially, GSA-YOLO secures mAP50:95 results of 0.531 and 0.679 on HiXray and PIDray, demonstrating 2.4% and 1.8% improvements over the baseline, respectively. Compared to other models, GSA-YOLO exhibits enhanced accuracy while maintaining computational efficiency, making it a promising solution for practical X-ray security inspection.

Dataset distillation (DD) compresses a large training set into a small synthetic set for efficient training, but most DD methods optimize only clean accuracy and leave robustness uncontrolled. Recent robust DD methods improve robustness, yet they often suffer from a poor accuracy-robustness trade-off because they (i) treat all adversarially perturbed examples uniformly, despite robust risk being dominated by near-zero robust margins, and (ii) do not explicitly increase inter-class separation in the decision boundary where attacks concentrate. We present Contrastive Curriculum for Robust Dataset Distillation (C$^2$R), a framework that couples an attack-aware curriculum with a contrastive robustness objective. From a robust-margin perspective, we derive a perturbation score that approximates each sample's robust hinge, enabling a curriculum that prioritizes the smallest-margin adversaries that most directly drive robust error. In parallel, a class-balanced contrastive robustness loss enforces adversarial invariance while explicitly widening boundary separation across classes. Experiments on CIFAR-10/100, Tiny-ImageNet, and multiple ImageNet-1K subsets under six attacks show that C$^2$R achieves the best robust accuracy, outperforming prior robust DD by $2.8$% on average.

Knowledge distillation (KD) transfers knowledge from a high-capacity teacher to a compact student by matching their predictive distributions, with temperature scaling serving as a central mechanism for smoothing teacher predictions and exposing informative "dark knowledge" beyond the hard label. However, the standard fixed-temperature design is inherently sample-agnostic. Since samples differ in logit scale and learning difficulty, a single global temperature produces teacher soft labels with highly inconsistent entropy: some predictions remain overly sharp and provide limited inter-class information, whereas others become over-smoothed and lose class-discriminative information. Moreover, sharing the same temperature between teacher and student further imposes rigid logit-scale alignment despite their capacity mismatch. To address these limitations, we propose CIST (Consistently Informative Soft-label Temperature), which assigns separate sample-wise adaptive temperatures to the teacher and student. This design produces consistently informative teacher soft labels while relaxing rigid teacher--student logit-scale matching. It also reweights the distillation objective according to teacher confidence and student learning difficulty. Theoretically, we show that teacher-label entropy is largely governed by the ratio between the maximum teacher logit and the temperature, providing a principled basis for adaptive smoothing. Empirically, CIST mitigates the inconsistency induced by fixed temperature, and experiments on both vision and language distillation tasks show consistent improvements over standard KD and strong baselines with negligible computational overhead.

Distillation transfers knowledge from a large model trained on broad data to a smaller, more efficient model suitable for deployment. In structured prediction settings, prior knowledge about the task can guide the choice of a target architecture that is algorithmically aligned with the underlying problem. Building on recent learning-theoretic analyses of decision-tree (DT) distillation (Boix-Adsera, 2024), we study when distillation succeeds for combinatorial optimization tasks. We focus on the case where the target model is a graph neural network whose architecture is aligned with a dynamic programming (DP) algorithm for the task. Assuming that the source model is sufficiently rich, formalized through the linear representation hypothesis (LRH) (Elhage et al., 2022; Park et al., 2024), we show that the distillation problem can be solved efficiently in the complexity parameters of the DP transition function, represented as a DT. Our results provide a rigorous sufficient condition for successful distillation in the flavour of algorithmic alignment.

We demonstrate that in knowledge distillation for diffusion models, the teacher network's highly complex denoising process - stemming from its substantially larger capacity - poses a significant challenge for the student model to faithfully mimic. To address this problem, we propose a coarse-to-fine distillation framework with LInear FiTtingbased distillation (LIFT) and Piecewise Local Adaptive Coefficient Estimation (PLACE). First, LIFT decomposes the objective into a "coarse" alignment and a "fine" refinement. The student is then trained on coarse alignment before proceeding to hard refinement. Second, PLACE extends LIFT to address spatially non-uniform errors by partitioning outputs into error-based groups, providing locally adaptive guidance. Our experiments show that LIFT and PLACE is effective across diffusion spaces (image/latent), backbones (U-Net/DiT), tasks (unconditional/conditional), datasets, and even extends to flow-based models such as MMDiT (SD3). Furthermore, under extreme compression with a 1.3M-parameter student (only 1.6% of the teacher), conventional KD fails to provide sufficient guidance for stable training, with FID scores often degrading to 50-200+, but our method remains stably convergent and achieves an FID of 15.73.

The exponential growth of big data has intensified the need for efficient and interpretable machine learning models that can handle diverse data characteristics while maintaining computational efficiency. Knowledge distillation has primarily focused on neural network-to-neural network transfer, leaving cross-paradigm knowledge transfer largely unexplored. This paper presents the first comprehensive study of bidirectional knowledge distillation between Random Forests (RF) and Deep Neural Networks (DNN), addressing critical gaps in ensemble learning and model compression for big data applications. We propose novel methodologies including progressive multi-stage distillation, multi-teacher ensemble distillation from diverse tree models, and uncertainty-aware cross-paradigm transfer mechanisms. Through 144 comprehensive experiments across 6 diverse datasets encompassing classification and regression tasks, we demonstrate that bidirectional RF-DL distillation achieves competitive performance while providing complementary benefits: interpretability from tree models and expressiveness from neural networks. Our results show that multi-teacher ensemble distillation consistently outperforms traditional approaches, with NN-COMPACT achieving 98.13% classification accuracy and NN-WIDE reaching 92.6% R^2 score in regression tasks. The proposed framework enables deployment flexibility in big data environments, allowing optimal model selection based on computational constraints and interpretability requirements. This work establishes a new research direction in cross-paradigm knowledge transfer with significant implications for interpretable AI and scalable model deployment in resource-constrained big data systems.

Tabular foundation models (TFMs) achieve strong performance on health datasets, but their inference cost and infrastructure requirements limit practical use. We study whether their predictive behavior can be transferred to lightweight tabular models through knowledge distillation. Since in-context TFMs condition on the training set at inference time, naive distillation can introduce context leakage; we address this with stratified out-of-fold teacher labeling. Across $19$ healthcare datasets, $6$ TFM teachers, $4$ student families, and several multi-teacher ensembles, we find that distilled students retain at least $90\%$ of teacher AUC, outperforming teachers in some cases, while running at least $26\times$ faster on CPU and preserving calibration and fairness critical for health applications. Moreover, multi-teacher averaging does not consistently improve over the best single teacher. Leakage-aware distillation is thus a viable route for bringing TFM-quality predictions into inference-constrained health settings.

Anomaly detection in batch processes is hindered by transient dynamics, scarce fault labels, and reliance on single-modality sensor data. This work introduces UTOPYA (Unified Temporal Observation for Physics-Informed Anomaly Detection and Time-Series Prediction), a 15.2M-parameter multimodal framework that jointly addresses anomaly detection, time-series prediction, and phase classification in batch distillation by fusing eight data modalities through Feature-wise Linear Modulation (FiLM) conditioned cross-modal attention and gated fusion. A physics-informed regularisation scheme introduced in this work enforces temporal smoothness and thermodynamic monotonicity, while curriculum learning introduces training samples in order of physical difficulty. On the 119-experiment multimodal batch distillation dataset of Arweiler et al. (2026), UTOPYA achieves a window-level test AUROC of 0.832 and 0.874 under multi-signal experiment-level scoring, substantially outperforming four external baselines (PCA, autoencoder, Isolation Forest, and LSTM autoencoder) evaluated under identical conditions (+0.147 window-level AUROC over the best baseline). A multimodal ablation over 15~architectural configurations shows that static context via FiLM conditioning is the key enabler, lifting experiment-level multi-signal AUROC by +0.145 over the unimodal baseline (0.729 to 0.874). Separately, a training ablation across 14 design choices reveals that several widely-adopted techniques, including instance normalisation, Mixup, ensembling, test-time augmentation, and stochastic weight averaging, fail to improve or actively degrade generalisation in this data-scarce setting. These negative results expose a fundamental tension between smoothing-based regularisation and anomaly detection, providing practical guidance for multimodal process monitoring deployment.

Deep neural networks have achieved impressive performance across a wide range of tasks, but this success often comes with substantial computational and storage costs due to large-scale training data. Dataset distillation addresses this challenge by constructing compact yet informative datasets that enable efficient model training while maintaining downstream performance. However, most existing approaches primarily emphasize matching data distributions or downstream training statistics, with limited attention to preserving high-level semantic information in the distilled data. In this work, we introduce a semantic-aware perspective for dataset distillation by leveraging Contrastive Language-Image Pretraining (CLIP) as a semantic prior for post-sampling. Our goal is to obtain distilled datasets that are not only compact but also semantically class-discriminative and diverse. To this end, we design three semantic scoring functions that quantify class relevance, inter-class separability, and intra-set diversity in a pretrained semantic space. Based on image pools generated by existing distillation methods, we further develop a two-stage strategy for effective sampling: the first stage filters semantically discriminative samples to form a reliable candidate set, and the second stage performs a dynamic diversity-aware selection to reduce redundancy while preserving semantic coverage. Extensive experiments across multiple datasets, image pools, and downstream models demonstrate consistent performance gains, highlighting the effectiveness of incorporating semantic information into dataset distillation.

Training long-horizon LLM agents with reinforcement learning is challenging because sparse outcome rewards reveal whether a task succeeds, but not which intermediate actions caused the outcome or how they should be corrected. Recent methods alleviate this issue by generating rewards or textual hints from turn-level action-output signals, or by using feedback-conditioned self-distillation. However, generating feedback at every turn is inefficient when many intermediate turns are already successful or neutral, and applying feedback at a fixed or misaligned turn often fails to supervise the actions that contributed to the failure. To bridge this gap, we propose HINT-SD, a targeted self-distillation framework that uses full-trajectory hindsight to select failure-relevant actions and applies feedback-conditioned distillation only on targeted action spans. Experiments on BFCL v3 and AppWorld show that our method improves over the dense per-turn feedback baseline by up to 18.80 percent while achieving 2.26$\times$ lower time per training step, suggesting that selecting where to distill is a key factor for both effective and efficient long-horizon agent training.

Knowledge distillation transfers knowledge from a high capacity teacher to a compact student using a mixture of hard and soft losses. On imbalanced data, a fixed weighting between hard and soft losses becomes brittle the learning process. Recent studies try to reweight these components in long-tailed settings. However, most of these methods do not adapt weights at the sample-wise level and do not take into account the students behavior during training. To address this, we propose BiKD -- a bilevel framework that dynamically balances hard and soft losses for each sample. We employ a weight generation network that produces adaptive per-sample weights, guided by a small balanced validation set. The student is now trained with an unconstrained combination of weighted hard and soft losses, allowing the student to relax both terms. We further propose a multi-step SGD strategy to optimize the weight model more accurately and efficiently. Experiments on long-tailed CIFAR-10/100 show that our approach surpasses recent balanced distillation methods across imbalance factors.

Query optimization in big data analytics remains computationally expensive, particularly for resource-constrained environments where traditional optimizers fail to satisfy memory and latency constraints. We present an agentic query planning system that combines a rule-based teacher planner, UCB1 bandit exploration, cost-aware prediction, and knowledge distillation to a lightweight student planner. Our teacher planner generates SQL plans using six key optimization strategies, while UCB1 bandit search efficiently explores the plan space under explicit resource constraints. A Random Forest cost model predicts query latency from plan features, enabling cost-aware decisions. A distilled student planner (Logistic Regression or Gradient Boosting) learns to mimic teacher-bandit decisions for fast inference. Evaluation on NYC Taxi and IMDB datasets demonstrates 23% latency reduction compared to default planners while maintaining 94% constraint satisfaction. The student planner achieves 89% accuracy in replicating optimal plans with 15x faster inference time. Our single-file implementation enables reproducible big-data analytics on resource-limited machines and is publicly available at https://github.com/mahdinaser/agentic-kd-planner.

Although modern multilingual Automatic Speech Recognition (ASR) systems support several Nigerian languages, their performance consistently lags behind high-resource languages like English and French. Nigerian languages present unique modelling hurdles, including acute data scarcity, inconsistent orthography, tonal diacritics, diverse accents, frequent code-switching, and localized named entities. To address these challenges, we developed a multilingual ASR framework utilizing a two-stage distillation process. First, we employ student-teacher knowledge distillation from existing monolingual models, conditioned on robust language-specific N-gram language models. Second, we perform iterative self improvement using pseudo-labelled data to further refine accuracy. Our method significantly bridges the performance gap, achieving on average a relative Word Error Rate (WER) reduction of 29 % over monolingual baselines. Our models also outperform state-of-the-art multilingual models across major benchmarks, including Common Voice and Fleurs. We introduce Sometin Beta Pass Notin (SBPN), a foundational multilingual ASR model covering Yorùbá, Hausa, Igbo, Nigerian Pidgin, and Nigerian English. SBPN is released in two sizes: SBPN-Base (120 M parameters) and SBPN-Large (600 M parameters). By releasing these as open foundation models, we aim to provide ASR resources for further research into the rich phonetic and cultural landscape of the region.

Open-vocabulary segmentation models such as SAM3 perform well across broad categories via text prompting, yet degrade when target classes are visually underrepresented in pretraining or depart from canonical depictions-limitations text prompts cannot resolve spatially. We present SegRAG, a training-free retrieval-augmented segmentation framework that grounds SAM3 with class-specific point prompts derived from a curated DINOv3 feature bank. Offline, dense patch-level descriptors are extracted from annotated references and filtered by Intra-Class Cohesion Distillation (ICCD), retaining only prototypes that reliably retrieve within-class foreground. At inference, Topographic Similarity Grounding (TSG) computes a cosine-similarity landscape against retrieved prototypes, identifies coherent high-confidence regions via connected-component analysis, and extracts peak locations through non-maximum suppression. The resulting point prompts are delivered jointly with class-name text in a single SAM3 forward pass. On four standard benchmarks, SegRAG consistently outperforms the text-only baseline, gaining up to +3.92 mIoU on LVIS. On AgML agricultural benchmarks under zero-shot domain transfer, it raises mean IoU from 25.27 to 59.24 (+33.97) and recovers individual classes from zero to over 95 mIoU. Ablations confirm that ICCD, TSG, and joint prompting each contribute independently and compound when combined. Code is available at (https://github.com/boudiafA/SegRAG).

Distilling vision-language models into faster hybrid architectures, such as 3:1 Mamba-2/attention mixes, is now standard practice for making inference efficient. Aggregate benchmarks suggest that this works but they hide selective failures. When we distill Qwen3-VL-8B-Instruct into a 3:1 Mamba-2/attention hybrid, student model stays within 2 points of the teacher across visual reasoning benchmarks like MMStar, MMBench, and MMMU-Pro, while dropping 13 points on optical-character-recognition and document tasks. The student can still understand the scene but loses the fine-grained text needed to answer. We localize much of the failure to a specific kind of position. In a high-resolution image, most patches are sky, wall, or smooth texture, while a small fraction carries text, edges, object boundaries, or other local details. In a token-level diagnostic, the top 10% highest-density patches have 3.6$\times$ larger residual drift than the bottom 10% lowest-density patches and 3.5$\times$ larger teacher-masking answer contribution. Uniform weighting devotes many loss terms to low-information background patches, whereas sparse answer-bearing patches receive no special protection. The required intervention is minimal: we replace uniform residual alignment with density-weighted residual alignment, using patch self-dissimilarity as a training-free proxy for position importance. We call this HEED. Compared with normal end-to-end distillation, HEED increases performance by 8.7 points on OCRBench v2 and 5.13 points on a 10-benchmark average. The gain is realized on different teacher models and hybrid architectures. After standard post-training, the student reaches teacher-level performance on the 10-benchmark average with a 4.12$\times$ throughput and a 68% memory saving at 128k context, with no additional parameters and no inference-time cost.

Supervised fine-tuning (SFT) is widely used to inject new knowledge into language models, but it often degrades pretrained capabilities such as reasoning and general-domain performance. We argue this forgetting arises because fine-tuning targets from humans or external systems diverge from the model's autoregressive distribution, forcing the optimizer to imitate low-probability token sequences. To address this problem, we propose MixSD, a simple external-teacher-free method for distribution-aligned knowledge injection. Instead of training on fixed targets, MixSD constructs supervision dynamically by mixing tokens from two conditionals of the base model itself: an expert conditional that observes the injected fact in context, and a naive conditional that reflects the model's original prior. The resulting supervision sequences preserve the factual learning signal while remaining substantially closer to the base model's distribution. We evaluate MixSD on two synthetic corpora that we construct to study factual recall and arithmetic function acquisition in a controlled setting, together with established benchmarks for open-domain factual question answering and knowledge editing. Across multiple model scales and settings, MixSD consistently achieves a better memorization-retention trade-off compared to SFT and on-policy self distillation baselines, retaining up to 100% of the base model's held-out capability while maintaining near-perfect training accuracy, whereas standard SFT retains as little as 1%. We further show that MixSD produces substantially lower-NLL supervision targets under the base model and reduces harmful movement along Fisher-sensitive parameter directions. These results suggest that aligning supervision with the model's native generation distribution is a simple and effective principle for knowledge injection that mitigates catastrophic forgetting.

Knowledge distillation is central to LLM post-training, yet its design space remains poorly understood, especially alongside reinforcement learning (RL). We show that the prevailing paradigms, off-policy distillation and on-policy distillation (OPD), implicitly couple two orthogonal choices: prefix source and token-level KL direction. This follows from decomposing sequence-level KL over autoregressive response distributions: forward KL pairs teacher prefixes with token-level forward KL, and reverse KL pairs student prefixes with token-level reverse KL. We argue this coupling is not intrinsic: decoupling the two axes yields four valid objectives. We establish gradient-level identities showing forward KL gives SFT-style cross-entropy matching with teacher soft targets, whereas reverse KL gives an RL-style policy-gradient objective with a dense teacher-student log-ratio reward, connecting them to off-policy SFT, DAgger-style on-policy SFT, offline-RL-style distillation, and OPD. We conduct an extensive controlled study on math reasoning, evaluating the four objectives both as standalone methods and as initializations for subsequent RL. The results reveal three tradeoffs: KL direction induces an accuracy-entropy tradeoff, prefix source a quality-compute tradeoff, and training length an accuracy-stability tradeoff. Motivated by these findings, we propose KL mixing and an entropy-gated length curriculum. KL mixing shows long-sequence distillation requires substantial forward-KL weight to prevent entropy collapse and length inflation without sacrificing accuracy. The entropy-gated length curriculum improves Avg@k and Pass@k by 3.6 and up to 5.8 points, and cuts average response length by roughly 3x versus fixed long-horizon training. Our results provide a framework and practical methods for designing reasoning distillation objectives that balance accuracy, diversity, compute, and RL behavior.

Detecting unauthorized knowledge distillation from a deployed LLM API is hard because the defender controls neither the attacker's training pipeline nor the next-token logits. Existing defenses operate on the teacher's output tokens -- biasing the next-token distribution (green-list watermarks, cryptographic schemes, antidistillation sampling) or rewriting outputs after generation. Recent work shows a paraphrasing attacker can strip these signals without losing the underlying knowledge. We propose interaction-layer antidistillation watermarks, which move the trace one layer higher, into the teacher's interaction behavior: the defender wraps the teacher with a system prompt that intermittently induces a behavioral marker -- an explicit follow-up question, a low-frequency variant, or a declarative restatement. An oblivious distiller inherits the behavior, and the defender audits via black-box queries with a human-validated LLM-as-judge (Cohen's kappa = 0.84/0.78 on strong/style rubrics). Across 63 LoRA-distilled students under a Llama-3.3-70B-Instruct teacher (35,343 judged samples), behavioral watermarks transfer at 88.9% (Gemma) / 80.9% (OLMo) / 45.2% (Qwen) relative fidelity (H1, H2). Under non-adaptive DIPPER paraphrasing, robustness decomposes into a teacher-self ceiling (about 66.4%) and student-relative retention of 21-112%, with OLMo preserving the watermark above the teacher itself (H3, F-Amp). Low-density (about 20%) explicit and implicit declarative variants transfer above per-family baseline (H4, F-Style). An N=20 in-lab study (pre-registered Latin-square) shows all marker variants within 0.22 Likert step of baseline; TOST, Friedman, and Bonferroni-Wilcoxon support H5. The interaction layer is a viable design locus for antidistillation watermarking, complementary to token-, model-, and reasoning-trace-layer defenses.

Fine-grained image recognition classifies subcategories such as bird species or car models. While state-of-the-art (SOTA) models are accurate, they are often too resource-intensive for deployment on constrained devices. Knowledge distillation addresses this by transferring knowledge from a large teacher model to a smaller student model. A key challenge is selecting the right teacher, as it heavily impacts student performance. This paper introduces a teacher selection metric, \textbf{Ratio 1-2}, based on teacher prediction ratios. Extensive analysis of over one thousand experiments across 3 students, 8 teachers, and 8 datasets under 4 training strategies demonstrates that our metric improves teacher selection by 18% over previous methods, enabling small student models to achieve up to 17% accuracy gains. Experiment codebase is available at: \href{https://github.com/arkel23/FGIR-KD-Teacher}{https://github.com/arkel23/FGIR-KD-Teacher}.

Distillation enables compact Vision-Language Models (VLMs) to obtain strong reasoning capabilities, yet the prompts driving this process are typically chosen via simple heuristics or aggregated from off-the-shelf datasets. We reveal a critical inefficiency in this approach: up to 69% of the prompts in standard chart / document reasoning datasets are effectively zero-delta, meaning the teacher and student already induce the exact same answer distribution. Training on these prompts provides minimal learning signal, causing student improvement to rapidly saturate regardless of data scale. To escape the zero-delta trap, we return to first principles: distillation fundamentally minimizes distributional divergence, and thus a prompt is valuable only if it exposes a functional capability gap between the teacher and student. We quantify this gap through answer divergence ($Δ$), demonstrating that non-zero divergence is critical for effective scaling. Building on this insight, we propose a staged synthesis pipeline that repurposes existing datasets as seeds, actively targeting student failure modes to produce better prompts. The result is DeltaPrompts, a diverse dataset of 200k synthetic, high-divergence reasoning problems. We evaluate DeltaPrompts across three distinct settings: on-policy distillation with the target teacher-student pair, transfer to a novel model family without regenerating the data, and off-policy fine-tuning of a non-reasoning model. Across all scenarios, DeltaPrompts drives substantial gains, yielding up to 15% relative improvement even on top of a highly-optimized reasoning model (e.g., Qwen3-VL-8B-Thinking) -- averaged over 10 benchmarks spanning chart, document and perception-centric reasoning.

Reinforcement learning from verifiable rewards (RLVR) suffers from sparse outcome signals, creating severe exploration bottlenecks on complex reasoning tasks. Recent on-policy self-distillation methods attempt to address this by utilizing language feedback to generate dense, token-level supervision. However, these approaches rely on a fixed, passive teacher to interpret the feedback. As the student policy improves, the teacher's zero-shot assessment capabilities plateau, ultimately halting further learning. To overcome this, we propose Variational Policy Distillation (VPD), a framework that formalizes learning from language feedback as a Variational Expectation-Maximization (EM) problem. VPD co-evolves both policies: in the E-step, the teacher is actively refined on trajectory outcomes via an adaptive trust-region update, translating textual feedback into a dynamically improved target token distribution. In the M-step, the student internalizes this dense distributional guidance on its own on-policy rollouts. By continuously improving the teacher's ability to extract actionable signals from textual critique, VPD overcomes the limitations of passive distillation. Evaluated across diverse sources of diagnostic feedback on scientific reasoning and code generation tasks, VPD consistently outperforms both standard RLVR and existing self-distillation baselines. Finally, by stress-testing our framework on rigid mathematical reasoning and cold-start regimes, we illuminate the fundamental bounds of feedback-driven self-distillation compared to pure environment-driven RL.

Electrocardiogram (ECG) monitoring in Internet of Medical Things (IoMT) networks is constrained by strict data-sharing regulations and privacy concerns. Federated learning (FL) enables collaborative learning by keeping raw ECG data on devices, but frequent transmissions of high-dimensional model updates incur heavy per-round traffic over bandwidth-limited links. To alleviate this bottleneck, federated distillation (FD) replaces parameter exchange with logit-based knowledge transfer. However, the performance of FD often degrades under the non-independent and identically distributed (non-IID) and long-tailed label distributions in ECG deployments. To address these challenges, we propose a bidirectional federated knowledge distillation (BiFedKD) framework that employs an aggregation-by-distillation pipeline with temperature scaling to produce a stable global distillation signal for cross-client alignment. Experiments on the MIT-BIH Arrhythmia dataset show that BiFedKD improves accuracy and Macro-F1 over the baseline by $3.52\%$ and $9.93\%$, respectively. Moreover, to reach the same Macro-F1, BiFedKD reduces communication overhead by $40\%$ and computation cost by $71.7\%$ compared with the baseline.

Accurately identifying student misconceptions is crucial for personalized education but faces three challenges: (1) data scarcity with long-tail distribution, where authentic student reasoning is difficult to synthesize; (2) fuzzy boundaries between error categories with high annotation noise; (3) deployment parado-large models overlook unconventional approaches due to pretraining bias and cannot be deployed on edge, while small models overfit to noise. Unlike traditional methods that increase diversity through large-scale data synthesis, we propose a two-stage knowledge distillation framework that mines high-value samples from existing data. The first stage performs standard distillation to transfer task capabilities. The second stage introduces a dual-layer marginal selection mechanism based on cognitive uncertainty, identifying four types of critical samples based on teacher model uncertainty and confidence differences. For different data subsets, we design difficulty-adaptive mechanism to balance hard/soft label contributions, enabling student models to inherit inter-class relationships from teacher soft labels while distinguishing ambiguous error types. Experiments show that with augmented training on only 10.30% of filtered samples, we achieve MAP@3 of 0.9585 (+17.8%) on the MAP-Charting dataset, and using only a 4B parameter model, we attain 84.38% accuracy on cross-topic tests of middle school algebra misconception benchmarks, significantly outperforming sota LLM (67.73%) and standard fine-tuned 72B models (81.25%). Our code is available at https://github.com/RoschildRui/acl2026_map.

Recent work has promoted task-aware layer pruning as a way to improve model performance on particular tasks, as shown by TALE. In this paper, we investigate when such improvements occur and why. We show first that, across controlled polynomial regression tasks and large language models, such pruning yields no benefit on in-distribution (ID) data but consistently improves out-of-distribution (OOD) accuracy. We further show empirically that OOD inputs induce layerwise norm and pairwise-distance profiles that deviate from the corresponding ID profiles. This leads to a geometric explanation of task-aware pruning: each task induces a task-adapted geometry, characterized empirically by the representation profiles observed on ID inputs. OOD inputs can introduce a distorted version of the task-adapted geometry. Task-aware pruning identifies layers that create or amplify this distortion; by removing them, it shifts OOD representational norms and pairwise distances toward those observed on the adapted distribution. This realigns OOD inputs with the model's task-adapted geometry and improves performance. We provide causal evidence through controlled distribution shifts and residual-scaling interventions, and demonstrate consistent behavior across model scales.

Single-frame Infrared Small Target Detection (ISTD) aims to localize weak targets under heavy background clutter, yet dense pixel-wise annotations are expensive. Point supervision with online label evolution reduces annotation cost; however, lightweight CNN detectors often lack sufficient semantics, leading to noisy pseudo-masks and unstable optimization. To address this, we propose a hierarchical VFM-driven knowledge distillation framework that uses a frozen Vision Foundation Model (VFM) during training. We formulate point-supervised learning as a bilevel optimization process: the inner loop adapts a VFM-embedded teacher on reweighted training samples, while the outer loop transfers validation-guided knowledge to a lightweight student to mitigate pseudo-label noise and training-set bias. We further introduce Semantic-Conditioned Affine Modulation (SCAM) to inject VFM semantics into CNN features at multiple layers. In addition, a dynamic collaborative learning strategy with cluster-level sample reweighting enhances robustness to imperfect pseudo-masks. Experiments on diverse challenging cases across multiple ISTD backbones demonstrate consistent improvements in detection accuracy and training stability. Our code is available at https://github.com/yuanhang-yao/semantic-prior.

Spiking neural networks (SNNs), which are brain-inspired and spike-driven, achieve high energy efficiency. However, a performance gap between SNNs and artificial neural networks (ANNs) still remains. Knowledge distillation (KD) is commonly adopted to improve SNN performance, but existing methods typically enforce uniform alignment across all timesteps, either from a teacher network or through inter-temporal self-distillation, implicitly assuming that per-timestep predictions should be treated equally. In practice, SNN predictions vary and evolve over time, and intermediate timesteps need not all be individually correct even when the final aggregated output is correct. Under such conditions, effective distillation should not force every timestep toward the same supervision target, but instead provide corrective guidance to erroneous timesteps while preserving useful temporal dynamics. To address this issue, we propose Selective Alignment Knowledge Distillation (SeAl-KD), which selectively aligns class-level and temporal knowledge by equalizing competing logits at erroneous timesteps and reweighting temporal alignment based on confidence and inter-timestep similarity. Extensive experiments on static image and neuromorphic event-based datasets demonstrate consistent improvements over existing distillation methods. The code is available at https://github.com/KaiSUN1/SeAl

The rise in deployment of large language models has driven a surge in GPU demand and datacenter scaling, raising concerns about electricity use, grid stress, and the impacts of modern AI workloads. Distillation is often promoted as one of the most effective paths to obtain cheaper, more efficient models, yet these claims rarely account for the full end-to-end energy and resource costs, including crucial teacher-side workloads such as data generation, logit caching, and evaluation. We present a comprehensive energy accounting framework that measures the complete computational cost of distillation pipelines via detailed stage-wise tracking of GPU device power consumption. In our experiments, we separate and log empirical energy use across distinct phases and systematically measure the energy and emissions of two common distillation methods: the classic logit-based knowledge distillation and synthetic-data supervised fine-tuning, constructing energy-quality Pareto frontiers that expose the previously ignored costs. From these measurements and analyses, we derive practical design rules for selecting distillation methods and hyperparameters under energy and budget constraints, and release an open-source measurement harness and accounting protocol to provide a standardized foundation for comparable, reproducible distillation research, explicitly accountable for complete pipeline energy impact.

Long chain-of-thought (Long CoT) reasoning improves performance on multi-step problems, but it also induces overthinking. This inefficiency is especially problematic in low-data fine-tuning regimes, where real applications adapt reasoning models with limited supervision and cannot rely on large-scale teacher distillation or heavy test-time control. To address this, we propose STOP (Structured On-policy Pruning), an on-policy algorithm for analyzing and pruning long-form reasoning traces. STOP constructs self-distilled traces from the model. Then it maps each trace into a structured reasoning interface through node segmentation, taxonomy annotation, and reasoning-tree construction. On top of this interface, we introduce ECN (Earliest Correct Node), which retains the shortest prefix ending at the earliest node. Experiments on DeepSeek-R1-Distill-Qwen-7B and DeepSeek-R1-Distill-LLaMA-3-8B across GSM8K, Math 500, and AIME 2024 show that STOP reduces generated tokens by 19.4% to 42.4% while largely preserving accuracy in low-data fine-tuning. Beyond efficiency, our analyses show that STOP induces much smaller distributional shift than teacher-guided pruning, improves the structural efficiency of generated reasoning, and reallocates reasoning effort away from redundant verification and backtracking toward more productive exploration.

Knowledge distillation is widely used to improve generalization in practice, yet its theoretical understanding remains elusive. In the standard distillation setting, a teacher model provides soft predictions to guide the training of a student model. We model teacher and student training as coupled stochastic processes and introduce a distillation divergence, defined as the Kullback-Leibler divergence between these two stochastic kernels. Within this framework, we derive two generalization bounds for the student model relative to the teacher's generalization gap: an upper bound under a sub-Gaussian assumption via algorithmic stability, and a lower bound under a central condition with sharper dependence on the distillation divergence. We further develop a loss-sharpness-aware bound with an explicit tightness regime, showing that the teacher's local flatness can strictly tighten the bound. Additionally, in a linear Gaussian case study, the distillation divergence admits an interpretable decomposition into bias, variance, and rank-bottleneck costs, yielding practical guidance for distillation design.

Distilling multi-step reasoning abilities from large language models (LLMs) into compact student models remains challenging due to noisy rationales, hallucinated supervision, and static teacher-student interactions. Existing reasoning distillation methods, including mentor-based approaches, predominantly operate in an open-loop manner, implicitly assuming uniform teacher reliability and consequently propagating erroneous intermediate reasoning. We propose GateKD, a confidence-gated closed-loop distillation framework that enables robust reasoning transfer by treating the teacher as a dynamic gatekeeper rather than a static oracle. GateKD introduces three complementary mechanisms: (i) confidence-gated soft supervision that selectively distills reliable predictive signals, (ii) gated hidden-state evolution that aligns intermediate representations only when teacher confidence is high, and (iii) reliability-filtered attention distillation that preserves stable reasoning structures while suppressing noisy patterns. These components jointly form a closed feedback loop in which teacher confidence continuously modulates the distillation process, reducing hallucination transfer and stabilizing student reasoning. Extensive experiments across commonsense, logical, and symbolic reasoning benchmarks, using T5 and Flan-T5 backbones of varying sizes, demonstrate that GateKD consistently outperforms strong open-loop distillation baselines. Notably, GateKD yields substantial gains in logical and symbolic reasoning, remains robust under low-resource distillation settings, and shows clear performance degradation when any gating component is removed. Our results highlight that confidence-gated closed-loop supervision is critical for building reliable and scalable small reasoning models.