Distillation 进展

Recent text-to-motion models have improved motion quality and instruction following, yet many-step denoising and large model components make deployment slow and memory-intensive. We present Terminal-Amplification-Controlled Distillation (TACD), an on-policy approach for training efficient motion generators from text prompts and pretrained teachers, without real-motion training data. Building on segmented on-policy flow distillation, we supervise clean-motion predictions along student-generated trajectories. We identify a failure mode in which velocity matching on a fixed supervision grid repeatedly overweights errors near the denoising endpoint, degrading few-step generation. TACD ties the latest teacher query to the student's step size, bounding the effective loss weights in clean-motion space without changing inference. Experiments on HumanML3D and KIT-ML demonstrate improved few-step generation, including a 58% reduction in eight-step HY-Motion student FID relative to distillation without this bound. For diffusion teachers, the endpoint-matching form of TACD yields four-step students with lower FID and matched or improved text-motion retrieval relative to their 50-step teachers on HumanML3D. On HY-Motion and Kimodo, eight-step students with compact components achieve 7.7-11.9x end-to-end speedups and reduce peak GPU memory by 3.8-6.7x relative to their teachers. Project page: https://vkgo.github.io/TACD/

Few-step neural text-to-speech models often rely on short- ened diffusion or flow-matching schedules, or on distillation from pretrained multi-step teachers. To avoid these depen- dencies, we present DriftTTS, a few-step mel-spectrogram generator trained without a generative teacher, distillation, or adversarial discrimination. DriftTTS uses a distribution- matching drift objective in a mel-domain feature space defined by raw mels and a frozen masked-autoencoder encoder pretrained on the same LJSpeech training split. On-policy rollout trains the decoder on its own interme- diate states and supports inference up to the trained roll- out depth. On LJSpeech, DriftTTS at NFE=4 achieves 3.87 dB MCD and 3.7% WER, compared with 3.85 dB and 3.4% for Matcha-TTS. In a fully paired blind listen- ing test, DriftTTS obtains 4.18 MOS, compared with 3.96 for Matcha-TTS and 4.22 for ground truth. These results demonstrate competitive few-step synthesis without a pre- trained generative teacher. Code can be found at https: //github.com/BASHLab/driftTTS.git

Streaming video generation has benefited from distribution matching distillation (DMD), which matches the joint distribution of video frames to a video teacher's approximation of the real video distribution. Although this joint matching mitigates drift during autoregressive rollouts, limitations remain in visual quality and semantic alignment. To address these limitations, we propose DuoMatching, a distribution matching framework that approximates the real video distribution through a unified joint-marginal formulation. On top of existing joint matching formulations, the additional marginal matching objective provides dedicated frame-level supervision from an image generator, transferring complementary visual and semantic priors from it. To apply this frame-level supervision in video generation, we introduce LatentBridge to resolve the latent representation mismatch between the video student and the image teacher. Latent Variation Sampling further distributes such frame-level supervision across distinct temporal segments, reducing redundancy. Experiments demonstrate that DuoMatching improves visual quality, composition, and semantic alignment while largely preserving motion dynamics. Human evaluations show overall preference rates above 80% against all evaluated baselines. The project page is available at https://johnzhan2023.github.io/DuoMatching/.

Video creation spans text-to-video (T2V), image-to-video (I2V), and condition-based generation, yet video diffusion models remain costly because they repeatedly evaluate large backbones during sampling. Distribution matching distillation (DMD) reduces this cost, but its reverse Kullback--Leibler (KL) objective can provide unstable or incomplete guidance when the student and teacher distributions have limited overlap. VDOT addressed this issue by adding optimal transport distillation (OTD), whose explicit coupling supplies geometric directions for condition-based generation. Balanced OTD, however, performs full-mass matching between the spatial tokens of each corresponding student--teacher frame pair. This assumption weakens for T2V and I2V, where one condition admits many valid outputs and spatial content need not align across different realizations. We present VDOT++, a unified distillation framework that applies the same training recipe separately to generators for the three task families. It makes OTD robust to output diversity through an asymmetric unbalanced formulation that allows unreliable student tokens to carry less mass while maintaining coverage of the teacher tokens. An $\ell_1$ ground cost further replaces mean-based aggregation with a more mode-preserving weighted median that limits the influence of distant transport targets. The two changes respectively determine whom to match and how the selected targets should be aggregated. We additionally combine distribution matching and adversarial refinement through sequential backward passes, and exploit the decoupled score networks for cross-scale distillation, where larger score networks improve a compact generator. Experiments on UVCBench, VBench, VBench-I2V, and the VACE benchmark show that the resulting four-step generators are competitive with many-step teachers and strong few-step baselines across all three task families.

Masked diffusion language models (dLMs) offer a promising parallel alternative to autoregressive models for complex reasoning. However, they face a distinct credit-assignment challenge, since a few commitments during denoising sharply reduce the uncertainty over the remaining masked positions and shape much of the response. Most post-training recipes for dLMs do not use this signal to decide which tokens to train on: they typically train on the final text or assign rewards to whole denoising steps, rather than selecting the individual commitments that shape the response. We introduce Pivot-SD, an efficient offline self-distillation framework that supervises only these high-impact commitments (pivots). Pivot-SD selects pivots using an information-gain metric measuring uncertainty reduction over the remaining masked positions. Pivots from successful trajectories are trained with cross-entropy, and pivots from failed trajectories with targeted unlikelihood, leaving the rest of the failed trajectory untouched. Using only 200 questions and four rollouts each, Pivot-SD improves LLaDA-8B-Instruct over full-sequence SFT and budget-matched diffusion RL baselines across math and code benchmarks.

Masked discrete diffusion models offer a promising alternative to autoregressive generation, but iterative sampling can be costly, and intractable sequence likelihoods complicate reward fine-tuning. We introduce IDRF, a framework for reward fine-tuning of few-step masked discrete diffusion generators. Starting from a standard reverse-KL-regularized objective, IDRF replaces the intractable sequence-level KL penalty with inverse-distillation regularization. With an optimal auxiliary denoiser, we prove that the population inverse-distillation loss upper-bounds the sequence-level KL divergence to the reference distribution. IDRF optimizes a trajectory-based surrogate of this loss without reference-model rollouts, so the student keeps its own few-step sampler. We view few-step generation as a finite-horizon Markov decision process and optimize reward with a clipped policy-gradient objective over the student's trajectories. Across DNA, image, and text generation, IDRF achieves high reward with up to $32\times$ fewer denoising steps than the reference while mitigating reward hacking and preserving sample quality.

Language model agents are deployed with a harness, the software around the model that manages its context, tools, and feedback. When such an agent is distilled into a smaller one, the harness stays in place, so the student mainly needs the teacher-specific abilities that the harness cannot provide, such as acting correctly on harness information. Standard distillation, however, imitates the teacher's full outputs and treats the harness as part of the input. We propose Harness-Aware Distillation (HAD), which focuses distillation on what the teacher adds beyond the harness. HAD complements on-policy distillation with two components: an action preference that contrasts the same teacher's actions with and without the harness information, scored after the student's own reasoning, and a validity check that drops preference pairs whose preferred action contradicts the harness records. We show that the contrast gives the student information that imitating the teacher alone cannot provide, and HAD needs no task rewards, success labels, or future information. Across multiple long-horizon agent benchmarks and models, HAD outperforms on-policy distillation baselines with the same fixed harness. Our analysis shows that HAD enters fewer unproductive loops and recovers from errors more often than the baselines, and suggests that it adaptively keeps learnable feedback in its weights while reading state information from the harness.

Diffusion Transformers (DiTs) can generate high-quality images and videos, but generating each sample requires multiple costly DiT forward passes. Two common ways to accelerate DiT sampling are step distillation, which reduces the number of sampling steps, and caching, which skips some DiT evaluations by reusing a tensor computed at an earlier step. Most caching methods decide in advance which tensor to reuse. After distillation, adjacent sampling steps are farther apart. Reusing a tensor across this larger gap introduces more error, so choosing what to cache becomes especially important. We therefore introduce AutoTarget, a method that chooses the cached tensor for a given model, solver, and reuse schedule. AutoTarget uses a small set of runs without cache reuse to measure the error caused by reusing each candidate tensor, then selects the candidate with the lowest error. We also analyze how an error at one reuse step affects the final sample. For Euler sampling, we identify cache targets that produce the same trajectory and show why a stored solver update may not. Experiments on distilled image and video DiTs show that the best cache target changes with the model, image resolution, and solver. AutoTarget reduces DiT evaluations and retained cache storage. Generation quality remains close to the corresponding uncached run. On the tested PixArt-LCM and FLUX.1-schnell settings, its calibration ranking matches the ranking from held-out cached runs. To help others reproduce the method, we provide its core implementation on GitHub at https://github.com/wali1024-offical/AutoTarget.

On-policy distillation (OPD) combines student-generated rollouts with dense token-level supervision from a teacher, but providing such supervision for every rollout requires substantial teacher computation. We introduce Success-Referenced On-Policy Distillation (SR-OPD), which reduces this cost by selecting which prompts and rollouts receive teacher supervision. When the student produces both successful and failed rollouts for the same prompt, a successful rollout can serve as a natural reference for selecting failed rollouts. SR-OPD therefore focuses on such prompts and prioritizes failed rollouts whose hidden-state trajectories show sustained divergence from a successful reference, while accounting for estimated teacher-input cost. Across three teacher-student pairs and six mathematical reasoning benchmarks, SR-OPD uses only 3.46-5.02% of the teacher-input tokens required by Vanilla OPD in the one-pass setting while maintaining comparable reasoning performance. Under a controlled setting matched to 5% of Vanilla OPD's teacher-input budget, further experiments support both key design choices: focusing supervision on prompts with both successful and failed rollouts, and using successful rollouts to guide failure selection. These results indicate that a student's own successful behavior can serve as a useful reference for allocating teacher supervision under a fixed teacher-input budget.

On-policy distillation (OPD) improves large language model reasoning by training students on their own rollouts with dense token-wise supervision from the teacher. However, token-wise OPD does not explicitly provide a coherent alternative reasoning step showing how the student's step could be revised to improve subsequent reasoning. Furthermore, this paradigm can become less effective when the student produces a degenerate reasoning prefix, as subsequent teacher supervision remains conditioned on that prefix and may reinforce poor reasoning patterns. In this work, we focus on learning reasoning revision with segment-wise OPD to rework intermediate reasoning steps and better support subsequent reasoning. Through controlled reasoning interventions, we find that replacing student segments with teacher redrafts improves subsequent reasoning accuracy. Therefore, we address the problem of turning teacher redrafts into explicit supervision for learning to revise reasoning. We propose Segment-wise On-Policy Distillation (Seg-OPD), which selects student segments based on an uncertainty metric and obtains corresponding teacher redrafts. Seg-OPD trains the student to prefer teacher redrafts over their paired student segments while retaining dense token-wise OPD supervision. Extensive experiments on mathematical reasoning and competitive programming tasks show that Seg-OPD-trained students achieve higher revision success rates than baselines. Seg-OPD consistently outperforms the compared state-of-the-art baselines in reasoning accuracy with an average relative improvement of 5.22% across diverse models and tasks. Code is available at https://anonymous.4open.science/r/Seg-OPD.

Foundation multimodal large language models are designed to support a broad spectrum of capabilities across diverse domains. Multi-teacher on-policy distillation (MOPD) provides an effective framework for consolidating domain-specific expertise into a single student model. However, MOPD training gradually drives the student away from its initialization model, and general capabilities decline as the displacement grows, resulting in capability interference. A direct remedy is constraining the student toward its initialization, but this suppresses the acquisition of domain expertise as well. We propose Slow-Fast Multi-Teacher On-Policy Distillation (SF-MOPD), which couples a fast model, the current student updated directly by each teacher, with a slow model, an exponential moving average of the student. The slow model absorbs the learning signal gradually, serving as a moving capability reference that fuses the general foundation with confirmed domain expertise. For each teacher, SF-MOPD computes the teacher-induced update in log-probability space and removes only the component that pushes the fast model further away from the slow model, while retaining aligned and orthogonal components. Experiments across multiple model scales demonstrate that SF-MOPD effectively mitigates capability interference, enhances specialized multimodal capabilities, and reduces the average degradation on general-capability benchmarks, consistently outperforming vanilla MOPD.

Reinforcement learning from verifiable rewards (RLVR) usually optimizes answer correctness, yet useful language-model behavior also requires high-quality reasoning and concise responses. Existing multi-reward post-training methods typically scalarize rewards or combine specialists without explicitly protecting a reward priority order. This is problematic when trade-offs are asymmetric: conciseness, for example, should not improve at the cost of correctness. We introduce Lexicographic Multi-Objective On-Policy Distillation (LMOPD), a multi-teacher method for integrating reward-specialized policies under explicit priorities. For each student rollout, LMOPD selects the specialist for the first objective whose gate detects a deficiency, then locally projects its centered log-policy correction to remove components that oppose higher-priority specialists. We evaluate 30B-A3B mixture-of-experts transformer models in two- and four-expert settings on three math benchmarks, measuring retained specialist gains. With two experts, LMOPD's point estimates fully retain the accuracy and reasoning-quality gains while acquiring $46.9%$ of the conciseness gain. With four experts, it retains $\approx90%$ of both the accuracy gain and reasoning-correctness gain, compared to only $\approx57%$ by the next best evaluated baseline. Matched four-expertablations show that lexicographic routing outperforms random routing and that projection further strengthens both top-priority capabilities. Across both scales, LMOPD preserves the highest-priority capabilities more effectively than the existing baselines we evaluate, demonstrating the value of explicit priorities for specialist integration.

Vision foundation models such as DINOv2, SigLIP2, and MASt3R develop complementary capabilities from different pretraining objectives, yet their knowledge remains distributed across separate, specialized models. Multi-teacher knowledge distillation offers a path toward consolidating these capabilities into a single agglomerative backbone, but existing approaches assume a fixed set of teachers, and incorporating a new teacher requires repeating expensive joint distillation over the entire teacher set. We introduce GRAFT, a continual multi-teacher distillation framework that enables a unified backbone to progressively acquire capabilities from an open-ended sequence of foundation models. When a new teacher arrives, GRAFT treats the previously distilled model as a teacher for preserving learned capabilities, while the current student jointly learns from both the previous model and the incoming teacher. Furthermore, to reconcile the incompatible representation geometries of heterogeneous teachers, we introduce Teacher Specific Readout Tokens, which grant each teacher an independent read-out of the shared encoder, together with Geometry Agnostic Relational Loss that aligns a vision-language teacher by matching image-text similarity structures rather than raw feature values. We provide GRAFT model, which is a single, continually extensible backbone that unifies five domains, including image understanding, 2D dense prediction, 3D human pose estimation, 3D vision, and vision-language, delivering strong performance across all of them while acquiring each new capability at the cost of a single distillation rather than a full re-distillation.

Flow-map distillation enables one- and few-step generation by learning finite-time transitions of a pretrained generative ODE. We investigate whether changing the teacher's time parameterization can make these transitions easier to learn. Motivated by the hypothesis that trajectory segments with large normal acceleration are harder to distill, we propose a geometry-aware time reparameterization that allocates more student time to these regions while preserving the teacher's geometric paths and terminal distribution. We derive a shared clock that equalizes a population normal-acceleration statistic under suitable assumptions, and construct a practical approximation from robust, regularized estimates across teacher trajectories. We incorporate this clock into Lagrangian flow-map distillation, using the transformed time coordinate to condition the student. The clock is estimated once before distillation and requires neither teacher retraining nor additional student parameters or inference-time network evaluations. Experiments on synthetic data, CIFAR-10, and CelebA-64 show improved sample quality over identity-time distillation at matched inference budgets, including improvements in one- and two-step image generation. The gains in one-step generation, where no intermediate sampling times can be adjusted, highlight the benefits of time reparameterization during distillation.

Zero-shot classifiers are useful for routing user requests to specialized LLM tasks, but scoring every request against a large candidate set is expensive: a zero-shot NLI classifier must evaluate one premise-hypothesis pair per label, so cost scales linearly with taxonomy size. We study a student-guided teacher distillation pipeline for a fixed taxonomy of 60 LLM task categories: a compact ModernBERT classifier predicts the full category distribution in one forward pass and retrieves a small top-k candidate set, and a larger DeBERTa-v3 zero-shot NLI classifier reranks only those candidates rather than all 60 labels; the resulting teacher labels iteratively improve the student, which produces sharper candidates for the next round. Unlike generic embedding retrieval or clustering-derived shortlists used in extreme multi-label classification, our candidate generator is trained end-to-end on the target taxonomy and is the same model serving production traffic, distinguishing it from LLM-routing work that routes between candidate models, and from concurrent System-1 encoder-classifier proposals (e.g. TypeSafe AI's Jev and the open-source Laya project) whose training methodology is undocumented or RL-based. Our best student checkpoint reaches 77.5% teacher agreement on a 200-example evaluation set, and preliminary coverage measurements show Coverage@16 of 91-100%, suggesting top-k sets retain most of the teacher's decision-relevant information. We further show truncated top-k teacher scores should not be treated as full 60-class soft targets for KL distillation: zeroing untruncated classes destroys the dark knowledge soft-label distillation depends on, introducing systematic bias rather than a harmless sparse approximation. A complete evaluation, including coverage at multiple k on a held-out set, an embedding-retrieval baseline, and a larger human-reviewed test set, remains in progress.

On-policy distillation (OPD) trains a student on the responses it generates. Existing LLM multi-teacher OPD transfers what specialists predict through their output distributions. We introduce Latent-MOPD, to our knowledge the first representation-level multi-teacher OPD method for LLMs. It integrates existing specialists through both their predictions and the hidden states used to compute them, without additional teacher training. To coordinate representation supervision from multiple specialists, we select late-layer targets according to the teacher-student relationship, bridge unequal hidden widths with a shared projection, and group updates by domain. Each teacher's supervision gradually shifts from hidden states to token predictions, with both channels using the same routed specialist. In our main same-family setting, Latent-MOPD outperforms the token-only, representation-only and uniform-averaging baselines on all nine benchmarks across math, code and logic. With the same parameter count as each teacher, the student also surpasses the per-benchmark best teacher on a majority of these benchmarks. With larger, separately developed cross-family teachers, Latent-MOPD outperforms both single-channel baselines on all benchmarks. A same-family all-layer representation-only control remains stable with domain-pure updates but collapses when teacher domains are interleaved within an update. Our results show that a single student can integrate capabilities from several specialists through both their output distributions and internal representations.

Distribution Matching Distillation (DMD) trains a few-step student from the difference between separately estimated target and student scores, so it must keep an auxiliary diffusion model fitted to the student's evolving distribution at extra memory and computation cost. We introduce DMAD, Distribution Matching as Adversarial Distillation, which recasts distribution matching as classification and learns the required log-density ratios directly. Two discriminator heads on a shared backbone distinguish real data and teacher samples from the student's, and linear losses on their logits train the student without auxiliary score fitting. We prove that at the discriminator optimum these losses recover the distribution-matching gradient underlying DMD, through the classical identity linking discriminator logits to log-density ratios. We further introduce gap-based reweighting, which adapts teacher supervision across noise levels from the real-data head's empirical logit gap between real and teacher samples. DMAD reaches a Fréchet Inception Distance (FID) of 1.04 with one-step generation on ImageNet-64x64, 14.47 with four-step SDXL on COCO-10K, and a VBench total score of 85.15 with four-step Wan2.1-T2V-14B, the best values among the compared few-step methods and the multi-step teachers. On MiniMax-H3-33B, our four-step student achieves overall human preference rates of 79.1% over DMD2 and 84.6% over rCM for joint audio-video generation, excluding ties. Our code, models and demos are available at https://yzmblog.github.io/projects/DMAD.

Training a single LLM agent jointly across diverse interactive environments has attracted increasing attention as a route to generalist agents. Existing curriculum and data-selection strategies often allocate training at the environment level or prioritize local reward-based signals, without explicitly considering relationships between current rollouts across environments for prompt-group selection. Meanwhile, as environments are learned at different rates, all-failure and all-success rollout groups can coexist within a batch, leaving those data without group-relative reward signals. Both challenges highlight limitations of relying solely on scalar rewards in multi-environment RL: they provide limited information about cross-environment relationships and no within-group reward contrast when rewards are identical. This motivates richer textual feedback, such as rubrics describing rollout behaviours, to guide learning. Beyond rubrics' usage as reward, we repurpose rubrics to guide both online data selection and policy supervision. An LLM judge tags each rollout using a predefined rubric vocabulary shared across environments. The resulting profiles guide the selection of data that aligns with the overall behavioural composition of the mixed-environment batch while limiting overlap with already-selected data. Available positive rubrics (describing desired behaviours) provide privileged context for an on-policy self-distillation teacher, supplying additional token-level supervision, while negative rubrics (describing undesired behaviours) guide subsequent rollout generation away from recurring failure modes. Together, these components form RISED. Across model backbones, RISED achieves the highest mean pass rate across environments and ranks first or second in every individual environment. Rubric-based analysis of RISED can further characterize the behavioural changes accompanying these gains.

On-policy self-distillation has recently emerged as an effective approach for improving language-model reasoning by supervising students with a frozen or EMA version of themselves that receives privileged information. Its application to multimodal large language models (MLLMs), however, remains largely unexplored. Recent approaches use privileged visual information, such as image crops corresponding to a question, to improve fine-grained perception, but their gains are confined to tasks that benefit from such visual zooming and require either human-annotated grounding data or external teacher models. We introduce a different form of on-policy self-distillation for MLLMs that provides the teacher with textual, spatially grounded guidance identifying the visual elements relevant to a query. We use procedurally generated scenes with automatically available object identities and spatial coordinates, enabling scalable and annotation-free post-training. The teacher uses this spatial guidance to locate and integrate evidence from multiple relevant image regions, while the student learns to reproduce the resulting behavior from the image and question alone. Our approach consistently improves performance on counting, document and chart understanding benchmarks across multiple models. Importantly, although post-training uses only synthetic scenes, the resulting improvements transfer to real-world perception benchmarks, yielding a 3.23-point gain in average performance across CVBench, V*, ZoomBench, BLINK, HR-Bench, and MME-RealWorld. These results show that spatially grounded privileged information can induce broader perceptual capabilities through on-policy self-distillation, enabling substantial synthetic-to-real transfer beyond the task and data distribution used for post-training. Project page: https://github.com/sirkosophia/Where-OPD

Multi-teacher on-policy distillation (MOPD) aims to combine the strengths of RL-trained teachers in a single student, but how teacher signals affect parameter changes remains underexplored. We study Qwen3-1.7B with four domain teachers trained with RL from the same initialization as the student, comparing gradients, optimizer updates, and task learning curves, with additional SmolLM3-3B diagnostics. We find that several factors influence teacher signals. First, loss averaging implicitly weights responses: token averaging favors longer responses, and equalizing domain contributions retains this weighting within domains. Second, Adam's first moment reduces differences in parameter updates: the cosine similarity is 0.83 between teachers and 0.96 between averaging rules, despite differences in raw gradients. Third, BF16 rounding hides small changes: about 97% of FP32 master weights differ from initialization, but only 7--11% of BF16 weights do. Finally, the top-64 intersection KL gradient closely matches Qwen's full-vocabulary gradient, but the effect on task performance depends on averaging: mathematics accuracy is 2.6 points higher than with sampled-token policy-gradient (PG) under response averaging and 2.1 points lower under global token averaging.

Knowledge distillation aims to transfer the factual knowledge of large language models to smaller models for efficient deployment. Yet a teacher may recall a relation in one direction while failing to generate the answer in the reverse direction. Distillation from its generated answers can therefore propagate this directional limitation to the student. The same teacher can nevertheless recognize such an answer by scoring the relation in the direction it knows. We introduce directional label distillation, in which frozen teachers score candidate answers in that known direction and the best-scoring candidate becomes the student's training target. On facts about parents and their children, known-direction scoring yields more accurate labels than scoring the requested direction, even after tuned corrections for name priors. With prior-corrected scores, the better direction depends on the facts rather than the template, and reverses on mined facts whose notable entity is the parent rather than the child. With the evaluated children's forward facts withheld, students trained on known-direction labels improve open-ended accuracy on their trained queries by 13 to 15 points over students trained on prior-corrected reverse labels. After generated answers are matched to a fixed name list by lexical similarity, students reproduce nearly all selected labels. Their accuracy largely follows label quality. The label advantage holds on unscreened queries and when candidates are retrieved without inserting correct answers. Our findings show that directional verification mitigates the transfer of errors from teacher-generated answers to students by providing more accurate training targets. Code is available at https://github.com/js-lee-AI/directional-verification.

Large audio-language models (LALMs) are sensitive to input perturbations, such as noise, waveform corruption, and adversarial injections. We propose AnchorPrompt, an efficient adaptation method that keeps the model frozen and learns a single block of prompt vectors inserted at the decoder input, between the audio and question embeddings. We train these vectors through self-distillation over diverse audio and text perturbations. To improve answer consistency and mitigate hallucination, we use the model's prediction on the clean recording as the target for answerable inputs, and assign a refusal target when the audio lacks sufficient evidence to answer. Furthermore, AnchorPrompt is perturbation-agnostic at inference, requiring no prior detection of perturbations and enabling zero-shot transfer to unseen distortions. We evaluate three LALMs across three benchmarks and show that AnchorPrompt improves answer consistency in most tested conditions. Clean accuracy improves in six of nine model-benchmark pairs, with minimal impact on the remainder of 1.2% at most. Crucially, AnchorPrompt reduces hallucinations under severe audio corruption while keeping false refusals on clean audio rare. Finally, these consistency gains transfer to unseen perturbations, such as choice permutations and reverberation.

Coding agents solve repository-level tasks through sequences of actions, where a single erroneous action can misdirect subsequent decisions and increase recovery costs. Existing approaches use execution feedback for recovery or specialized checks to block errors, but deciding before execution whether intervention will benefit eventual task completion remains challenging. To address this challenge, we propose HiSentinel, a hindsight-distillation framework that trains lightweight 0.6B and 1.7B sentinels to select pre-execution interventions aimed at improving task completion rather than correcting every imperfect action. A privileged teacher uses recorded execution outcomes as evidence for intervention judgments, which are distilled into a causal student that receives only the pre-action context and proposed action. Beyond identifying whether and when to intervene, the sentinel must also provide actionable feedback that helps the coding agent recover or obtain necessary human input. To support these capabilities, we introduce SWE-Intervene, an action-level dataset constructed from software-engineering trajectories that annotates whether an action should be allowed, autonomously redirected, or paused for human assistance, together with corresponding intervention feedback. Across SWE-bench Verified Mini and Ask or Assume, HiSentinel consistently improves task completion across Sentinel scales and coding-agent families, with gains of up to 14% and 10%, respectively, while maintaining competitive token consumption. These results demonstrate that lightweight pre-execution intervention can effectively prevent error propagation and improve the reliability of autonomous coding agents.

Long-video question answering is limited by the high cost of visual tokens and by the fixed context width of current VLMs. A long-video question may require broad temporal coverage, but the answer is often supported by only a compact set of moments. To locate these moments efficiently, we propose token-budgeted Video Evidence Indexing (VEI): given a dense low-resolution Video Preview, the model constructs a compact high-resolution Evidence Set for final reasoning. We treat VEI as a policy that must jointly solve evidence localization, which finds question-relevant moments, and budget planning, which decides where to spend the limited high-resolution frame budget. We implement this idea with an inference pipeline: the Video Preview provides cheap global coverage, Video Evidence Indexing constructs the Evidence Set, and Answer Generation combines both inputs for final VQA. To address missing frame-level supervision, we adopt privileged self-distillation, where an answer-aware teacher guides the normal test-time policy on student-generated indexing traces. We explore previews at 1, 6, 12, and 24 visual tokens per frame, training a single policy that supports all four resolutions. Experiments show that Video Evidence Indexing improves accuracy under limited visual budgets, and self-distillation further improves both QA accuracy and temporal evidence localization.

Agentic reinforcement learning (RL) trains a large language model (LLM) to act over long, multi-step interactions. However, a single localized error can cause task failure, while trajectory-level rewards provide limited guidance for assigning credit to individual decisions. To address this limitation, we introduce Segment-level Hindsight Advantage Reweighting for Policy Optimization (SHARPO), a credit-assignment mechanism that refines Group Relative Policy Optimization (GRPO) at the level of environment-facing segments. Inspired by the existing on-policy self-distillation (OPSD) method, SHARPO computes teacher-student log-probability gaps within each segment and uses the resulting signal to compute a bounded multiplier on the GRPO advantage. This multiplier is shared by all tokens within the segment, allowing credit to vary across different segments. With Qwen2.5-7B-Instruct, SHARPO outperforms existing baselines on the ALFWorld and WebShop benchmarks, including GRPO, SDAR, RLSD, and StepOPSD.

We present Gumbel Straight Flow (GSF), a continuous flow map language model that leverages the noise-data coupling of a pretrained autoregressive language (AR) model. We theoretically demonstrate that the coupling between Gumbel noise and one-hot token sequences induced by an autoregressive model yields non-intersecting linear paths connecting the noise to the sequence representations. To further enhance high-quality few-step path sampling, we use a flow map semigroup objective where the tangent (velocity) condition is guided directly by the AR teacher. Across various benchmarks, including pretraining and downstream tasks, GSF can outperform current few-step language generation baselines.

On-policy distillation (OPD) is a promising approach for training language agents, providing dense teacher supervision on student-generated trajectories. However, in multi-turn interaction, an incorrect action changes the states the student encounters later, so errors compound across turns. In preliminary experiments across three Qwen3 models (8B to 235B), we find that more than half of the failed rollouts contain a pivotal mistake, an action that moves the agent farther from completing the task, and this mistake typically occurs early. These pivotal mistakes often remain recoverable: guiding the model for only a few turns after the pivotal turn can restore task success. We therefore propose PivotOPD, an on-policy distillation framework that jointly trains the student to prevent pivotal mistakes and to recover from the states they create. At each pivotal mistake, a teacher model provides a gold action and then names a recovery action at each of the next few turns. Preventive distillation uses the gold action with reverse KL to steer the student away from the pivotal mistake, while recovery distillation uses the recovery actions with forward KL to transfer recovery behaviors that the student rarely samples. Against 13 baselines on ALFWorld, WebShop, and Search-based QA, PivotOPD achieves the strongest average performance for both Qwen3-1.7B and Qwen3-8B students, improving over the strongest baseline on ALFWorld by +5.5% with the 1.7B student. The gains also transfer to another model family on the software engineering domain, where PivotOPD raises the resolve rate of a Nemotron-3.5 student on SWE-Bench Verified by +3.2%. Project page: https://research.nvidia.com/labs/lpr/pivotopd/

Length scaling during reinforcement-learning (RL) post-training is often viewed as a sign of improved reasoning ability, especially on difficult problems, but may also make responses to already-solved problems unnecessarily verbose. We quantify this side effect as the length-scaling tax (LST): excess response length on already-solved queries without a commensurate accuracy gain. To mitigate LST, we propose Length Self-Distillation (LSD), which routes solved prompts to on-policy distillation and retains the original RL objective for unsolved prompts. LSD uses an exponential moving average of the online policy as its teacher, requiring no external model. We find that LSD achieves comparable or better performance than RL across multiple variants, while substantially curbing response-length growth on easy queries. LSD reduces LST from 19.0% to -3.7% on single-turn reasoning and from 31.4% to 13.7% on multi-turn agentic tasks, demonstrating that LSD effectively preserves concise response patterns on easy queries while supporting efficient exploration on difficult queries during RL post-training.

Reinforcement learning (RL) has substantially improved the reasoning ability of multimodal language models through verifiable rewards and increasingly fine-grainedvisual or temporal credit assignment. In video reasoning, however, current RL methods typically train with a fixed sparse frame budget: increasing the number of frames makes autoregressive rollouts expensive, while too few frames may miss temporally localized events and fine-grained visual details. We present Frame Differential On-Policy Self-Distillation (FD-OPSD), which transfers the useful evidence of dense frame observations to a sparse frame policy during RL training. FD-OPSD compares the policy's token level preferences for the same sampled response under sparse and dense views, and distills the resulting frame differential signal without an external teacher or dense autoregressive rollout. The method preserves sparse-frame rollouts and leaves inference unchanged. Across Qwen2.5-VL-7B and Qwen3-VL-4B on six video reasoning benchmarks, FD-OPSD yields higher overall average performance than the strongest corresponding GRPO, T-GRPO, or Video-KTR baselines across the 16, 32, and 64 frame evaluation settings. These results show that dense visual evidence can be transferred selectively during training through token level self-distillation while retaining sparse frame rollouts and unchanged inference.

Reinforcement learning with a verifiable reward (RLVR) offers a scalable approach to training language-model agents, yet sparse outcome rewards can leave early training with little signal for policy improvement. We identify an On-Policy Acceleration Phenomenon: in our main comparisons, RLVR initialized with on-policy distillation reaches high performance earlier in training and achieves both higher average performance during subsequent RLVR and higher final performance than the alternative baselines. Motivated by this observation, we study On-Policy Warmup (OPW), a teacher-guided stage in which the student trains with teacher supervision on its own interaction trajectories before transitioning to RLVR. Unlike imitation on fixed teacher-generated trajectories, OPW targets states induced by the student's own decisions, including imperfect actions and recovery situations. We provide a theoretical explanation by connecting on-policy reverse-KL distillation to trajectory-level distribution matching. Under a competent teacher and sufficiently small population distillation loss, this connection yields a lower bound on initial verifier success and a corresponding bound on reward-discovery complexity. For group-relative RLVR, we further characterize when increased success probability produces more reward-informative groups. Together, our findings support on-policy distillation as an effective warmup for agentic RLVR and identify initial reward discovery as a mechanism that can contribute to the observed acceleration.