Overview

Self-distillation uses a model—or a related version of it—as a teacher, transferring knowledge through output probabilities, hidden representations, or generated experience. Common approaches include teacher–student consistency with exponential-moving-average or earlier-checkpoint teachers, on-policy learning from the model’s own trajectories, and privileged or hindsight feedback that supplies extra context during training. The methods are used for language-model reasoning and tool use, as well as vision, audio, video, time series, and faster generative inference; evaluation often spans math and code reasoning, agent tasks, and domain-specific perception benchmarks. A central challenge is that teacher signals can be biased or misaligned with task success, reinforcing errors, reducing exploration, or causing forgetting, so generalization, capability retention, training stability, and compute cost remain important open concerns.

Latest papers

151 papers23 papers in the last four weeks, down 23% on the four weeks before. 0.2% of all new papers.

Nov 12, 2026

Model Annealing: Data-Efficient Learning through Review and Consolidation

A small labelled dataset is memorized long before the training budget runs out, and the one-hot labels that remain to be fitted then mostly sharpen confidence. Model annealing spends the rest of the budget on review and consolidation, relearning from the network's own knowledge. Training is divided into episodes whose learning rate is raised again and cooled, every cooled state joins a memory of past selves, and each new episode learns from the labels, from the tempered predictions of this memory, and from adversarial dreams sought where student and memory disagree most. The memory averages predictions rather than weights, and the deployed network is the last student, so inference costs nothing extra. Across CIFAR-100, Tiny-ImageNet and Imagewoof, four architecture families and 5% to 100% of the labels, model annealing raises the top-1 accuracy of the AdamW-trained ConvNeXt, ViT and MLP-Mixer over cross-entropy training in 28 of 33 configurations, by 1.3 points on average, and is on par in accuracy with snapshot distillation, the closest competitor. On these families it is better than both in likelihood, calibration and robustness to corruptions and small adversarial perturbations. For an SGD-trained ResNet-18 it lowers accuracy but still improves robustness to small adversarial perturbations. Separating the reheated schedule from relearning shows that relearning helps every family and that the losses come from reheats, which undo most of this network's fit to its training set. How early the accuracy of a cross-entropy run saturates largely separates the AdamW families from ResNet-18 but does not predict the gain within them. With little data, relearning from the latest snapshot without dreams, as snapshot distillation does, compounds overconfidence in the AdamW families from episode to episode. The memory and the dreams each curb most of this drift, and together they remove it.
Nov 11, 2026

Hint-Informed Self-Distillation for Reasoning Segmentation

Reasoning segmentation requires identifying the target referred to by a query among visually similar objects, yet trajectory-level rewards cannot directly indicate which tokens provide relevant visual evidence. We propose Hint-Informed Self-Distillation (HISD), which uses target-disambiguation evidence for token-level credit assignment. An external model generates Segment Hints offline from training images, queries, and ground-truth target annotations, organizing target identity, discriminative attributes, confusable objects, and anchoring paths. A hint-conditioned exponential moving average teacher scores the student's own trajectories, and HISD computes the teacher-to-student evidence ratio over a token support set to adjust only the magnitudes of positive group-relative advantages, leaving nonpositive advantages unchanged. This design encourages concise reasoning while retaining the information needed for target disambiguation, and the student requires neither hints nor a teacher at inference. Under a unified training protocol, HISD improves ReasonSeg test gIoU and cIoU over our reproduction of DPAD by 3.00 and 3.55 percentage points, respectively. Across the five evaluated splits of ReasonSeg and the RefCOCO family, HISD achieves higher segmentation metrics than this baseline and reduces the average token count of valid reasoning spans by 80.17 to 91.83 percent.
Nov 9, 2026

ECOS: Environment-Calibrated On-Policy Self-Distillation for Language Agents

On-policy self-distillation lets a language agent act as its own Teacher by conditioning on privileged context such as skills written from its own trajectories. Its guidance, however, is weighted by how much this context shifts token likelihoods rather than by whether it improves behavior, so errors in self-written skills return through later updates. We propose environment-calibrated on-policy self-distillation, which uses environmental outcomes to decide how much to trust the Teacher's current guidance and which self-written skills to reinforce. Command-level advantages, estimated by comparing commands executed at similar observations, scale each Teacher preference and reward reflections and skills whose induced preferences favor higher-advantage commands. Without skills at inference, 3B and 7B agents reach 90.5% and 93.3% macro success on seen ALFWorld tasks and 75.5% and 80.5% success on WebShop. On unseen ALFWorld tasks, they outperform GRPO under the same protocol by 11.0 and 7.6 percentage points, respectively. In an ablation study, supervising only executable commands rather than the Actor's full response raises 3B WebShop success from 52.3% to 75.0%.
Nov 9, 2026

Speculative Self-Distillation enables Efficient Knowledge Internalization

Many real-world applications require language models to internalize knowledge that was absent from pretraining, to incorporate continually updated facts, or to learn new domain skills. Self-distillation has emerged as a natural approach for this task. A model conditioned on the source material acts as the teacher, and an unconditioned student is trained to reproduce the teacher's behavior closed-book. We propose Speculative Self-Distillation (SSD), an efficient per-token mixed-policy method for self-distillation. The student generates each rollout by default, while the teacher intervenes only at positions where its next-token distribution substantially diverges from the student's. Unlike off-policy distillation with teacher-generated prefixes, this design keeps training rollouts close to the student's test-time behavior, while avoiding the low-signal states that fully on-policy rollouts drift into. In effect, SSD lets the teacher steer the rollout exactly where the privileged context changes the prediction. On three closed-book QA benchmarks measuring fact recall, knowledge updates, and skill acquisition, SSD nearly matches on-policy distillation while using 38-69% fewer supervised tokens and consistently outperforms off-policy distillation on the held-out test set. On a forgetting stress test, SSD retains the most general capability while also achieving the best in-domain accuracy and using less supervision than the on-policy baseline. Overall, SSD establishes a new accuracy-efficiency-forgetting Pareto frontier for self-distillation.
Nov 1, 2026

Search-Aware On-Policy Self-Distillation with Sibling Alignment for Generative Recommendation

Generative recommendation retrieves items by autoregressively generating semantic IDs, typically via beam search. On-policy self-distillation (OPSD) enhances training supervision by transferring privileged self-teacher knowledge at student-generated states. However, beam-search retention depends on competition among prefixes: visited states differ in importance, and initially reliable decisions may approach the pruning boundary as training evolves. Moreover, conventional distribution matching couples target-token confidence with preferences among competing continuations, potentially weakening relational supervision when the self-teacher strongly favors the target. To address these challenges, we propose Search-Aware On-Policy Self-Distillation (SA-OPSD), a framework that transfers privileged self-teacher knowledge to the states visited during the student's own beam search. Specifically, we introduce Evolving-Boundary Sampling, which prioritizes search-critical states by considering both current pruning-boundary proximity and initial-reference retention-margin decline under a controlled distillation budget. Furthermore, we develop Confidence-Decoupled Branch Distillation, which separates target-probability calibration from informative sibling relations to better preserve the teacher's preferences among competing branches. Experiments on Industrial, Office, and Instruments demonstrate strong performance for SA-OPSD.
Oct 31, 2026

Steer, Don't Tell: Breaking Reasoning Plateaus with Reachability-Aligned Self-Distillation

Post training for large reasoning models remains insufficient on challenging problems, whose correct trajectories are hard to sample. Though prior works address this difficulty by integrating supervised fine tuning on expert trajectories with reinforcement learning, such external experts are not always available. Existing self improvement methods instead use reference solutions to create a privileged self-teacher. Yet privileged teacher correctness does not imply student reachability. Solution conditioning shifts the teacher distribution toward actions unlikely under the unprivileged student, producing seemingly "correct" trajectories that are unreachable or even mal-formed learning targets. Accordingly, we identify reachability alignment as the key to self constructed supervision and propose Self-Steered Reasoner (SSR). SSR realizes this principle through student support projection, which transforms the privileged teacher distribution into a directional signal only over reasoning actions locally reachable under the unguided student. Verification then retains correct trajectories sampled from this reachable space. To further boost the sampling efficiency, an iterative refinement framework is adopted, using diagnostic guidance based on current trajectories. On problems with pass@64 = 0 even after extensive RL, SSR recovers correct and learnable trajectories for 74.6% without parameter updates. Fine tuning on these trajectories followed by resumed reinforcement learning improves aggregate accuracy by 4.9 percentage points on AIME25/26 and HMMT26. These results show that privileged guidance becomes effective self-improvement data when aligned with the reasoning actions the student can reach.
Oct 28, 2026

Knowledge-structured and evidence-aware semantic self-distillation for glomerular pathology foundation model

Visual foundation models learn rich pathological morphology but lack explicit concept-level supervision. Vision--language pretraining can provide this supervision but requires costly image--text pairs. Here, we introduce GloSAGE, a semantic self-distillation framework that uses lesion profiles predicted by a pretrained visual teacher to construct structured pathology descriptions. Reliability-guided selection identifies image--text anchors, while Knowledge-Structured Pathology Text Synthesis and Evidence-Guided Negative-pair Construction account for partially observed lesions. Image--text alignment trains projection layers while retaining the frozen visual encoder; an intermediate Semantic Transfer Layer supplies features for downstream tasks. Developed on GloMillion and evaluated across seven glomerulus- and case-level cohorts, GloSAGE supported zero-shot lesion recognition, image--text retrieval, hard-negative discrimination and concept-intervention analyses. It outperformed its source visual model in most lesion-recognition tasks and improved lesion grading, while supporting competitive case-level disease classification, injury grading and baseline renal function estimation. Ablations supported the combined contributions of knowledge-structured text and evidence-guided negatives, and the intermediate transfer layer improved mean lesion-recognition ROC-AUC relative to the complete projection output. A frozen language model conditioned on GloSAGE features generated glomerular descriptions. These findings suggest that teacher-derived lesion profiles can provide annotation-efficient semantic supervision for extending a glomerular visual foundation model.
Oct 28, 2026

What Can Firms Do in the Face of Model Distillation?

Model distillation has become a major concern in the rapidly growing generative AI discourse. In its simplest form, distillation is the phenomenon where smaller firms can use data distilled from transcripts of conversations with powerful proprietary closed-weight models to train their own models, thus achieving reasonably high model quality but at much lower costs. Besides the legitimate safety and security concerns associated with distillation, another major issue is how it can also distort marketplace outcomes. This motivates us to initiate an exploration of the problem from an economic theory lens. We consider a setting with two players: an incumbent and a distilling competitor and characterize how the incumbent's model characteristics drive the incentives around how much the competitor should distill. We also consider the impact of distillation on key market outcomes like user-side surplus and the incumbent's utility. Our key contribution, however, is to study interventions for preventing or limiting distillation attacks. In this context, we study two settings: i) when it is possible, with a certain probability (which grows in the extent of distillation conducted), to detect an attack on the incumbent's model; and ii) when a distillation attack cannot be reliably detected. In the first setting, we show that penalty-based regulations can be very effective in completely deterring distillation, although it can have potentially adverse effects on market competition and user surplus. In the latter setting, we show that the incumbent can use self-distillation as a mitigation strategy. Importantly, the effectiveness of self-distillation depends on how cost-efficiently the incumbent can deploy a new self-distilled model on the market. On the positive side, self-distillation can preserve market competition thus protecting user surplus to some extent.
Oct 27, 2026

Uncovering Shortcut Amplification in Video RLVR through the Lens of Perceptual Uncertainty Collapse

Reinforcement learning with verifiable rewards (RLVR) has emerged as an effective post-training paradigm for enhancing the video reasoning capability of Large Multimodal Models (LMMs). Despite its empirical effectiveness, we show that outcome-level optimization can introduce an undesirable shortcut amplification effect, whereby models increasingly favor text-prior answers even when they are not supported by the visual evidence. To characterize this behavior, we investigate the training dynamics of video RLVR through the lens of perceptual uncertainty, analyzing how predictive uncertainty evolves under video-conditioned and text-only contexts. Our analysis identifies a perceptual uncertainty collapse phenomenon, in which RL training continuously sharpens the model's answer distribution while the uncertainty reduction attributable to visual evidence progressively diminishes. A further gradient-direction analysis reveals that such confidence enhancement becomes increasingly less aligned with visually grounded uncertainty reduction and is instead dominated by text-prior compression. Motivated by these findings, we propose Bifocal Self-distillation Policy Optimization (BSPO), a self-distillation framework that regulates policy updates with two complementary distributional signals. Specifically, BSPO derives empirical reasoning principles from intra-group correct and incorrect rollouts to identify reasoning-sensitive tokens, and introduces shortcut-aware visual perturbation to penalize predictions that remain invariant when task-relevant visual evidence is weakened. By adaptively reweighting the original RL advantage, BSPO preserves the task-solving benefits of RLVR while promoting updates that are both reasoning-sensitive and visually grounded. Extensive experiments validate the effectiveness of BSPO in improving video reasoning performance and strengthening visual faithfulness.
Oct 27, 2026

Self-Regularized Flow Language Models

We introduce SERF (Self-Regularized Flow Language Models), a simple self-supervised framework for improving continuous diffusion language models by exploiting the denoising trajectory itself. While continuous flow-based language models offer a promising alternative to autoregressive generation, their training objectives typically rely on supervision from clean target tokens, leaving the denoising trajectory underexploited as a learning signal. SERF addresses this limitation through noisy-to-clean self-distillation: an exponential moving average (EMA) teacher receives a cleaner version of the student's noisy input along the same interpolation path and provides additional supervision. We investigate two variants: SERF-L, which aligns the student's intermediate representations with deeper teacher representations, and SERF-P, which aligns their output token distributions, either at the student's output or at an early student layer. All variants augment the standard training objective without requiring external pretrained models or modifying the inference procedure. Experiments across three continuous language-model paradigms (one-hot FLM, latent ELF, and hyper-spherical S-FLM) demonstrate substantial improvements in training efficiency, generation quality, and few-step generation. In particular, SERF achieves up to 7.7x wall-clock training acceleration on OpenWebText, improves MAUVE by 24.3% and GSM8K accuracy by up to 28.8%, and matches baseline performance with up to 64x fewer sampling steps. These results establish noisy-to-clean self-distillation as an effective approach to improving continuous diffusion language models.
Oct 27, 2026

WithinVAE: Compact Visual Autoencoding with Self-Distilled Semantics

Visual autoencoders must balance reconstruction fidelity, compactness, and semantic structure. Existing semantic tokenizers typically inherit semantics from pretrained recognition encoders, whose invariances can discard the appearance detail required for inversion. We show that this inheritance is unnecessary: semantic structure can be learned inside an autoencoder's own reconstructive representation. We introduce WithinVAE, a compact variational visual autoencoder with a sequential bottleneck, DINO/iBOT-style self-distillation from an exponential-moving-average copy of the autoencoder itself, and position-dependent token masking to achieve a 4-fold compression ratio, which is free of pretrained representation encoder or external semantic target. On ImageNet-1K at 256×256256\times256, WithinVAE improves all four reported reconstruction metrics over RPiAE including PSNR by 1.36 dB, while using half its latent scalar capacity. Under frozen evaluation, the same latents improve linear probing by 6.5 points over RPiAE and 5.1 points over our reconstruction-only baseline, attributing the semantic gains to self-distillation rather than architecture or external supervision. An auxiliary latent-to-latent model reproduces the ordered-compression trend, and a pixel-to-latent model transfers the objective to a pretrained generative space. These results separate the origin of semantic supervision from its use: a compact visual autoencoder can organize semantic structure without borrowing a recognition representation.
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 25, 2026

Spatial Audio Generation and Editing with Executable Plans and Self-Distilling Transfusion

Spatial audio authoring coordinates source decisions with soundfield synthesis. For editing, plans specify what to change and retain, while reference audio supplies further acoustic detail. We present AMBIT, a unified framework for native First-Order Ambisonics (FOA) generation and editing with three connected contributions. First, a shared Transfusion backbone couples autoregressive planning with rectified-flow rendering through executable ScenePlans that specify source content, activity, and 3D trajectories. For editing, the plan describes the full target scene, including retained sources; contrastive features ground the requested change, while full reference latents supply residual acoustic detail for preservation. Second, an FOA-aware continuous VAE combines shared-decoder omnidirectional recovery, asymmetric grouped KL, and spatial covariance matching to preserve content and inter-channel structure. Third, joint Transfusion on-policy self-distillation (OPSD) refines planning on student-generated prefixes and rendering at sampling states reached under student plans, with target-conditioned teachers supervising both modes of the shared student. Evaluations across speech, music, and sound show improved generation content and spatial quality, acoustic fidelity during instruction-guided editing, and unedited-source preservation over the compared baselines. Project page: https://wjm-george.github.io/AMBIT/
Oct 24, 2026

Distilling Dense Trajectory Teachers For Large-Motion Video World Models

Video world models are expected to generate videos that accurately follow specified camera trajectories while maintaining consistent 3D structures. However, existing models are predominantly trained on densely sampled videos with small inter-frame camera motions. When applied to sparse, large step trajectories, they often produce frozen frames, inaccurate camera motion, and severe geometric drift. We attribute this degradation to a mismatch between the gradual viewpoint transitions seen during training and the large viewpoint changes required at inference time. To address this problem, we propose Dense Trajectory Teacher Distillation (DTTD), a post-training framework that transfers a model's generation capability from dense trajectories to large camera motions. Given a sparse trajectory, DTTD interpolates additional camera poses to construct a dense trajectory covering the same path. The frozen base model then generates reliable teacher videos along this familiar dense trajectory, and pose aligned teacher observations are used to evaluate candidate videos generated under the original sparse trajectory. In this way, the model serves as its own generation teacher, requiring neither additional video data nor a stronger external video generator. We further combine dense teacher supervision with geometry aware rewards that assess cross frame consistency and camera trajectory following, and use the resulting preferences for offline preference optimization. To stably expand the model's motion range, we introduce a cumulative curriculum that progressively incorporates larger motion steps while retaining simpler trajectories to reduce capability forgetting.
Oct 21, 2026

Self-Distilling Future Video Representations for World Action Models

World Action Models (WAMs) have emerged as a promising approach to learning generalist robot policies using rich spatiotemporal priors from pre-trained video diffusion models. A recent line of WAMs improves inference efficiency by conditioning a separate action decoder on intermediate future video representations extracted from a single denoising step, avoiding costly iterative denoising required for video generation. This design makes action prediction depend directly on the quality of these future representations. In this paper, we introduce Future REpresentation learning via self-Distillation with temporal context (FRED), a framework that trains WAMs to improve the quality of future video representations from noise in a single denoising step. Specifically, FRED uses self-distillation to predict future representations from noise, using representations extracted from ground-truth future observations as targets. We further condition the video model on historical observations, providing temporal context for predicting future representations better. Across diverse real-world manipulation scenarios, FRED consistently outperforms competing WAMs. It achieves particularly large gains on long-horizon and out-of-domain tasks under zero-shot environment generalization, and substantially improves human-to-robot skill transfer, reaching 50.0% task success on tasks absent from robot training data and demonstrated only in human videos, compared with 19.4% for the best competing WAM.
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 20, 2026

Self-Evolving Agents through Privileged Retrospective Distillation

Self-evolving agents aim to improve from their own experience rather than remain fixed after deployment, but naturally collected trajectories often lack ground-truth solutions or verifier-derived rewards. We introduce Retrospective Self-Distillation (RSD), a multi-round framework that turns such experience into dense supervision using privileged retrospective context. For each on-policy trajectory, the same underlying model acts as both student and teacher: the teacher re-scores the student's realized actions while conditioning on cross-task precedents and lessons, together with turn-level reflection and future hindsight from the completed trajectory. The resulting teacher distribution is distilled back into the policy through on-policy self-distillation. Across rounds, the experience bank grows and the retrospective context is refreshed, allowing the model, its data, and its supervision to co-evolve. We further introduce frontier-guided context-policy evolution, where a stronger frontier model proposes constrained edits to how privileged context is constructed. Across ALFWorld, ScienceWorld, and SearchQA with Qwen3-8B and Qwen3-14B, RSD improves performance by up to 59.7% relative to the base model, while frontier-guided context-policy evolution adds up to a further 44.7% relative improvement over a fixed context policy. These results show that naturally accumulated experience can be converted into persistent model capability without ground-truth trajectories or verifier-derived rewards for the core policy update.
Oct 19, 2026

FLMNet: Self-Ensemble Distillation from Masks to Seeds for Single-Tower CTR Prediction

Click-through rate (CTR) prediction underpins recommender systems and online advertising, where high-dimensional categorical features interact in combinatorially many ways. Several recent CTR models, such as FinalNet and QNN, report higher accuracy when a network also learns from its own averaged predictions, for example by running the same network under different dropout patterns and training each run toward the average. Why this self-distillation helps, and whether its random sampling is needed, remains unclear. This work shows that as the number of dropout patterns grows without bound, the objective reduces exactly to the label loss of the logit-averaged prediction plus a penalty on how much the dropout copies of the network disagree. When dropout does not affect how a network gates its features, this penalty can be approximated in closed form, and FLMNet is designed on this principle and trains without sampling. Networks trained from different seeds disagree in the same form, and distilling several of them into one network recovers part of their ensemble gain. On five benchmarks against seventeen baselines, the resulting models take the top two places in accuracy with a single network at inference, and the gains of the penalty and of distillation, where measurable, largely add up.
Oct 18, 2026

Self-Distillation for Supervised Fine-Tuning

Supervised fine-tuning treats the expert demonstration or gold response as a one-hot target, suppressing valid alternative continuations that the model would otherwise assign probability to. We propose Gold-Conditioned Distillation (GCD), a method that turns expert demonstrations into model-native supervision. GCD uses one model as both teacher and student: the teacher conditions on a prompt containing the gold response in context, the student conditions on the original prompt as at inference, and the student is trained to minimize the forward KL divergence to the teacher's next-token distribution at each gold response token. This lets the model use its prior knowledge to teach itself the new task: correctness is anchored by the gold response, while the soft target preserves plausible alternatives. Across various training tasks on Qwen and Llama models, GCD improves new-task accuracy over SFT and generally reduces degradation on general benchmarks.
Oct 17, 2026

DiPLO: Self-distilled Diffusion Planning for Loco-manipulation

Humanoid loco-manipulation is particularly challenging because it requires coordinating locomotion, balance, and object interaction, making it difficult to learn from limited demonstrations or from scratch through Reinforcement Learning (RL). We introduce DiPLO (self-distilled Diffusion Planning for Loco-manipulation), a framework that combines a self-distilled diffusion planner with a low-level whole-body tracking policy. Our planner is pretrained via supervised learning on motions derived from human-to-humanoid retargeting, while the tracker is trained via RL to execute them. DiPLO centers around two complementary post-training stages. First, we introduce Self-Distillation Diffusion Policy Optimisation (SD:DPO) to remove the planner's dependence on strong task conditioning, which we identify as a crucial ingredient for effective pretraining, as downstream fine-tuning does not recover comparable performance when it is omitted.. We then fine-tune this distilled model with Diffusion-Augmented Proximal Policy Optimization (DA:PPO), a maximum-entropy RL method for diffusion models, to improve task success and align generated plans with the frozen tracker's execution capabilities. Together, these stages enable motion generation through state-history and goal conditioning alone, enabling 99% task success in pick-carry-place tasks, compared to just 45% for the pretrained model. The resulting planner also generates motions for unseen object goals, with further evaluations on kicking and pushing/pulling tasks. Finally, we deploy our pipeline on the hardware, using our DA:PPO planner-tracker on a Unitree G1, achieving repeated successful executions without any real-world fine-tuning.
Oct 17, 2026

AtomSD: Procedural Atom-Guided Self-Distillation for LLM Agents

Large language models (LLMs) possess broad factual knowledge and strong reasoning capabilities, but solving complex tasks also requires procedural knowledge about how to act. Existing approaches often represent procedural knowledge through trajectories or bundled skills, limiting flexible reuse across interaction histories. This paper introduces Procedural Atoms, fine-grained and independently applicable units that represent Intent, When, Do, and Avoid guidance. Building on this representation, we propose AtomSD, an Atom-guided Self-Distillation framework in which the learner continually extracts Procedural Atoms from its own successful trajectories and distills them into the evolving policy through Atom-conditioned supervision and task-reward optimization. Comprehensive experiments show that AtomSD consistently outperforms strong reward-based and distillation methods, exceeding the state-of-the-art baseline by up to 9.31 points in average performance. Further analyses demonstrate that the proposed Procedural Atoms provide effective guidance, highlighting the value of atom-level procedural knowledge for policy learning.
Oct 15, 2026

Progressive Misalignment Can Emerge from Benign Recursive Self-Distillation

Training models on their own outputs has attracted widespread attention as a possible route to advance AI capabilities. One prominent approach is self-distillation, where a model is trained to match the token-level predictions it would have produced when given access to additional context, such as a cue, feedback, or instruction. However, the safety properties of this training procedure are poorly understood. In this paper, we present a surprising safety failure of recursive self-distillation, where each self-distilled model becomes the teacher for the next round of training. We observe that when a safe model is recursively self-distilled on its own (safe) generations while receiving an additional non-harmful cue, it becomes increasingly misaligned across a diverse set of harmful queries; we call this Progressive Misalignment. We show this failure mode in both the Qwen3 and Phi-4 model families on SORRY-Bench and StrongREJECT. Progressive misalignment also occurs in a more realistic setup, where the fixed cue is replaced by feedback from a simulated user with benign intent. Using a gradient-based method, we are also able to find benign prompts that lead to progressive misalignment when applying recursive self-distillation, showing that harmful training prompts are not necessary for this failure mode to emerge. We hope our work encourages future investigations of the safety properties of AI systems recursively trained on their own data.
Oct 14, 2026

More Methods, Better Agents? A Reality Check on Self-Distillation in Agentic Reinforcement Learning

Self-distillation in agentic reinforcement learning has produced a growing number of methods. However, existing evaluations often reuse published baseline scores under differing experimental conditions, provide few direct comparisons among contemporaneous methods, and give limited attention to training stability and computational cost. These limitations obscure the source of reported gains and the practical strengths and weaknesses of competing methods. Costly agent training further motivates a common basis for comparison. To address these gaps, we compare ten representative methods against GRPO on ALFWorld, ScienceWorld, Search-QA, and WebShop across three model sizes. Our study comprises 132 training runs totaling 12,124.4 recorded H800 GPU-hours, with repeated evaluations over eight rollout seeds for the pretrained backbones and all eleven trained methods. Beyond final task performance, we examine training progress, stability, and measured computational cost. Our evaluation yields three key findings. First, gains over GRPO remain limited in scope, with no method consistently outperforming it across settings. Second, strong final performance can mask unstable training. Third, additional training cost does not consistently translate into better performance. We establish this benchmark to support fair, comprehensive comparison and guide future research toward reliable and cost-effective improvements in agent training. Project page: https://aim-uofa.github.io/Self-Distillation-AgentRL/
Oct 14, 2026

Function-Space Temporal Self-Distillation for Policy Optimization

Reinforcement Learning with Verifiable Rewards (RLVR) improves reasoning in language models, but scalar outcomes provide coarse supervision over response trajectories. Self-distillation offers denser token-level guidance by comparing an unassisted policy with a feedback-informed self-teacher, yet the teacher evolves with the student, creating unstable targets and motivating additional regularization. We introduce Function-Space Policy Optimization (FSPO), a single-policy framework that combines current privileged adaptation with function-space temporal consolidation. FSPO stores sparse historical next-token distributions as fixed temporal targets, providing temporal separation without maintaining a persistent teacher model. It further introduces privileged disagreement correction to explicitly suppress locally overvalued actions, reliability-adaptive grounding that strengthens reference supervision when self-generated evidence repeatedly fails, and quality and coverage-aware temporal memory that retains verified behaviors and improves privileged demonstrations. Across reasoning and tool-use benchmarks, FSPO achieves the highest average accuracy under a 5-hour training budget, improving over the strongest baseline by approximately 4% points on average while reducing peak memory from about 70 GB to 56 GB on Olmo3-7B-Instruct.
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

Teach Me After You Learn It: Lookahead Self-Distillation for RLVR

On-policy self-distillation (OPSD) has emerged as a popular training paradigm for LLMs, providing dense, token-level supervision by aligning a model's own distribution with the distribution induced by privileged information. However, existing OPSD methods derive supervision solely from privileged information of the same instance, leaving inter-instance supervision largely unexplored. We propose Lookahead Self-Distillation (LEAD), which introduces inter-instance distillation by allowing the teacher to first learn from other on-policy rollouts before teaching the student. LEAD constructs lookahead teachers through virtual RL updates on disjoint subsets of on-policy rollouts, and uses each teacher to provide cross-group token-level supervision for the other subset. To reconcile the complementary yet potentially conflicting signals from OPSD and RLVR, we introduce harmonious modulation, which adaptively modulates the GRPO advantage based on the agreement between the two signals. Extensive experiments on scientific reasoning and interactive-agent benchmarks show that LEAD achieves the best in-domain average performance across model scales while maintaining out-of-domain generalization comparable to GRPO.
Oct 13, 2026

Training Generalist Navigation Policies via Action Self-Distillation

Language provides a natural interface for humans to specify and reason about goals, but training language-conditioned robotic policies and reasoning typically requires expensive manual annotations. We propose Action Self-Distillation (ASD), a method for learning language-conditioned reasoning for robotic navigation without human-annotated language or reasoning labels. ASD uses an open-source VLM, which contains broad world knowledge and reasoning capabilities, as both a teacher network and a student policy. During training, the self-teacher receives privileged trajectory feedback and infers the underlying navigation goal together with goal-oriented reasoning. The student is then trained to condition on the inferred navigation goal, predict future ego-frame trajectory, and provide a description of the scene semantics, plan, and reasoning, from onboard observations alone. Across 13 challenging indoor and outdoor real-world navigation tasks, our method improves success rate over the strongest baseline by 30%30\% in indoor settings, 30%30\% on multi-stage tasks, and achieves 95%95\% success rate on outdoor tasks, validating that privileged trajectory feedback can enable self-distillation for scalable generalist navigation policies without human annotations.
Oct 10, 2026

Decoupling Knowledge Forgetting and Refusal Learning in MLLM Unlearning

Current evaluations of multimodal unlearning typically focus on whether a model stops outputting target information. This surface-level criterion, however, masks two critical failure modes. First, knowledge suppression techniques often induce confabulation, leading the model to hallucinate facts about the unlearned subject. Second, refusal training merely alters observable behavior without erasing the underlying knowledge; such models might decline open queries but still identify correct answers under multiple-choice probing. To overcome these limitations, we propose Forget-Aware Self-Distillation (FASD). This method is designed to achieve honest ignorance, a state where the model genuinely loses the target knowledge and learns to acknowledge its absence. FASD separates these goals into two complementary channels. The behavior channel distills appropriate refusals from an ignorance-prompted copy of the model itself. Crucially, target tokens are masked from the teacher distribution to avoid accidentally reteaching the forgotten concepts. Simultaneously, the knowledge channel directly suppresses ground-truth answer spans independently of the model's generated responses. Experiments demonstrate that FASD successfully unlearns target knowledge, prevents information leakage in open-ended tasks, and preserves performance on retained data. The resulting model not only forgets the specified information but accurately recognizes the boundaries of its own knowledge. Code and data will be released upon publication.
Oct 9, 2026

A Privileged Teacher That Only Scores Cannot Search: Critique-Conditioned On-Policy Self-Distillation for Reasoning

On-policy self-distillation promises to replace the sparse outcome signal of verifiable-reward RL with dense token-level credit: a teacher conditioned on the reference solution scores the student's own rollout, and the resulting distributional shift is distilled back into the policy. We investigate whether this scoring signal actually contains corrective credit. Using gold error spans in student rollouts, we test the two signals consumed by self-distillation objectives, the full-distribution KL and the token log-ratio, for two properties: whether they localize errors and whether they push probability away from erroneous student tokens. They exhibit neither of these properties nor can they reliably distinguish correct from incorrect rollouts within the same problem. We argue that this failure is structural: locating an error in a reasoning trace is itself a search problem, while a scoring teacher is never explicitly asked to perform that search. The same model, with the same privileged information, localizes errors substantially better when allowed to generate a critique. This forms the motivation for critique-conditioned self-distillation: first generate a critique of the current rollout, then condition the scoring teacher on that critique. The resulting signal localizes errors, is corrective within error spans, and shifts probability toward tokens that initiate backtracking such as Wait'', But'', and ``Actually''. We further show that the critique is useful both at test time, where resuming a wrong rollout with the critique in context converts math problems the model cannot solve by sampling, and in cold-start GRPO, where it provides an advantage signal on prompts for which verifiable rewards alone provide none.
Oct 9, 2026

Continually Post-training Small Models via Synthetic Self-distillation

Continual post-training faces a dual challenge of acquisition and retention: models must learn new capabilities while preserving prior knowledge. While Self-Distillation Fine-Tuning (SDFT) mitigates this using self-supervision with in-context demonstrations, it relies on strong in-context learning and instruction-following abilities. Consequently, SDFT fails on smaller models critical for low-latency, on-device use. To enable continual learning at smaller scales, we first identify the root cause of failure. We show that poor acquisition and forgetting correlate linearly with the post-trained model's distribution shift, revealing the necessity of in-distribution learning. We thus introduce Synthetic Self-Distillation (SynthSelf), imitating SDFT's in-distribution rewrite by outsourcing the task to a minimally larger, same-family aligner. Coupled with off-policy supervision and rejection sampling, this provides the high-quality supervision small models require without breaking their original distribution. Evaluations spanning 4 acquisition datasets (including a mixture) and 14 retention benchmarks show SynthSelf consistently outperforms 6 methods (including off- and on-policy distillation) across 3 small model scales. Extensive further analyses confirm SynthSelf paves the way toward effective small-model continual learning.