Overview

Knowledge distillation studies how a student model can learn from a teacher’s outputs, internal representations, examples, or behavior, often to reduce inference cost or transfer capabilities. Common approaches match softened class probabilities, hidden features, attention or relational structure, and use teacher-generated data; newer language-model methods also train on student-generated rollouts. The field spans model compression, cross-architecture and multimodal transfer, and settings such as federated learning, robotics, and low-resource tasks. A central challenge is choosing supervision that fits the student and remains reliable: stronger or denser teacher signals can help, but can also cause negative transfer, forgetting, or loss of capabilities not captured by the benchmark.

Latest papers

887 papers87 papers in the last four weeks, down 49% on the four weeks before. 0.7% of all new papers.

Oct 27, 2026

You Don't Need Aligned Representations: Knowledge Distillation via Random Prototype Spaces

A widely held assumption in knowledge distillation is that teacher and student representations must occupy a compatible or explicitly aligned feature space. We propose RPKD (Random Prototype Knowledge Distillation), which discards this assumption by comparing both networks against a shared bank of randomly initialized, frozen prototype vectors. Alignment is therefore carried out in an external reference space rather than in the teacher's native feature coordinates: the teacher side is projected by fixed random maps that carry no trainable parameters, and only the student adapts. Two branches handle the transfer: a logit branch aligning global prototype similarity distributions and a feature branch matching spatially-resolved prototype activation maps. Both operate within the same fixed prototype space, which can be interpreted as a random feature embedding that approximately preserves inter-sample similarity structure, so the prototype bank and the distillation objective stay unchanged across teacher-student pairs while only the input projections are instantiated per architecture. In low-category settings, decoupling the prototype vocabulary from the task label space yields richer supervision than the label space alone, an advantage absent from logit-based methods, whose supervisory signal collapses as class count shrinks. This advantage is especially relevant in real-world scenarios where the label space is typically constrained. Extensive experiments demonstrate the effectiveness of RPKD, achieving a maximum gain of +7.98% over OFA on CIFAR-100, +12.44% on ImageNet-100, and +1.28% on ImageNet-1K. Source code is available at https://github.com/HamzaEtcibasi/RPKD
Oct 27, 2026

What Survives Distillation? A Mechanistic Interpretability Study of GPT-2 and DistilGPT-2

Knowledge distillation compresses large language models into compact students that preserve aggregate metrics, yet the mechanistic consequences of depth-reduction distillation remain largely uncharacterized. This work applies linear CKA, logit lens analysis, attention pattern decomposition, sparse autoencoder feature statistics, and activation patching to compare GPT-2 and its six-layer student DistilGPT-2 across two canonical circuits: the Indirect Object Identification (IOI) circuit and the Greater-Than (GT) circuit. The results reveal a compression-collapse dichotomy: the IOI circuit suffers complete mechanistic collapse (DistilGPT-2: 35.2% accuracy vs. 94.8%, zero name-mover preservation), while the GT circuit survives with 87.7% task performance recovery through active remapping of the teacher's layer-9 comparison head into student layer~4 (mean Pearson r = 0.945). These findings demonstrate that circuit compressibility governs mechanistic survival under distillation, and that aggregate metrics are insufficient diagnostics for circuit integrity.
Oct 27, 2026

Ground-to-Satellite Knowledge Distillation for Crop Classification under Temporal and Geographic Shift

Continuous and up-to-date crop type mapping is essential for agricultural monitoring but is often hampered by limited reference data. The growing availability of ground-level field photographs offers a promising way to address this challenge. When combined with satellite time series, these images provide complementary perspectives on the same parcel, capturing crop characteristics from distinctly different perspectives. However, field photographs are not always available for each satellite pixel at inference time, limiting their potential in scalable crop type mapping. We investigate whether ground photographs can provide privileged supervision for crop classification built on pretrained satellite embeddings. A photo teacher based on frozen CLIP features supervises classifiers trained on AlphaEarth and TESSERA representations; photographs are used only during training. On a paired field photo and satellite imagery dataset with 17 crop classes, knowledge distillation consistently improves over satellite-only classification; for frozen embeddings the margin grows under temporal shift, reaching up to +5.5 macro-F1 points, and widens in geographic holdouts (+3.7 vs +3.0 in-distribution). These results show that ground-view supervision can complement satellite embeddings and improve their robustness without changing inference requirements.
Oct 27, 2026

Dynamic Merging Mutually-Distilled Multi-View Networks for RGB-Event-based Representation Learning

Multi-modal fusion algorithms combining conventional frame-based cameras and event cameras have recently attracted substantial research attention. Typically, existing approaches convert event streams into frames, voxels, or tensors, and then learn the required spatiotemporal information from these transformed event representations for fusion with conventional frame-camera data. Obviously, such paradigms only exploit the limited representational capacity of event data. Moreover, standard training pipelines fail to fully leverage diverse model parameters obtained throughout the learning process. In this paper, we propose a novel mutual distillation method leveraging multi-view teacher network weights to better integrate multiple event representations and achieve model-parameter level fusion. Specifically, we employ multiple encoders to extract different types of event representations and perform mutual knowledge distillation at both the feature-level and logit-level in this process. Afterwards, we perform a dynamic hierarchical model merge strategy on the model parameters from the first stage to yield a unified multi-modal RGB-Event representation model. Experimental results across multiple event-based datasets demonstrate that our proposed framework achieves significant performance improvements.
Oct 26, 2026

How Much Does the Hint Matter? Scope Conditions for Hindsight Self-Distillation in Financial Reasoning

Financial institutions operating under data-residency constraints are restricted to open-weight models small enough to locally host, which trail frontier systems on multi-step quantitative reasoning. On-policy distillation narrows such gaps by supervising a student on its own rollouts, but traditional OPD requires a strong external teacher, and on-policy self-distillation (OPSD) conditions on a full reference solution. We ask how much of that signal is necessary. We present Fin-SD, a localized hindsight self-distillation recipe for multi-step reasoning. Given a problem the student fails, the student itself identifies the first incorrect reasoning step and writes a short critique of that step. We truncate the student's trajectory immediately before the error (aiming to obtain a correct prefix), condition a frozen copy of the student on the same prefix plus the hint, and distill that hint-conditioned teacher into the student over a short rollout under reverse KL. No external model contributes a gradient. We apply Fin-SD to GPT-OSS-20B on a custom multi-step financial reasoning benchmark that requires deriving formula inputs from realistic scenarios via accounting conventions. We find: (i) replacing the critique with a generic-feedback control matches Fin-SD's entire 3.03 pp accuracy gain over the base model on the FinanceReasoning Hard subset, suggesting the gain comes from where supervision is applied, not hint content; (ii) the model writing the hint does not need to be stronger than the model receiving it: hints from same-family GPT-OSS-120B, at roughly 6x the parameters, yield no materially greater benefits than self-written hints; and (iii) Fin-SD matches on-policy self-distillation from full reference solutions on accuracy while generating 9.5% fewer tokens. With Fin-SD, general reasoning on MMLU-Pro and a fourteen-metric compliance audit on FinTrust show no material regression.
Oct 26, 2026

Joint Latent Reasoning and Observation for Agentic Video Understanding

Effective video agents must learn not only how to answer, but also where to look next. Transferring this capability to smaller models is challenging because each observation serves two purposes: supporting an answer and guiding the search for further evidence. Textual descriptions may capture what is currently visible while omitting cues needed for subsequent acquisition. We introduce Latent Agentic Reasoning for Video Analysis (LARVA), a framework that distills multi-round video agents through jointly learned latent reasoning and visual observations. Rather than communicating through captions, an observer passes compact continuous representations to a reasoner that selects the next acquisition action or produces the final answer. Crucially, action and answer supervision propagate through this interface, training the observer to communicate information useful for downstream decisions. Observation reconstruction and reasoning-distribution alignment provide auxiliary supervision from textual teacher trajectories. Across five video-understanding benchmarks, LARVA achieves strong performance with 4B and 7B backbones. On VideoHolmes, Video-MME, and LVBench, the 7B model outperforms a textual CoT student trained on the same teacher trajectories by 5.1, 2.5, and 3.7 percentage points, respectively. These results highlight the value of jointly learning how agents represent evidence and act on it when distilling video-understanding capabilities.
Oct 26, 2026

MuSID-KD: Multigranularity Knowledge Distillation for Generative Recommendation with Semantic IDs

Generative recommendation with semantic IDs (SIDs) represents each item as a sequence of discrete tokens and autoregressively predicts the next item's SID. Despite strong recommendation performance, repeated decoding over multiple SID positions incurs substantial inference cost. Knowledge distillation transfers knowledge from a powerful generative teacher to a lightweight student, yet conventional objectives do not fully account for multi-token SID generation. We identify two limitations: item-level distribution alignment may not adequately preserve the teacher's ranking of candidate groups sharing SID prefixes, while uniform alignment across SID positions may affect final item ranking differently. We propose MuSID-KD, which combines Hierarchical Candidate Distribution Alignment (HCDA) and Ranking Impact Guided Levelwise Distillation (RIGD). HCDA aggregates item probabilities according to shared SID prefixes and aligns candidate distributions at successive levels, from candidate groups to individual items. RIGD evaluates the reduction in pairwise ranking surrogate loss when one student score contribution is replaced with the teacher's contribution, and allocates supervision accordingly. Experiments on real-world recommendation datasets using three lightweight student architectures demonstrate improved recommendation accuracy and better alignment of prefix marginal distributions and pairwise item rankings, without modifying the student's online inference architecture. Code is available at https://anonymous.4open.science/r/m7k2-review-83f9/.
Oct 26, 2026

DAC-Loop: Distillation-Aware Compression for Remote Sensing CNN

While large-scale models achieve strong results in remote sensing scene classification, deploying them on edge devices remains difficult due to limited compute, memory, and energy budgets. Compact CNN backbones are therefore a practical choice in many real-world settings, offering efficient inference with well-optimized implementations on embedded hardware. To push this efficiency further, we investigate how far a knowledge-distilled student model can be compressed while preserving high accuracy. We propose a Distillation-Aware Compression Loop (DAC-Loop) that integrates (i) knowledge distillation applied both before and after structured channel pruning, and (ii) an iterative compression schedule that alternates structured pruning with unstructured sparsification. Across EuroSAT and NWPU-RESISC45, the DAC-Loop consistently improves the accuracy-compression trade-off relative to standard one-shot pruning and single-stage KD recovery, maintaining strong accuracy even beyond 90% parameter reduction. These results demonstrate that combining distillation and pruning in a unified iterative loop offers an effective path toward compact remote sensing models.
Oct 26, 2026

RSKD: Grounding Semantic Consistency in Driving VLMs via Distillation

Knowledge distillation can transfer large multimodal models to compact models for autonomous driving, but uniform token-level matching can dilute planning-relevant semantics and amplify teacher errors. We propose RSKD, a semantic-consistency-guided distillation framework that routes supervision according to both the task importance of an answer phrase and its semantic consistency with the teacher prediction. Task-critical semantic phrases are localized in the ground-truth answer, and a phrase-level consistency score controls the balance between Jensen--Shannon divergence (JSD) distillation and ground-truth supervision. Teacher information is emphasized only where it is semantically consistent, while semantically inconsistent phrases receive stronger ground-truth supervision. We evaluate RSKD on DrivingVQA, LingoQA, and MapLMv2. RSKD improves over the compared student-based baselines on DrivingVQA and LingoQA and remains competitive on MapLMv2.
Oct 25, 2026

Knowledge distillation across teacher architectures for compact diabetic retinopathy grading

Diabetic retinopathy grading from retinal fundus photographs is important but difficult to deploy on resource-constrained hardware. We conducted a controlled study of offline response-based knowledge distillation for five-class diabetic retinopathy grading on APTOS 2019, comparing ResNet-50, MobileNetV3-Large, and Swin Transformer-Tiny (Swin-T) teachers distilled into a MobileNetV3-Small student. A 3 x 4 x 2 grid of teacher, temperature, and hard-label weight was screened at one seed, while six configurations and the no-distillation baseline were evaluated across three seeds. The highest observed single-seed quadratic weighted kappa was 0.9093 (macro-F1 0.6851) for Swin-T at temperature 4 and hard-label weight 0.7; the corresponding three-seed mean was 0.9024, compared with 0.8804 for the baseline. The student used 1.52 M parameters and 5.94 MB, representing an 18.07x parameter reduction relative to Swin-T. Temperature and hard-label weight effects were teacher- and configuration-dependent, while small performance differences were sensitive to random seed. Knowledge distillation produced a compact student with competitive performance under the evaluated protocol, but independent external validation remains necessary.
Oct 24, 2026

TAD: Distilling Vision-Language-Action Policies from Teacher Trajectories Alone

Online distillation of vision-language-action (VLA) policies commonly relies on teacher action scores that an inference-only interface may not expose. Trajectory-only Adversarial Distillation (\gad{}) instead turns a fixed bank of teacher joint-position features into adaptive feedback: a discriminator compares feature marginals from teacher and student rollouts and supplies rewards for on-policy learning. The student interacts with the environment without online teacher queries or environment task rewards. On ManiSkill PutOnPlateInScene25Main, \gad{} improves OpenVLA-OFT (7B) from 26.2%26.2\% to 76.6%76.6\% \successonce{} (+50.4+50.4 percentage points) and from 20.3%20.3\% to 58.6%58.6\% \successend{}. It trails white-box on-policy distillation (OPD) at 81.6%81.6\% / 59.4%59.4\% by 5.05.0 / 0.80.8 percentage points (n=256n{=}256 evaluation trajectories per checkpoint). The tested four-epoch continuation achieves the highest \gad{} result without additional teacher data. This best run uses 5×5\times OPD's student interaction budget and 12.5×12.5\times its complete-rollout batch-passes. The comparison demonstrates online VLA improvement without teacher scoring while quantifying unequal interaction and optimization budgets.
Oct 23, 2026

Knowledge Distillation across Item and Code Spaces for Parallel Generative Recommendation

Generative recommendation with semantic IDs represents each item as a sequence of discrete codes derived from its semantic representation and predicts these codes to generate recommendations. Existing semantic-ID recommenders are typically trained with hard labels for the observed next item. Although such supervision identifies the target, it provides no explicit distributional information about alternative items conditioned on the same user history. Sequential recommenders can supply this item-level predictive knowledge, but transferring it to semantic-ID generation is challenging because the teacher and student operate in item and code spaces, respectively. We propose two complementary knowledge-distillation objectives for parallel semantic-ID recommenders, which predict all code positions simultaneously without autoregressive prefix conditioning. Item Distribution Matching (IDM) reconstructs a catalogue-level item distribution from the student's parallel code predictions and aligns it with the teacher's item distribution. Code Marginal Matching (CMM) projects the teacher's item probabilities onto each code position using fixed item-to-code assignments, then matches the resulting code marginals to the corresponding student heads. Both objectives transfer the teacher's soft distributional knowledge while preserving the student's parallel prediction structure. Experiments on benchmark datasets show that both methods consistently outperform existing state-of-the-art approaches. On average, IDM improves Recall@10 by 15.39%, while CMM improves NDCG@10 by 8.28%. These results demonstrate that distillation in either reconstructed item space or code-marginal space can substantially strengthen hard-label supervision for parallel generative recommendation.
Oct 23, 2026

Yalda-Embedding: Open Development of a Competitive Persian Retrieval Model

Despite advances in multilingual text representations, effective dense retrieval for Persian remains constrained by the limited availability of language-specific training resources and compact embedding models. We present Yalda Embedding, a 595.8-million-parameter Persian retrieval model adapted from Qwen3-Embedding-0.6B using more than one million query-document pairs. On eight Persian retrieval tasks from FaMTEB, the model achieves the highest macro-average among the evaluated 0.6B models, scoring 0.5488 and outperforming Qwen3-Embedding-0.6B and F2LLM-v2-0.6B on every task. It also approaches Qwen3-Embedding-4B, winning on four tasks, tying on one, and trailing by 0.0075 in macro-average. An incremental ablation study examines contrastive batch size, optimization stabilization, and relational knowledge distillation from F2LLM-v2-8B. Under the one-epoch training setup, contrastive fine-tuning alone yields a modest mean improvement over initialization and regressions on two tasks. The final distilled configuration improves all eight tasks over the non-distilled model, raising the mean from 0.4813 to 0.5488 without increasing inference-time model size. These gains include hard-negative evaluation tasks despite training without explicitly mined or supplied hard negatives. We release model weights, training and evaluation code, the pairwise dataset, and its teacher-embedded counterpart, enabling inspection of data composition and training procedures, independent evaluation, and reuse without repeating expensive teacher inference.
Oct 22, 2026

Beyond On-Policy vs. Off-Policy: At What Level Should LLM Agent Distillation Follow the Student?

Distilling interactive LLM agents requires choosing between teacher-generated demonstrations and student-generated experience, corresponding broadly to off-policy and on-policy distillation. While both have proven effective, we argue that this distinction is inherently hierarchical: the policy that induces an interaction history need not be the same as the policy that generates the token prefix of the current action. We introduce a hierarchical agent distillation framework that formalizes these two levels of on-policiness and enables explicit control over their pairing, mixture proportions, and training schedule. We study this problem through self-distillation, where a frozen teacher has access to task guidance unavailable to the student, and conduct a three-stage controlled study across three models and four interactive environments. We find that student-induced histories paired with teacher-generated action prefixes generally provide the strongest standalone training signal, whereas following the student at both levels generally performs worst. Nevertheless, fully on-policy data becomes beneficial when incorporated into a mixture with other configurations. Moreover, mixtures with identical teacher/student proportions at each level can yield substantially different performance, showing that how the two levels are paired matters beyond the overall amount of on-policy data. Finally, front-loading fully teacher-generated contexts and delaying fully on-policy contexts improves over a static mixture. These findings provide practical principles for deciding where, how much, and when agent distillation should follow the student.
Oct 22, 2026

HERD: Harness-Evolving Recursive Distillation for Continual Learning

Continual learning requires a model to keep acquiring new skills and knowledge without compromising its existing capabilities. Parametric methods suffer from catastrophic forgetting and loss of plasticity. Non-parametric methods, such as prompts, skills, and harnesses, leverage in-context learning but are limited by context length and lossy plain-text compression. To effectively combine their complementary strengths, we propose HERD (Harness-Evolving Recursive Distillation), a hybrid continual learning framework that learns efficiently in the harness and internalizes experience into cumulative capabilities via distillation. On the harness side, HERD continually refines its skills to capture reusable patterns, while retaining high-residual examples that the skills fail to capture in a bounded memory. On the parameter side, HERD performs drift-constrained distillation with dual anchors, substantially reducing forgetting over long recursive runs. On the 25-stage task stream of TRACE+17, under fixed memory and compute budgets, HERD matches the strongest parametric baseline in new-task acquisition while exhibiting the least forgetting and the strongest retention of general capabilities.
Oct 22, 2026

One Teacher, Many Views: Reducing Reference Sensitivity in On-Policy Self-Distillation

On-policy self-distillation (OPSD) improves reasoning by training a model on its own trajectories with dense token-level supervision from a privileged copy conditioned on a reference solution. Existing methods typically use a single teacher distribution conditioned on the complete reference. We identify a limitation of this formulation: a reference is only one realization of the reasoning process. Although it provides evidence about the answer, it also encodes an ordering, phrasing, and solution path. Full-reference conditioning can therefore produce correct targets that depend excessively on information unavailable to the student at inference. We show that replacing one valid reference with another, while holding the problem and student state fixed, can substantially alter the teacher distribution, a phenomenon we call reference-induced target instability. To address this instability, we introduce Monte Carlo Privileged Distillation (MCPD), which models privileged context as a distribution over partial reference views and distills the marginal teacher distribution. MCPD evaluates the teacher under multiple views and averages the resulting next-token distributions. We show that this probability-space average is the exact target induced by minimizing the expected teacher-to-student KL divergence over privileged views. Across mathematical and scientific reasoning benchmarks, we show that MCPD produces more reference-consistent supervision and consistently improves over full-reference OPSD.
Oct 22, 2026

Reverse-KL Policy Distillation: Statistical Guarantees and Limits

Reverse Kullback-Leibler (KL) distillation has shown empirical promise for training language models, but its statistical guarantees and limits remain poorly understood. In particular, the effects of two design choices remain unclear: Does access to the teacher's full action distribution provide an advantage over observing only the probability it assigns to the sampled action? Must training trajectories come from the student (online distillation), or can teacher-generated trajectories (offline distillation) suffice? Our work takes a first step toward understanding the theory underlying these questions. We study the sample complexity of online reverse-KL distillation under two feedback models: the teacher reveals either its full action distribution or only the probability of the student-generated action. We establish minimax bounds for the former and show that substantially more data may be required for the latter. In contrast, offline reverse-KL distillation can attain the same minimax rate whether the teacher reveals its full distribution or only its probability of the teacher-generated action. The key distinction is that teacher-generated actions themselves reveal the teacher's preferences, while queries at student-selected actions can leave those preferences unobserved. This highlights an additional statistical challenge for online distillation compared to offline distillation. These statistical guarantees do not follow simply from optimizing standard distillation objectives: our lower bounds show that common online objectives can fail when candidates are evaluated only on past student-generated trajectories. We develop trajectory-loss and importance-ratio truncation techniques together with policy-space learning to achieve optimal online distillation guarantees.
Oct 22, 2026

GIFT: Gradient Informed Filtering for Distillation Augmented Reasoning RL

Training large language models for complex reasoning tasks relies on two complementary signals: sparse, outcome-grounded reinforcement learning (RL) for correctness, and dense, token-level distillation for guidance. However, complementarity in supervision does not guarantee compatibility in optimization; imitation-focused distillation can conflict with the outcome-driven RL objective, and existing hybrids lack a principled resolution. We propose GIFT (Gradient-Informed FilTering for Distillation-Augmented Reasoning RL), a novel method that directly addresses the allocation problem of which teacher updates to apply. GIFT performs a token-level geometric analysis, decomposing the distillation gradient into a component parallel to the RL gradient, which measures first-order conflict and provides a safety filter, and a component orthogonal to it, which quantifies novel directional guidance and provides a measure of value. This decomposition yields a principled scoring rule to rank token-level guidance. We combine this with a progressive coverage annealing schedule that shrinks the budget of tokens receiving distillation, ensuring that the most conflicting updates are discarded first while the most complementary ones are retained at full strength. This facilitates a smooth transition from dense, geometrically-filtered guidance to a pure, outcome-grounded objective at convergence. Extensive experiments on mathematical and logical reasoning benchmarks show that GIFT consistently outperforms strong baselines across multiple model scales, achieving relative gains of up to 9.5% on difficult mathematical reasoning tasks. Code: https://github.com/lilifree/GIFT
Oct 22, 2026

CODI: Improving Self-Supervised Vision Transformers with Cross Distillation

Self-supervised Vision Transformers learn strong image-level representations, but their patch-level representations can be unreliable for dense prediction. We study this mismatch and find that the class token concentrates global semantics on a small number of representative patches, while dense prediction errors often occur on low-attention regions. We provide a local gradient analysis showing that, along the [CLS] attention readout, image-level self-distillation gradients are weighted by [CLS]-to-patch attention mass. To provide masked patches with global semantics without changing the architecture, we propose Cross Distillation (CODI), which mixes a small amount of the teacher class-token distribution into each teacher patch-token target. CODI gives every masked patch an attention-independent semantic correction while preserving the standard self-distillation architecture. Experiments on dense and image-level benchmarks show that CODI improves both dense and image-level performance under aligned pretraining settings. Code is available at https://github.com/EldercatSAM/CODI.
Oct 21, 2026

Benchmarking Model Compression Techniques Across Heterogeneous Edge Platforms: A Predictive Maintenance Use Case for Rotating Machinery

To enable Industry 5.0, where autonomous systems are widely adopted and integrated with people to promote sustainable development in industries, it is imperative to adapt complex deep learning models to operate efficiently on resource-constrained hardware for edge inference. However, adapting these models remains challenging due to the hardware heterogeneity of edge computing in industrial settings, performance degradation caused by model compression, and the lack of clear guidance on which compression techniques are practical and effective for applications such as predictive maintenance. To address these challenges, this paper presents a comprehensive comparative analysis of four state-of-the-art model compression techniques applied to a high-performance convolutional neural network for predictive maintenance using a real-world dataset of rotating machinery, equipment typically used in industry. The compressed models were evaluated across a heterogeneous set of nine computing platforms, ranging from highly constrained edge computing devices to general purpose and high-performance processors, reflecting the diversity and fragmentation typical of industrial edge environments. By jointly prioritizing classification performance, inference speed, and compression ratio, we found that compression using knowledge distillation emerged as the most effective approach. The model had more than 90% speedup across all devices while maintaining a 91.22% balanced F-score, with a very small drop in predictive performance relative to the uncompressed model, along with a 98.9% size reduction. Through successful deployment across multiple edge platforms and a systematic comparison of compression strategies, we demonstrated a clear path to achieving Industry 5.0 targets, drastically cutting computational costs and, consequently, energy consumption without sacrificing predictive performance.
Oct 21, 2026

CORD: Conflict-Aware Optimization for Joint RL and Distillation in Agents

Joint reinforcement learning and self-distillation combines task rewards with dense teacher guidance for agent training, but their gradients can partly cancel. A mixing coefficient controls relative scale without selectively removing opposing components. We present CORD, a general framework for conflict-aware optimization of two objectives. CORD separates the choice of protected objective from the condition that triggers projection, and combines the gradients once per optimizer mini-batch. A nearest-point formulation yields a closed-form update and characterizes its first-order effect on the protected objective before optimizer transformations. Gradient diagnostics connect conflict in agent training to lost local optimization progress and motivate selective correction. As a gradient-combination mechanism, CORD can be added to different joint training recipes while retaining their objectives. Across three such recipes, it improves ALFWorld success by 1.5-15.9 percentage points. On WebShop, it improves two of the three pairs, with SDAR rising from 60.7% to 64.6% success. Our analyses explain why protection direction and projection gate must be chosen together, and why the distillation objective and its mixing weight remain consequential after projection.
Oct 21, 2026

How Far Can Foundation Models Be Distilled for Efficient Jet Tagging?

Jet tagging is one of the most important applications of machine learning at the Large Hadron Collider. Recent models build in more physics knowledge or rely on large-scale pretraining, but the experiments place strict limits on their inference cost. Knowledge distillation can transfer their performance to small students that are cheap to run. We ask which properties of a teacher help the student, how small a student can get and how close it can come to the state of the art. On a jet tagging benchmark, we distil six teachers that differ in architecture, symmetry and pretraining, including two foundation models, into five small students. Every teacher improves every student, and the gain depends mainly on the strength of the teacher. We then shrink our two strongest distilled students, Transformer and L-GATr-slim, in width and in depth. With 20 thousand parameters, a distilled L-GATr-slim outperforms every published jet tagger trained from scratch at a fraction of the inference cost of every teacher. Even with two thousand parameters, it matches ParT, a model with a thousand times as many parameters. We also study what the students learn from their teachers. Our benchmark of inference cost shows that students of the same size can differ by almost an order of magnitude in speed. We release the code and the weights of all trained models.
Oct 20, 2026

D-Skill: Distilling Teacher Knowledge through Agent Skills

Frontier language-model agents excel at complex tasks, but high inference costs make transferring their capabilities to smaller models essential. Traditional knowledge distillation achieves this through parameter updates, requiring access to the student's weights and additional training compute while encoding the transferred knowledge implicitly in the parameters. We introduce D-Skill, a framework that treats the teacher as a black box and distills its behavioral advantage into an interpretable natural-language skill by analyzing paired rollouts and identifying behavioral divergences and agreements. The resulting lessons are consolidated offline, with any compatible skill-update backend, into a skill document that is provided to the student at inference time. D-Skill constructs the skill entirely offline from fixed trajectories, without labels, further environment interaction, or weight updates after rollout collection. Across SpreadsheetBench-Verified, ALFWorld, and chess puzzles spanning multiple teacher-student pairs, D-Skill closes 75%-95% of the teacher-student gap without touching model weights. After a single offline pass over the paired traces, D-Skill outperforms one-epoch SFT baselines and remains competitive with multi-epoch SFT. Under the same model family, the resulting student also incurs 5.7x-8.3x lower inference cost per task than the teacher. Our results demonstrate that natural-language skill distillation offers a performant, interpretable, and efficient alternative to weight-based capability transfer.
Oct 20, 2026

CSD: Common-Support Distillation for Cross-Tokenizer Knowledge Transfer

Cross-tokenizer logits distillation enables knowledge transfer across model families, allowing students to learn from stronger or more specialized teachers. However, differences in vocabulary and segmentation make teacher and student next-token distributions difficult to align. Existing logits-based methods are easy to integrate into standard distillation pipelines, but often rely on restrictive heuristics or tokenizer-pair-specific choices. They may restrict supervision to shared vocabulary subsets, require manually selecting different objectives for different tokenizer pairs, or match only the observed chunk and its complement rather than individual alternative continuations. We introduce Common-Support Distillation (CSD), a unified logits-based framework that constructs a shared support between the teacher and student at each aligned input chunk and matches the induced distributions with a single objective. For one-to-one alignments, CSD uses explicit common-token pairs as the shared support. For other alignments, it constructs a virtual support whose probabilities are computed using the autoregressive chain rule. A residual event captures probability mass outside the explicit support, ensuring that both induced distributions remain normalized. Across four teacher-student settings comprising Llama-3.2-1B and Gemma3-1B-PT students with Qwen and Phi teachers, CSD consistently achieves the highest average performance among the evaluated cross-tokenizer methods, outperforming the strongest alternative by up to 1.8 points and surpassing both continued pretraining and same-tokenizer distillation. For Llama-3.1-8B, CSD improves average accuracy over the strongest evaluated baseline by up to 4.2 points for the base student and 1.6 points for the instruction-tuned student.
Oct 20, 2026

De-Diffusing WAMs: Action-Sufficient Direct Prediction for Robot Control

Cascade world-action models (WAMs) use predicted visual futures to condition robot actions, but inherit costly iterative inference from diffusion-pretrained experts. We ask what a single-step predicted future must preserve to support accurate single-step action prediction. In the evaluated WAM, resampling video noise has little effect on downstream actions, yet naively truncating both streams to one step substantially reduces task success. Video-space supervision only partially closes this gap, and a local sensitivity analysis shows that vision prediction errors have highly unequal consequences for action accuracy. These findings motivate DirectWAM, which converts pretrained video and action experts into single-step predictors without stochastic sampling through action-sufficient distillation, while retaining the explicit future-to-action cascade. After learning a one-step action model, we train the future predictor by decoding actions from its own predicted futures and backpropagating demonstration action error through the expert, with an optional future-latent reconstruction loss. The resulting objective evaluates the student-generated future through its downstream control consequences rather than relying on representation matching alone. DirectWAM achieves 97.1% average success across four LIBERO suites and 85.4% on RoboTwin 2.0, outperforming Flash-WAM at the same one-evaluation-per-stream budget. This advantage has also extended to three real-world tasks that we designed. Compared with the full iterative teacher, it reduces inference latency by 28.0×28.0\times.
Oct 20, 2026

On-Policy Transition Distillation

On-policy distillation (OPD) consolidates expert capabilities into a single student by matching teacher and student final-layer distributions on student-generated rollouts. We investigate whether depthwise prediction transitions, the changes in predictions between layers, can provide additional task-relevant supervision. Our diagnostic study finds that stronger teacher--student agreement in intermediate transitions is associated with higher answer correctness within questions, with correct responses exhibiting lower mean transition KL across all measured intermediate layer pairs. These observations motivate On-Policy Transition Distillation (OPTD), which augments final-layer distribution matching by aligning teacher and student layer-to-layer log-probability increments on student-generated rollouts. Evaluations across three backbones and eight benchmarks spanning mathematics, coding, and instruction following show that OPTD achieves the best or tied-best overall performance in both dual- and three-domain distillation. In the three-domain setting, OPTD recovers ≈90%\approx 90\% of the overall teacher--student capability gap, compared with ≈75%\approx 75\% for OPD.
Oct 20, 2026

Building Telecom Language Models with RSOD: Domain Data, Specialist Post-Training, Memory-Conscious Distillation, and Networking Evaluation

Telecom language models must combine broad networking knowledge with specialist behavior across routing, broadband services, access technologies, and operational diagnostics while remaining practical to deploy as a single model. We present an end-to-end framework that builds source-grounded synthetic training data from structured networking taxonomies and expert graphs, materializes 212,559 routed training records across fourteen networking routes, trains six frozen specialist LoRA teachers over one shared continued-pretrained backbone, and consolidates their behavior into one deployable student adapter using Routed Specialist On-Policy Distillation (RSOD). In RSOD, the committed student generates each trajectory, the route-assigned specialist teacher-forces the exact native token sequence, compact teacher hidden states are reconstructed through the shared frozen output head, and only the student LoRA is updated. A controlled reconstruction validation shows mean absolute logit error of 7.26 x 10^-6 with 100% top-1 agreement between native and reconstructed teacher logits. On four RTX 6000 Ada GPUs, the completed 2,000-step run produced 16,000 trajectories with zero observed student-version or token- ID mismatches. Under the primary route-conditioned evaluation over 2,850 held-out questions, the raw Meta-Llama-3.1-8B-Instruct model scores 12.8661/25 micro (13.0172/25 macro), the combined routed six-teacher system scores 21.7641/25 micro, and the consolidated RSOD student, released as Llama-CalixCLM v1, reaches 21.7112/25 micro (21.8006/25 macro) on the same five dimension LLM-judge rubric. Under a separate neutral deployment-like prompt, the same student reaches 20.0028/25 micro (20.0778/25 macro). The consolidated student remains within 0.0529 micro-score points of the combined routed-teacher system while retaining a numerically validated and auditable train-serve execution path.
Oct 20, 2026

Staged Continual Adaptation of Multimodal Foundation Models for Japanese Financial Documents

Staged post-training of multimodal foundation models chains vision-language alignment, reasoning distillation, and domain tuning, yet how each phase trades capabilities against the others remains under-characterised. Tracking an 8.4B-parameter model from an un-fine-tuned baseline through three training phases on Japanese financial documents (evaluated on five text-input benchmarks), we find the phases interact non-additively: math-only distillation overwrites the financial knowledge acquired earlier even as procedural math skill survives, and the domain phase repairs more than it installs. A skip-phase ablation is best for domain knowledge but forfeits numerical-reasoning gains, so per-phase contributions are not decomposable and the best checkpoint is task-dependent, with implications for checkpoint retention and scheduling rather than added stages.
Oct 19, 2026

Understanding the Benefits of Scaffolds through Computational Separations

Inference-time harnesses, i.e., external procedures that call a model many times and keep state across calls, can substantially improve model capabilities. In this paper, we investigate the benefits of using such harnesses and the extent to which these benefits can be internalized through training. More precisely, we conceptualize a harness A\mathcal{A} as an arbitrary wrapper around a base model MM and attempt to distill a student model from the harness's traces. Empirically we find that, even with simple scaffolds and small models, the inference-time harness significantly outperforms a distilled student. We then prove that such a gap is due to a fundamental representational limitation: inference-time harnesses can compute a closure of their per-call primitive that is \NL\NL-complete even when the primitive is a trivial \AC0\AC^0 lookup, and that no fixed-depth forward pass can represent under the widely believed conjecture that \NL⊈\TC0\NL\not\subseteq\TC^0. Our theory suggests that expanding the representational capacity of the student relative to the task at hand, e.g., by enlarging the model, making the task easier, or lengthening its chain-of-thought, is necessary to close the gap, which we verify empirically. Finally, we scale to an agentic harness on competition math, where the same gap reappears. Overall, our results show that a harness confers on the base model a profound gain in representational power, one that training a same-class student on its traces cannot transfer back into the weights.
Oct 19, 2026

Teacher Supervision, Fixed Representations and Calibration: Exact Risk Calculations for Candidate-Scoring Decision Models

Candidate-scoring decision models return a probability distribution over a presented set of answers. Public documentation of one such system, Jev, and of an open counterpart, JevAny, describes three training ingredients that determine what these probabilities estimate: soft targets from larger models, a frozen pretrained backbone and temperature calibration. We analyze each ingredient in Gaussian surrogates where the finite-sample risk can be computed exactly. A student trained on teacher targets inherits the teacher's bias, which more student data does not remove and agreement with the teacher does not reveal. A fixed representation lowers the risk only below an explicit sample-size threshold. In a probit model the exact recalibration map is a temperature scaling, which removes all excess risk only when the teacher's coefficient has the correct direction and otherwise leaves a residual that calibration diagnostics do not detect. Under Luce's choice axiom, candidate scores are coherent across refinements of an event if and only if they aggregate by log-sum-exp. On 231,521 credit card complaints from the Consumer Financial Protection Bureau, where the decision is whether a complaint ends in monetary relief, a student trained on a logistic teacher's probabilities ends at the teacher's decision cost while a student trained on outcomes goes 8% below it, a 30B model read zero-shot remains worse than the teacher after recalibration, and that model departs from the coherence condition by 17.5 decibans. Pretrained representations reduce the labeled data required and calibration corrects the scale of the probabilities, but neither corrects a biased teacher, so the probabilities have to be checked against realized outcomes.