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 13, 2026

CAFE: Confining Feature Distillation to Language-Routed Adapters for Compact Multilingual Automatic Speech Recognition

Feature distillation is usually expected to add to logit distillation. We find that this expectation does not hold for two common feature losses in compact multilingual automatic speech recognition (ASR), and analyze why through their gradients. The gradient of a projection-based loss is orthogonal to the task gradient, while that of Centered Kernel Alignment (CKA), a similarity-based loss, is weakly aligned with it. CKA lowers the word error rate (WER) on non-English languages, but at the expense of data-rich English. Three loss-level remedies (one trained, two probed) all fall short, and the CKA gradient overlaps with the subspace English uses in the shared encoder by 97.5% to 99.5% (about 81% by chance). Guided by this diagnosis, we propose CAFE (Confined-Adapter Feature distillation), which confines the CKA gradient to adapters that English bypasses, addressing the trade-off in the parameters rather than in the loss. On the Open ASR Leaderboard, CAFE lowers the average WER from 8.55% to 8.02%, beats an adapter-only capacity control at 8.51%, and makes English more stable across checkpoints. At 254M parameters, it trails the 0.6B teacher by 2 WER points and is competitive with much larger models across 48 test sets in 24 languages. We use 564k hours of training data and state-of-the-art models to keep the findings relevant to practice.
Oct 13, 2026

GAD: Geometry Attention Distillation for Consistent Video Generation

Video diffusion models generate visually compelling short-term sequences but often struggle to preserve scene geometry and object structure over time. In this paper, we first demonstrate that current video attention cannot identify reliable correspondences across frames, with responses dispersed across multiple regions even after recognizable scene structure has emerged. We argue that reliable cross-frame associations are essential for information exchange as scene structure forms in video generation. Thus, we introduce Geometry Attention Distillation (GAD), which transfers correspondences from a frozen geometry model's attention to video attention. Extracted from clean videos, these correspondences link the same static surface or moving object across frames. Since structure formation begins in deeper layers at higher noise and shallower layers at lower noise, we apply attention distillation at a student layer selected according to the noise timestep. After training, video attention focuses more consistently on corresponding locations. Experiments across all three video backbones demonstrate improved long-range scene consistency and dynamic temporal consistency, with geometric gains across generation lengths. On Wan2.2 5B, GAD improves cross-frame matching accuracy by at least 30% over the Wan baseline on both long-range benchmarks. Notably, GAD achieves these gains without additional inference modules, offering a promising approach to efficient and consistent video-based world simulation. Project page: https://fantasy-amap.github.io/GAD/.
Oct 13, 2026

Does self-distillation improve reasoning?

Self-distillation is an emerging approach that trains a model to imitate its own predictions when conditioned on privileged information, such as correct answers or environmental feedback. Recent methods have reported better sample efficiency and generalization than reinforcement learning (RL), but it remains unclear whether these gains extend beyond in-distribution evaluations and whether they stem from privileged distillation. In this paper, we revisit two representative methods, Self-Distillation Policy Optimization (SDPO) and On-Policy Self-Distillation (OPSD), across code, science, tool use, and math reasoning tasks. We find that their reported gains do not consistently generalize to unseen problems and are largely explained by overfitting or transferring existing reasoning behavior, rather than effectively using privileged information. Moreover, self-distillation gradients show weak alignment with policy gradients that directly optimize expected reward, questioning whether they provide an effective direction for policy improvement. Our experiments across Qwen3 models (1.7B-14B) and OLMo-3-7B suggest that effectively leveraging privileged information to improve reasoning remains an open challenge. Our code is available at https://github.com/sleepypandu/does-sd-improve-reasoning
Oct 13, 2026

Imprinting: Knowledge Distillation Can Shift Student Behavior

Knowledge distillation trains student models on teacher-generated responses to transfer task capability, but those responses may also carry behavioral tendencies of their source. We investigate whether the source teacher can shape student behavior beyond the capability that distillation is intended to transfer. We construct harm-accepting and harm-avoiding variants of the same base teacher, use them to generate responses to ordinary reasoning and knowledge questions unrelated to safety, and distill those responses into students with different base models. Students generally acquire task capability while their downstream decisions shift toward the source teacher's safety-pragmatism disposition, a phenomenon we call imprinting. Distillation from the harm-accepting teacher reduces harm avoidance, whereas distillation from the harm-avoiding teacher increases it. Imprinting varies across teacher-student pairs and does not extend uniformly across different forms of safety behavior. Excluding a sparse set of divergence tokens reduces imprinting, whereas response rewriting preserves the source-dependent effect and can introduce additional rewriter-dependent shifts. Post-distillation preference optimization raises harm avoidance but does not eliminate the source-dependent difference. These results show that behavioral tendencies can be intentionally imprinted through the source teacher, even with ordinary task supervision.
Oct 13, 2026

DisaggKD: Efficient Knowledge Distillation through Teacher-Student Disaggregation and a Fused Loss

Knowledge distillation (KD) frameworks place teacher and student on the same GPUs, so per-GPU memory limits the teachers that fit. DisaggKD separates them across nodes: teacher replicas stream hidden states to the student, a KD loss over row windows and vocabulary shards equals the full-vocabulary loss in exact arithmetic without forming the full-vocabulary logits, and a prefetch ring paces producer and trainer. Closed-form memory and traffic models locate the memory limits of each layout. On the Qwen3.5-35B-A3B to 9B pair, the disaggregated layout matches the LM loss of the colocated run with the same loss and is faster than that run; our fused loss also lets a colocated run train on as few nodes at faster speed, while the split is memory efficient when the student spans several nodes. On the larger pairs, colocated runs with a fused loss still run OOM where the split trains, because teacher weights and student states fill the budget.
Oct 13, 2026

RELAY: Selective Self-Distillation with Adaptive Hand-off for Rubric-Based Reinforcement Learning

Rubric-based reinforcement learning grades a response against many written criteria but collapses them into one reward, which discards the detail the rubric carries and gives no GRPO reward gradient when a group's responses tie. A frozen teacher that reads the rubric as privileged context provides token-level supervision, but we observe an early plateau in the gains over GRPO. RELAY uses advantages and rewards to select positive-advantage and low-scoring zero-advantage rollouts, and monitors the distillation loss to decide when to retire the teacher. Stage 1 combines selective distillation from a frozen teacher with GRPO; a plateau detector then removes the teacher, and Stage 2 continues with GRPO alone. Calibrated on one model-domain pair and reused unchanged in three others, the plateau detector fires between steps 122 and 171. RELAY scores highest in-domain among the compared methods in all four settings, ahead of GRPO on RaR-Medicine by 4.62 points with Qwen3-1.7B and 6.02 with Qwen3-4B. On Qwen3-1.7B/Medicine, removing the teacher from a shared checkpoint outperforms keeping, weakening or refreshing it. Distilling selected rollouts covering about half the response tokens also outperforms distilling all rollouts at a matched average coefficient.
Oct 12, 2026

From a Large-Scale Teacher to Efficient Multimodal Students for Search Relevance Prediction

Relevance prediction between user queries and items is a fundamental component of e-commerce search systems. Traditionally, high-quality relevance labels are obtained through human judgment, which is reliable but expensive, slow, and difficult to scale. Recent advances in large language models (LLMs) enable automated relevance judgment at scale, achieving performance comparable to human annotators. However, such systems often rely on large commercial LLMs with complex prompts, resulting in high inference costs and limited throughput. In this work, we propose a framework that bridges this gap by distilling a high-performing LLM-based relevance judgment system into efficient multimodal student models. We first curate a small, high-quality human-annotated dataset to serve as ground truth. Using this dataset, we develop a reliable LLM-as-a-judge pipeline through careful prompt engineering and iterative refinement. This pipeline is then used to generate large-scale relevance labels, which serve as training data for compact multimodal models. Our distilled models achieve performance close to the teacher system while significantly improving latency, cost, and scalability. This approach enables practical deployment of high-quality relevance prediction systems in production environments.
Oct 12, 2026

Overtaught, Underlearned: Learner-Centered Distillation from Multi-Agent Discussion

Multi-agent discussion produces reasoning collaboratively, while deployment still relies on a smaller student that must answer alone. Full-graph distillation transfers such discussions by treating each correct solution as a separate training target, but what teachers provide need not match what the student learns. This mismatch appears in two ways: several correct solutions may amount to the same lesson in the student's representation (over-teaching), while some training questions remain missed after learning (under-learning). We quantify both with rates that are high on discussion graphs. To address them, we introduce FAR (Fold--Attempt--Revisit), a learner-centered distillation procedure that organizes supervision around the student's own state. Before training, FAR folds solutions within each discussion state into representative lessons. After training, the student's own attempts identify missed questions, while poorly represented clusters indicate where folding may have lost useful variation; FAR then writes targeted lessons from retained discussion contexts. Across two students and four reasoning tasks, FAR outperforms full-graph distillation with far fewer training targets and lower under-learning. Further analyses examine the behavior of FAR and the sources of its gains. Overall, the results show that organizing discussion supervision around the student can make distillation both more effective and more efficient.
Oct 12, 2026

Avoiding What Your Teacher Already Knows: Entity Familiarity for Deep-Search Distillation

Deep-search agents are commonly trained by distilling tool-use trajectories from stronger teacher models. However, structurally complex questions may remain relatively easy for teachers familiar with key entities in the underlying entity graph, resulting in lower search demand and weaker supervision. We refer to this teacher-specific variation in question difficulty as teacher-relative difficulty and investigate entity familiarity as an early predictor. We first introduce Evidence-Verified Entity Familiarity (EVF), which estimates familiarity by verifying a teacher-generated factual brief against online evidence. Across multiple teachers, EVF is significantly correlated with teacher-relative difficulty with AUCs of 0.60--0.70. However, its high cost makes it impractical for constructing entity graphs at scale. To address this limitation, we propose Self-Reported Entity Familiarity (SEF), a lightweight gating signal that directly asks a teacher whether it is familiar with an entity before admitting that entity into the graph. Across multiple teachers, trajectories from SEF-gated synthesis exhibit greater observed search demand and yield higher downstream student accuracy than those from ungated synthesis. Because familiarity is teacher specific, we further study cross-teacher reuse. We screen graphs collected for other teachers using Kimi-K3's SEF judgments and use the retained graphs to produce 5,411 additional trajectories. Combined with 2,200 trajectories from newly constructed, SEF-gated graphs, these data improve Qwen3.8-27B's accuracy from 44.10% to 66.41% on a BrowseComp subset, and achieving state-of-the-art performance among models of comparable size on BrowseComp, SEAL-0, HLE-text, and xbench. Overall, SEF provides a low-cost way to synthesize more informative supervision for deep-search distillation.
Oct 11, 2026

Representational Difference Distillation

Knowledge distillation transfers representational knowledge from a large teacher network to a compact student. Existing relational and contrastive distillation methods treat all samples uniformly, spending equal gradient on well-matched and poorly-matched regions of the representation space alike. We propose Representational Difference Distillation (RDD), which diagnoses where teacher and student representations structurally disagree and concentrates the distillation signal on those regions. RDD uses two asymmetric affinity matrices that capture opposing directions of misalignment: one identifies samples the student incorrectly considers similar, the other identifies teacher relationships the student has yet to learn. These affinities determine which sample pairs the distillation loss aligns and how strongly it corrects each misaligned sample. On CIFAR-100 with resnet32x4 to resnet8x4, RDD reaches 75.58% test accuracy in our runs, on par with CRD (75.45%) and above KD and RKD. We additionally explore curriculum-based and sampling-based variants, finding that weighting the loss is more effective than filtering the data.
Oct 11, 2026

Radioactive Feedback: Tracing Models Trained by Label-Only LLM Supervision

Conventional distillation transfers capability through teacher-generated demonstrations. LLM-as-a-Judge knowledge distillation offers a distinct and increasingly practical alternative. In the label-only variant studied here, the student generates every response and improves from scalar scores alone. Prior work shows that this non-demonstrative feedback can improve student capability, but it creates a fundamentally different provenance problem. No teacher-generated answer enters the student's training set, so any evaluator-specific trace must pass through a low-bandwidth reward signal rather than copied text. Existing watermarks and distillation detectors do not cover this score-only channel. We introduce Radioactive Feedback, which encodes a registered pattern of behavioral preferences in the rewards themselves. The evaluator applies bounded pointwise adjustments to quality-qualified responses, allowing repeated optimization to transfer the pattern to a student trained by another party. Experiments on code and mathematics show that the registered preferences transfer through scores alone and across datasets, enabling black-box detection against a fixed reference pool.
Oct 11, 2026

The Gradient That Top-k Distillation Never Computes

Many deployed distillation pipelines show the student a truncated view of the teacher: keep the top-k probabilities, renormalize, and minimize a divergence on what remains. Renormalization is assumed to account for the missing mass. We prove that it removes the visibility of tail mass, not the freedom to move it. The renormalized objective depends only on log-odds among the retained tokens, so at each position its gradient along the directions that control tail mass is exactly zero; on a real teacher that component measures 22-3×10−153\times10^{-15} of the gradient. The full-vocabulary gradient equals the truncated gradient plus a residual weighted by the tail masses, and training acts only on the first term. Without the cross-entropy term, a top-k student's tail mass drifts away from the teacher's while a full-vocabulary control moves slightly toward it; with it, the gap shrinks about fourfold but persists. On seven open-weight teachers, the retained set determines the teacher's distribution only from k=d+1k=d+1 (dd the hidden dimension): recovery is exact from there on all seven, and commonly cached sizes sit below it. In our off-policy runs the practical effect is real but smaller than the choice of teacher-student pair. Renormalization is a normalization, not a safeguard. The code is available at https://github.com/KurbanIntelligenceLab/top-k-distillation.
Oct 10, 2026

VC-OPD: Visual Counterfactual On-Policy Distillation for Grounded Vision-Language Reasoning

GRPO-style reinforcement learning has advanced vision-language reasoning, but its sequence-level rewards provide limited credit assignment for individual reasoning tokens. On-policy distillation (OPD) addresses this issue by providing dense supervision on student-generated trajectories, yet standard OPD assigns similar importance to all teacher-preferred tokens, including template phrases and visually irrelevant continuations. We introduce Visual Counterfactual On-Policy Distillation (VC-OPD), a grounding-aware distillation method that prioritizes tokens whose correctness depends on image-specific evidence. After a teacher-supervised warm start that stabilizes response formats and reduces noisy rollouts, VC-OPD performs on-policy distillation with two teacher evaluations for each student trajectory: one under the original image and one under a counterfactual image with corrupted instance-specific evidence. The original-image teacher distribution determines the distillation target, while the original--counterfactual teacher gap estimates token-level visual dependence and softly reweights the loss. To avoid unstable response-level update scales, VC-OPD further applies mass-preserving normalization to the token weights. This design preserves the dense supervision of OPD while shifting learning capacity toward visually grounded reasoning tokens. Experiments on multimodal reasoning benchmarks show that VC-OPD improves visual reasoning performance and produces more effective token-level distillation.
Oct 10, 2026

Reward-Aligned Distillation

Reinforcement learning with verifiable rewards improves the reasoning capabilities of large language models, but remains computationally expensive and sample inefficient. On-policy distillation (OPD) provides dense token-level supervision from a stronger teacher, yet standard OPD distills every token regardless of whether the teacher's preference actually improves task reward. Existing token-selection methods rely on proxies such as uncertainty or teacher--student divergence, which need not identify reward-improving supervision. We introduce Reward-Aligned Distillation (RAD), a bilevel framework that weights distillation signals based on their predicted effect on the student's post-update reward. RAD differentiates through a temporary optimizer step to score each token's supervision, then updates the token weights by projected ascent from the initial token weights. The resulting weights can strengthen, suppress, or reverse individual distillation-gradient contributions. At a matched training-response budget, RAD improves mean accuracy on six competition-math benchmarks over the strongest baselines by 6.5 percentage points for Qwen3-1.7B (non-thinking) and 2.5 points for Qwen3-4B-Instruct-2507. RAD-Simple, a simplified variant, achieves comparable math accuracy at the same training-response budget. On LiveCodeBench-v6, RAD reaches 54.69% and 58.67% pass@1 with Qwen3-1.7B (non-thinking) and OLMo3-7B-Instruct, respectively, exceeding the strongest baselines by 8.7 and 6.4 points. These results support reward-guided weighting of teacher supervision across model families and tasks.
Oct 9, 2026

Knowing Whether, Not Just Where: Anchor-Probing Distillation for Open-Set Remote Sensing Visual Grounding

Open-set visual grounding in remote sensing requires models to determine not only where a queried object is located, but also whether the image provides sufficient evidence for the query at all. Existing closed-set grounding models, however, tend to return plausible boxes even for absent objects, blurring localization confidence with genuine Visual Support. We introduce AnchorProbe-Distill (APD), a teacher-student framework that recasts open-set grounding as Visual Support learning. APD builds on the insight that supported queries should induce coherent spatial anchors, whereas unsupported or mismatched queries tend to produce unstable or misleading localization patterns. APD-T therefore uses the Original Query as a Reference Anchor to interpret query perturbations: Semantic-Preserving Queries retain the annotated target, Near-Category Ambiguous Queries are resolved according to whether they form a reliable alternative anchor or collapse onto the Reference Anchor, and Far-Category Counterfactual Queries provide rejection supervision. This reference-aware reasoning is distilled into APD-S, enabling unified localization and rejection from a single image--query pair. Experiments show that APD enables the resulting model to reject absent targets while also improving its localization of present targets.
Oct 9, 2026

Beyond Domain-Routed Supervision: Cross-Teacher Guarding for Multi-Teacher On-Policy Distillation

Multi-teacher on-policy distillation (MOPD) extends on-policy distillation (OPD) by routing domain-specific data to specialized teachers, integrating their capabilities through dense token-level supervision on student-generated trajectories. However, domain routing only distinguishes the supervision sources for different domains; updates from all domains still act on shared parameters. When teachers differ substantially in response style and domain expertise, supervision that is effective in isolation may still lead to persistent gradient conflicts and capability degradation during joint optimization. To address this challenge, we propose Cross-Teacher Guarding for MOPD (Road MOPD), which introduces non-routed teachers as auxiliary cross-domain critics on the same student trajectories and coordinates multi-teacher supervision through token-level credit assignment and gradient updates. Specifically, cross-teacher-guided credit assignment uses teacher agreement to strengthen token-level corrections jointly supported by the teachers and attenuate conflicting signals; preventive multi-teacher gradient projection uses auxiliary gradients from the same trajectories to constrain potentially conflicting updates and adaptively controls projection strength to retain target-domain learning signals. In experiments, Road MOPD mitigates degradation in the disadvantaged domain, gains over 5 percentage points under high conflict, and improves overall performance in low-conflict and three-domain settings.
Oct 9, 2026

CV-OPD: Contrastive Vision On-policy Distillation

On-policy distillation (OPD) reveals versatility and convenience during VLM post-training, providing dense supervision for student from teacher. However, we observed an anomaly that the distilled VLM could still answer the questions accurately even when the related visual regions were masked or replaced with noise. Building on this observation, our quantitative and qualitative experiments demonstrate that OPD, which uses the answer as a hint, has failed to teach students to genuinely seek out and rely on visual evidence. Our theoretical analysis suggests that this failure arises from a shortcut enabled by the lack of visual constraints during distillation. To address the issue, we introduce Contrastive Vision On-policy Distillation (CV-OPD), which compares teacher distributions induced by positive and negative crops while conditioning on the same student prefix and augments optimization with a margin-based contrastive constraint, which is modulated with discrepancy gating across token and sample level. Specifically, CV-OPD first introduces a negative teacher unrelated to the answer and designs a Contrastive Divergence Loss (CDL) that constrains the feature distribution of student by contrasting the positive and negative teachers, correcting how the student exploits visual regions. The CV-OPD introduces a Discrepancy Aware Gate (DAG) that uses token-level and sample-level discrepancies to select visually relevant tokens for optimization, directing the student's attention toward visual content. Extensive experiments across a broad range of fine-grained visual benchmarks demonstrate that CV-OPD substantially alleviates shortcut and consistently improves the visual reasoning capabilities of existing VLMs.
Oct 9, 2026

FPD-Drive: Future Policy Distillation for End-to-End Autonomous Driving

End-to-end autonomous driving models are typically trained on real-world driving logs, which record only the trajectory actually executed in each scene. This provides limited supervision for learning alternative valid behaviors. Multi-target supervision can mitigate this limitation, but broader trajectory coverage makes safe candidate selection more challenging. We propose FPD-Drive, an end-to-end planner that distills privileged future knowledge into a visual driving policy to improve both multi-trajectory supervision and safety selection. Specifically, we introduce Future-Privileged Multi-Target Distillation to transfer diverse, high-quality trajectory targets from a teacher with access to recorded future observations to the student. These targets extend supervision beyond the single recorded human trajectory. For candidate selection, we further introduce Safety-Aware Score Correction to correct ranking errors between nearby safe and unsafe candidates and raise underestimated safety scores of safe candidates. At deployment, the student requires neither future observations nor teacher inference. Our base model achieves 94.9 PDMS on NAVSIM v1 navtest and 55.9 EPDMS on NAVSIM v2 navhard. The scaled FPD-Drive-Pro improves these scores to 95.5 PDMS and 58.1 EPDMS, respectively, achieving SOTA performance among the compared methods on both benchmarks. Zero-shot closed-loop evaluation on HUGSIM further yields an HD-Score of 45.1 and RC of 55.3.
Oct 9, 2026

Redirecting Teacher Supervision: Temporal-Response-Guided Latent Distillation for Video-Language Models

On-policy distillation (OPD) transfers capabilities through teacher predictions on student-generated sequences, as studied by GKD and MiniLLM for language models and Video-OPD for video tasks. We ask how internal supervision can further support temporal learning. Direct hidden-state alignment does not explicitly use the student's temporal responses to choose a supervision direction. We propose Temporal-Response Guidance (TRG), which measures representation responses to local temporal perturbations, adds the teacher correction's projection onto their span back to the original correction, and restores its original norm. The projection is not normalized separately, and correction components outside the response space are retained. With an orthogonal tangent projector and a detached target, TRG changes the supervision direction while preserving the cosine-loss value and aligned-representation gradient norm at target construction. On TempCompass, Video-MME, MVBench, and Video-MMMU, TRG improves the 4B student's macro accuracy by 2.86 percentage points over the reported output-only OPD baseline.
Oct 9, 2026

Hindsight Structured Advice Distillation: Steering Executor Agents with Advisors Trained on Their Failures and Task Characteristics

Language-model agents are used for specialized work such as financial research, healthcare and law. Such agents are often capable enough for a task but still miss its details. In finance, for example, these details include which filing reports a figure or how a company discloses a metric. Retraining the agent to learn these details is expensive, and often impossible when the agent is a closed or shared system. Writing the missing know-how into the agent's prompt makes the prompt long and identical for every input, although different inputs need different guidance. We propose to learn to generate this guidance instead. A small advisor model reads each input and writes structured advice, which is appended to the prompt of a frozen executor. The advisor is trained with hindsight, following work on distillation with privileged information. On a training pool with graded answers, we analyse the executor's failures, run it under several strategies to see which one worked on each input, and let a teacher that sees the grades and the traces write the advice that would have helped. The advisor then learns, by supervised fine-tuning, to write that advice from the input alone. Because the student is a separate small model rather than the executor, it is cheap to train and adds little latency. We call this approach hindsight structured advice distillation (HSAD). On FrontierFinance, a public benchmark of expert-written financial research queries, an 8B advisor raises the qualification rate of Samaya's proprietary research agent, which already leads the benchmark's leaderboard, from 52.9% to 56.6%. This is above the agent's own high-effort configuration, at 37% lower cost per query. We study the generalization of the proposed approach, and in particular show that HSAD also improves results on WebShop, BIRD text-to-SQL and ExpertLongBench.
Oct 9, 2026

COFFEQA: Counterfactual-Guided Few-Shot Distillation for LLM Question Answering

Knowledge distillation aims to train a smaller student model to mimic a larger teacher model for resource-constrained environments. However, the process of distillation is typically data-intensive and computationally expensive, particularly for generative tasks. Few-shot distillation offers an alternative by transferring teacher knowledge using limited training samples. However, we observe that few-shot distillation performance is heavily dependent on the choice of these training samples. In fact, we demonstrate that using more uncertain samples near the teacher's decision boundary can significantly deteriorate distillation performance. Motivated by this observation, we propose Counterfactual-guided Few-shot Distillation in Open-book Question Answering (COFFEQA), a novel strategy that aptly uses counterfactuals to provide additional supervision around challenging factual samples. In the generative QA setting, we define a counterfactual as a minimally modified context that induces the teacher to change its response to an alternative answer. We provide geometric insights on how factual-counterfactual pairs can constrain the deviation between the teacher's and student's decision boundaries. We further introduce a new metric, Context Following Difficulty (CFD) that further validates counterfactuals prior to distillation based on agreement with the intended alternative answer and dependence on the modified context. COFFEQA uses CFD-validated factual-counterfactual pairs for few-shot distillation. Experiments across five in-domain and three out-of-domain QA benchmarks using T5 and Qwen2.5 families show that COFFEQA outperforms factual-only distillation baselines under the same few-shot budget. For instance, on TriviaQA, COFFEQA improves over the factual-only baseline by 8.77 average EM points and 9.98 average F1 points under the T5 setting with 32 few-shot pairs.
Oct 8, 2026

Universal Textual Teaching for LLMs

Knowledge distillation (KD) transfers knowledge from stronger Teacher models to weaker Student models, but most methods require training the Student parameters, thereby binding the distilled knowledge to a specific architecture and checkpoint. This implicit representation is difficult to interpret or reuse across models and limits KD for API-only or costly-to-train models. This paper studies knowledge transfer for large language models (LLMs). We introduce Universal Textual Teaching (UTT), a parameter-update-free framework that distills observed Teacher-Student knowledge gaps into a textual, interpretable, and reusable natural-language artifact called Primer. Specifically, UTT first identifies representative gap cases through paired evaluations, and iteratively updates the Primer via multi-role interactions: the Student attempts each task, the Prompter turns evaluation feedback into a teaching instruction, the Teacher provides a targeted demonstration, and the Synthesizer consolidates validated lessons. Empirically, on the challenging math (Omni-MATH-2) and code generation (KernelBench) tasks, extensive results confirm the effectiveness of the method: UTT remarkably raises the Student's accuracy from 9.4% to 48.6% and Fast1 accuracy from 9% to 35% on KernelBench, while increasing mathematical reasoning accuracy from 27.6% to 51.7%. UTT also performs better than representative prompt engineering and parameter-based KD methods. Of note, UTT is shown to be generalizable across different Teachers and Students: a Primer synthesized for one Teacher-Student pair can generalize to other Students that do not participate in the synthesis.
Oct 8, 2026

DIAL-OPD: Learning More from Fewer Tokens in On-Policy Distillation

On-policy distillation (OPD) supervises student-generated trajectories with token-level teacher signals. Its sampled-token variant avoids the cost of full-vocabulary probabilities. Yet we find that training on fewer tokens can outperform full-token OPD, challenging the intuition that more supervision improves learning. This motivates selecting tokens by learning value. Existing disagreement-based criteria ignore probability scale: tokens assigned negligible probability by both models, termed low-low tokens, can receive large log-ratio rewards and hinder learning. We propose DIAL-OPD, a token-selection method that bridges log-probability and probability spaces by weighting reward magnitude with the logarithmic mean of teacher and student probabilities. A parameter beta controls this weighting, and the highest-scoring tokens are retained. Across 4 teacher-student pairs and 7 mathematical reasoning benchmarks, we compare DIAL-OPD with 9 baselines. Retaining only 40% of tokens, it outperforms Vanilla OPD and its full-token variants, with mean accuracy gains reaching 5.25 percentage points over Vanilla OPD, and doubles AIME25 Pass@16 from 13.33% to 26.67%. It also achieves up to an 18% relative improvement in mean accuracy over the strongest token-selection baseline at matched retention ratios. With a 4B teacher, DIAL-OPD surpasses the strongest full-token baseline using an 8B teacher at both student scales, showing that effective supervision allocation can outweigh teacher scaling. Further analysis shows that moderate beta balances suppressing low-low tokens against preserving useful disagreements. Token-level evidence reveals that DIAL-OPD filters high-reward tokens with limited reasoning value while preserving supervision critical to reasoning correctness.
Oct 8, 2026

TAM: Task-Aware Memory Distillation for Efficient Spatiotemporal Prediction

Knowledge distillation enables efficient spatiotemporal prediction by transferring knowledge from an accurate teacher to a compact student. However, matching outputs or features independently for each sample leaves cross-sample predictive structure underused. Exploiting this structure requires representations and historical references that reflect the dynamics of each task. We propose TAM, a Task-Aware Memory Distillation framework that organizes a frozen teacher's knowledge into a bounded, retrievable history. Memory entries encode latent features, forecast changes, or flow residuals, while task-specific selection rules identify relevant historical references. The student either matches the teacher's similarity distribution over shared references or regresses observation-conditioned residual prototypes. These objectives complement supervised prediction and conventional distillation. The teacher, memory, and auxiliary adapters are used only during training, leaving student inference unchanged. We evaluate TAM on video prediction, weather forecasting, and traffic flow prediction across multiple teacher-student configurations. Averaged over four paired runs, adding TAM improves SSIM on all six video datasets and reduces MSE on five relative to the corresponding KD baselines. Mean paired MSE reductions reach 1.86% on KittiCaltech, 1.93% on WeatherBench with a gSTA teacher, and 1.01% on TaxiBJ. These results demonstrate the utility of historical teacher supervision across distinct forecasting tasks without additional student inference cost.
Oct 8, 2026

SpatialOPSD: Self-Distilling Spatial Intelligence from Verified Coding Agent Traces

Spatial coding agents significantly improve spatial reasoning in Multimodal Large Language Models (MLLMs) by using external tools to generate verified execution traces. However, this paradigm inherently suffers from prohibitive inference-time overhead and external dependencies. In this paper, we explore whether an MLLM can internalize this agentic capability to operate entirely tool-free. We begin with a simple observation: prompting an MLLM with summarized execution traces of a spatial coding agent naturally unlocks the model's internal spatial Chain-of-Thought (CoT). Motivated by this, we introduce SpatialOPSD, an on-policy self-distillation framework that internalizes spatial reasoning into a standalone MLLM by formulating verified agent traces as privileged information. To mitigate privileged-information leakage during distillation, we introduce Repetition-Aware Distillation, which combines repetition masking with unlikelihood regularization. Experiments across multiple benchmarks demonstrate that self-distilling SpatialOPSD achieves higher average accuracy than SFT and GRPO on both spatial and OOD datasets, exhibiting superior performance and generalization.
Oct 8, 2026

Efficient Multi-Granularity Knowledge Transfer for Radiology Report Generation

Radiology report generation can automatically generate clinical descriptions from X-ray images, thereby significantly improving the efficiency of radiologists. This task is challenging because it requires medical knowledge to accurately identify diseases and describe them in a professional manner. However, existing methods often overlook the importance of enhancing medical knowledge in describing pivotal areas, a capability that requires models to effectively extract and aggregate knowledge at multiple levels of granularity. Accordingly, we herein propose a novel and compact Efficient Multi-Granularity Knowledge Transfer (EMGKT) method to address the above issues. First, we encode global knowledge embeddings using a medical vision-language model, which provides contextual medical knowledge. Moreover, we devise a novel Fine-Grained Knowledge Distillation (FGKD) training task which efficiently extract fine-grained knowledge. Specifically, the FGKD training task contains teacher embeddings and student embeddings. Teacher embeddings are encoded using extra priors; while student embeddings are learned from the teacher embeddings through knowledge distillation. During inference, the student embeddings are used to enhance fine-grained knowledge while the teacher embeddings are discarded, resulting in negligible computational costs and no need for extra priors. Finally, we further develop a mixture of disease diagnosis expert classifiers to enhance knowledge extraction. The classifiers are initialized using disease embeddings and are modeled as different experts to address various granularity features. Notably, EMGKT can be efficiently applied to most existing methods. Extensive experiments are conducted on two widely-used public datasets and various baselines, which demonstrates the effectiveness and transferability of EMGKT.
Oct 8, 2026

When Do We Need On-Policy Distillation? Distilling on Offline Student Rollouts Is Often Better

On-policy distillation (OPD) has become increasingly popular for transferring teacher capabilities to student models. In this work, we ask a critical research question: Is on-policy sampling always beneficial for distilling arbitrary teacher-student pairs? We show that a simple alternative, Semi-OPD, which distills from offline rollouts generated by the initial student, can often outperform OPD in both accuracy and training efficiency. Across 17 teacher-student pairs ranging from 1.5B to 235B parameters, Semi-OPD outperforms OPD in 14 cases, with up to +13.6% accuracy and 11.4x training speedup. We further find that the choice between OPD and Semi-OPD depends on the alignment between the initial teacher and student, quantified by an output-token overlap ratio: OPD is beneficial only when the two are highly aligned with high overlap ratios. Our deeper investigation suggests that effective distillation requires on-policyness w.r.t. both the student and the teacher. For misaligned pairs, student rollouts can become increasingly off-policy w.r.t. the teacher as context length grows, weakening the distillation signal. In contrast, Semi-OPD is often more stable, as it distills on shorter contexts while covering full trajectories and exposing the student to more teacher-preferred tokens. Beyond proposing Semi-OPD as an efficient alternative, our work motivates the community to rethink when to use OPD and to study stronger OPD variants with meaningful teacher-student pairs.
Oct 8, 2026

PIVOT: Perplexity-Informed KD-to-RL Transition Scheduling for Vertical-Domain Few-Shot Distillation

Vertical-domain few-shot classification remains challenging for small language models, as limited supervision makes it difficult to acquire domain-specific decision knowledge. On-Policy Distillation (OPD) can improve teacher-guided adaptation by supervising student-generated rollouts, while GRPO-based reinforcement learning can further refine downstream predictions. However, existing KD-to-RL pipelines typically rely on globally fixed transition schedules, ignoring that different samples may require different amounts of teacher-guided acquisition before reward-driven refinement. We propose PIVOT (Perplexity-Informed Transition Optimization), a dynamic transition framework that routes samples between OPD and GRPO according to teacher-evaluated sequence perplexity. PIVOT moves low-perplexity samples to GRPO for reward-driven refinement while keeping high-perplexity samples under OPD for continued domain knowledge acquisition. Experiments on Banking77 and HWU64 show that PIVOT consistently outperforms continued OPD and globally synchronized OPD→\rightarrowGRPO baselines under the same number of post-warm-up student optimization steps, achieving stronger downstream performance and more stable training dynamics.
Oct 8, 2026

CARE: Constrained Attention Refinement for Fine-Grained Visual Classification via Teacher-Student Distillation

Fine-grained visual classification requires models to recognize subtle local traits while exposing the visual evidence behind their predictions. Class-specific attention pathways provide a natural basis for interpretable recognition, but their constrained prediction structure limits discriminative capacity and underuses intermediate representations from strong pretrained backbones. To address this problem, we propose CARE, a constrained attention refinement framework for interpretable fine-grained recognition via teacher-student distillation. CARE keeps the final prediction and explanation within a class-specific attention student, while introducing a training-only auxiliary query teacher that reads selected intermediate DINOv2 layers with learnable queries. The teacher fuses multi-level representations and transfers logit-standardized class-discriminative knowledge to the student. To further refine the explanation pathway, we design diversity and sparsity terms to regularize student attention heads, reducing redundancy and encouraging compact trait localization. Experiments on CUB, Oxford-IIIT Pet, Stanford Dogs, and Stanford Cars show that CARE achieves strong classification performance under an interpretable frozen-backbone setting, reaching 78.5% Top-1 accuracy on CUB. Faithfulness analysis with insertion and deletion metrics further indicates that the top-ranked attention regions retain class-relevant evidence for explanation.
Oct 7, 2026

KDFP: A first-principles approach to knowledge distillation in large language models

Knowledge distillation is an established technique for improving the capabilities of small, efficient student models by training them with the representations of larger, more capable teacher models. Much of the recent work in the distillation of large language models (LLMs) has focused on distilling abilities learned during post-training, such as instruction following, chain-of-thought reasoning, and tool usage. This has left a large research gap in general knowledge distillation for LLMs, which is essential for developing efficient and private systems suitable for deployment on edge devices. We take a first-principles approach, evaluating previous lessons from prior works and conducting new explorations to develop a distillation methodology suitable for modern LLMs. We present KDFP, a novel methodology for white-box general knowledge distillation in LLMs. We demonstrate that KDFP outperforms existing methods by 1.6% −- 4.9% across 9 benchmarks while increasing training efficiency by up to 99.1% through ephemeral parameter reduction.