Distillation 进展

Speech Language Models (SLMs) inherit strong instruction-following capabilities from pretrained language models, yet ASR specialization can substantially degrade them. To address this ASR--QA trade-off, we propose Task-Specific On-Policy Distillation (TS-OPD), which leverages models before and after ASR specialization as complementary QA and ASR teachers. The student generates separate task-conditioned trajectories for ASR and QA, each supervised only by its corresponding teacher, thereby reducing direct competition between the two supervision signals. Experiments on basic ASR, contextual ASR, and QA demonstrate that TS-OPD improves recognition while preserving QA capability. Moreover, TS-OPD remains robust across different balancing coefficients and continues to benefit from increased distillation data.

Query understanding (QU) plays a critical role in production search systems, translating raw user queries into search execution plans that drive downstream retrieval and ranking. While large language models (LLMs) have enabled QU to be framed as a structured multi-task generation problem (e.g., intent classification, query expansion), optimizing such models to produce search-engine-coupled outputs remains challenging: static, label-based supervision fails to capture how each component actually interacts with the underlying search pipeline to affect downstream performance. We present a search-aware reinforcement learning (RL) framework for QU based on a distill-then-RL paradigm. Teacher-student supervised fine-tuning (SFT) first yields a well-formed, schema-compliant policy initialization. The RL stage then optimizes each QU component with rewards derived from live interaction with the search engine, tailored to that component's operational role, rather than a single reward tied to the final search outcome. Experiments on Roblox search show that this component-specific optimization improves both per-component utility and downstream search quality, raising NDCG@20 by 8.9 points over the SFT policy and by 3.5 points over training with a single end-to-end reward.

Direct On-Policy Distillation (Direct-OPD) transfers reinforcement-learning-induced policy improvements from a small model to a larger student by using the token-level log-ratio between post-RL and pre-RL checkpoints as dense supervision on the student's own rollouts. This transfer rewards the policy shift at every state, yet the log-ratio measures only relative change: it can stay fixed even as the probability mass that both checkpoints assign to the student's candidate tokens vanishes. Through an exact construction, we show that the Direct-OPD reward and its update can remain unchanged while the Jensen-Shannon divergence (JSD) and both KL directions between the checkpoints vanish with this mass, and we note that a small JSD bounds how much the teacher's behavior changed. Motivated by this analysis, we propose Selective Supervision for Direct-OPD (S$^2$D-OPD), which ranks student-sampled states by their teacher-reference JSD and masks Direct-OPD supervision at low-divergence states, retaining only the top 10% of states per response. Across two teacher pairs and four student models ranging from 1.7B to 8B parameters, S$^2$D-OPD improves held-out accuracy over dense Direct-OPD on AIME and HMMT benchmarks in seven of eight settings and matches it in the eighth, without extra forward passes.

We find that language models can transfer capabilities through task-unrelated text. Post-training typically improves language models using task-specific data. Prior work on subliminal learning shows that information about these updates can pass through unrelated generations, but has largely focused on traits or preferences using extensive teacher outputs. We introduce Active Taskless Distillation (ATD), which achieves capability transfer using only a single word from the teacher per prompt. ATD probes the behavioral shadow of post-training by selecting prompts where the teacher and student's shared public ancestor is nearly indifferent between two ordinary words. A student initialized from this ancestor learns solely from the resulting prompt-word pairs, without target-task examples, teacher logits, or teacher parameters. In the primary coding experiment with Qwen2.5-1.5B, 5,664nses yield a 5.34 pp gain on HumanEval+ over an exact nuisance-matched control thadisrupts prompt-resperiments showtransfer in scientific knowledge, commonsense reasoning, and reading comprehensins across additional model generations, sizes, and families. Functional analyses show that the learned sid composable, andthat its strength tracks the teacher's update strength.

Video diffusion is computationally expensive, as it requires executing a large model across many denoising steps. Even with step-distillation, inference remains expensive because every distilled step still requires a costly model evaluation. We present TRACK: TRajectory-Aware Capacity routing via top-K selection, a heterogeneous denoising strategy that switches between compatible large and small models at selected steps, reducing the average cost per denoising evaluation. The switching steps are determined using a calibration process. TRACK first rolls out a reference trajectory with the large model. Then at each step, the small model's prediction is also collected and compared against the large model's prediction to obtain a relative disagreement score. Both models receive the same latent, timestep, conditioning, and guidance inputs. Aggregating this signal over a calibration set produces a disagreement score map across diffusion steps, which determines a switching policy for an efficient inference process: quality-sensitive steps keep using the large model, while steps with low disagreement scores are routed to the small model. Inference executes only the selected model at each step, requiring no retraining, architecture or scheduler changes, or online dual-model evaluation. Across Wan 2.1, Cosmos 3, TurboDiffusion, and FastVideo, TRACK yields $1.95\times$, $2.04\times$-$2.73\times$, $2.69\times$, and $2.17\times$ speedups, respectively, with comparable aggregate quality and high diversity retention. TRACK thereby establishes automated, training-free model switching as a practical acceleration paradigm for video diffusion.

On-policy self-distillation (OPSD) has become a popular recipe for post-training LLM agents. It supervises the agent model at the token level with a stronger teacher view of the same model, obtained by conditioning on privileged information (PI). In this work, we show that in multi-turn agents, this paradigm teaches the student to act with confidence but without the information behind it. The trained agent behaves as if it had privileged information it never observed, and its performance falls well short of plain RL, in the worst case below the untrained base model. Therefore, we propose Privileged Self-Practice (PSP), which keeps the PI and moves it from the loss to the sampler. When the student's rollouts on a task mostly fail, we inject a short per-task instruction written by an analyzer model, sample the task again with the instruction in context, and train on the result with an unchanged GRPO objective. The privileged information stays in the prompt and never enters the loss. Across AppWorld and SWE-bench Verified, with three different student models, PSP obtains the best average score in every setting and is the only method that consistently outperforms plain GRPO, improving task-goal completion by up to 65% on AppWorld and the resolved rate by up to 61% on SWE-bench Verified.

Few-step autoregressive (AR) video diffusion enables low-latency streaming generation, but existing post-training methods predominantly rely on Distribution Matching Distillation (DMD), requiring both a large pretrained teacher and an online critic to estimate distributional discrepancies through diffusion scores. In this work, we ask whether this resource-intensive teacher--critic stack can be eliminated by post-training only the generator against a precomputed target distribution. Drawing inspiration from representation distribution matching (RDM) for one-step image generation, we systematically study its transfer to few-step causal video generation and identify three key barriers: a memory-intractable gradient path, a distinct video optimization regime, and representation distributions that underconstrain temporal dynamics. We introduce ViRDM, a teacher- and critic-free video post-training recipe that addresses these barriers sequentially. By coupling RDM with stochastically truncated clean-exit supervision, a lightweight VAE decoder, and staged vector--Jacobian products, ViRDM makes representation distribution matching memory-feasible for multi-step causal video rollouts. We further establish effective generated-population and initialization regimes for video RDM, and introduce lightweight dynamics regularization to compensate for the underconstrained temporal dynamics. ViRDM turns three-network distillation into generator-only post-training, reducing GPU memory use and training time while improving video quality. With only 20 generator updates, the recipe reaches 84.87 on the official VBench evaluation, outperforming the previous best few-step causal baseline by 0.36, while requiring 16 A100 GPU-hours. We additionally report exploratory results demonstrating the potential of the same recipe for lower causal sampling budget and for one-, two-, and four-step bidirectional generation.

Distribution Matching Distillation (DMD) enables high-quality diffusion sampling in only a few steps, but its optimization dynamics remain dominated by coarse, low-frequency signals, delaying the recovery of fine-grained details. We identify a pronounced concentration of spectral amplitudes at low frequencies in the DMD directional error, where dominant low-frequency components overwhelm weaker mid- and high-frequency signals. To address this issue, we propose Spectral Amplitude Purification for Distribution Matching Distillation (SAP-DMD), a plug-and-play approach that adaptively modulates the amplitude spectrum of the DMD directional field. By suppressing the dominant tail of the amplitude spectrum, SAP-DMD reduces low-frequency dominance and promotes more effective recovery of fine structures and textures. Experiments on PixArt-$α$, SD3, and SD3.5 demonstrate that SAP-DMD accelerates training convergence and improves generation quality under both 2-step and 4-step sampling.

Reinforcement learning (RL) improves reasoning, but its performance depends on the policy from which training begins. We study on-policy distillation (OPD) as a preparation stage for RL and ask whether its benefits extend beyond improvements in the distilled model's initial accuracy. Under shared RL settings, students initialized with OPD reach higher final performance than those trained with direct RL or supervised fine-tuning followed by RL. This advantage can emerge even when OPD produces little immediate improvement in accuracy. Pre-RL Pass@k does not fully explain the benefit: similar or even higher values do not necessarily lead to better performance after RL. Behavioral analyses point to alignment with the teacher's distribution beyond top-1 agreement as a possible explanation. Such alignment may favor higher-quality reasoning paths while retaining alternatives that RL can further refine using outcome feedback. We further examine how trajectory sources and divergence objectives affect the value of distillation for subsequent RL. Standard reverse-KL OPD performs better before RL, but forward-KL OPD overtakes it afterward. Student rollouts outperform teacher rollouts under both objectives before and after RL. These findings highlight the importance of both objective choice and the states receiving supervision for subsequent RL. Our results support OPD as preparation for RL and favor forward KL when OPD is followed by RL in our comparison.

Audio-language models (ALMs) can exploit textual shortcuts to answer questions while overlooking acoustic evidence, weakening audio understanding. On-policy distillation (OPD) trains compact ALMs by supervising student-generated responses with teacher predictions, but does not explicitly distinguish acoustic support from linguistic predictability. We propose Reward-Tilted On-Policy Distillation (RT-OPD) to strengthen acoustic grounding. Given the same question and student-generated text, a frozen teacher predicts the next token with and without audio inputs. Their log-probability contrast defines a reward that reshapes the teacher distribution for reverse-KL distillation, emphasizing the additional evidence provided by audio. Across two compact students and three benchmarks, RT-OPD consistently outperforms Vanilla OPD. Experiments with silenced and replacement audio further suggest that RT-OPD strengthens the student's reliance on acoustic evidence. Our 3B model achieves 72.72% accuracy on MMAU, the highest among the compared 3B models and competitive with several 7B and 8B models. Code and model checkpoints are available at https://github.com/KaiyangLi1992/RT-OPD.

Multi-constraint instruction following requires a model to respond to a query under many simultaneously active constraints. Even strong instruction-tuned models still routinely violate some of them. Existing approaches either augment supervision with sequence- or token-level RL rewards from external verifiers or learned graders, or use on-policy distillation (OPD) against a single full-context teacher whose probability mass becomes diluted as more constraints become simultaneously active. We propose CC-OPD (Counterfactual Constraint-Conditioned On-Policy Distillation), which inverts the standard supervision-generation direction in distillation. Rather than enriching the teacher with information beyond what the student sees, CC-OPD ablates each constraint from the teacher's conditioning in turn, and constructs the per-constraint signal from the resulting per-token probability differentials. The resulting per-token leave-one-out log-likelihood shifts are summed, clipped, and added to the vanilla OPD reward as a token-level shaping term. All shaping terms are obtained from the frozen teacher, without an external verifier during distillation, and the reward equals vanilla OPD wherever the aggregate shift is zero. Across two Qwen model pairs and seven benchmarks, CC-OPD achieves the highest average among all evaluated student-training methods. A 1.5B student trained with CC-OPD surpasses its own 7B RL-trained teacher on the MulDimIF benchmark.

Reinforcement learning with verifiable rewards (RLVR) supervises mathematical reasoning through final-answer correctness, but provides little guidance on individual tokens. On-policy distillation (OPD) supplies dense feedback on student-generated responses, yet teacher preference need not reflect correctness. Recent hybrids combine OPD and verifier-derived advantages or reweight task credit using teacher ratios. However, teacher guidance enters after verifier-based group normalization, and token reweighting need not preserve the total task credit assigned to each response. We introduce Unified Entropy-Calibrated Credit Redistribution for GRPO (UECR-GRPO), which integrates verifier and teacher signals within a single GRPO-style update at both the response and token levels. Path-Utility Unification (PUU) combines verifier reward and a teacher-to-anchor path log-ratio in a single KL-regularized objective. Its on-policy implementation uses a length-normalized teacher score and combines both rewards before group normalization and PPO clipping, allowing teacher evidence to influence the response ranking. Entropy-Calibrated Redistribution (ECR) then uses the signed teacher--old-policy token gap to redistribute the verifier-derived component. Full-vocabulary teacher entropy attenuates uncertain guidance, while a response-wise zero-sum projection preserves the total task credit and its token-wise sign before clipping. Across five mathematical reasoning benchmarks, UECR-GRPO achieves average \(\mathrm{Avg@12}\) accuracies of 17.21% and 65.09% with Qwen3-1.7B and Qwen3-4B students, respectively, exceeding the strongest baseline at each scale by 0.89 and 0.56 percentage points.

On-policy distillation (OPD) corrects a student on the responses it writes, but its signal is the teacher's next-token distribution: it tells the student what the teacher says but misses how it thinks. Latent supervision promises the missing part by aligning the student's latent states to the teacher's. Recent methods such as OPRD bring this signal into on-policy distillation. However, we observe two failures of this recipe when distilling Qwen3-4B and Qwen3-8B into Qwen3-1.7B-Base. Early gain, late collapse: latent supervision alone lifts MATH-500 accuracy from 25 to 46 in 10 steps, but subsequent training degrades performance down to 11 with no recovery. Better alignment, worse behavior: although the alignment metric steadily improves throughout this collapse, the most aligned model turns out to be the worst performing. Further analysis suggests a mismatch in how the latent signal is applied: layers paired by depth play different roles in the two models, so continued alignment may pull the student toward teacher states it cannot understand. To address this, we propose LastOPD, which applies the latent signal only at the last-layer state, the common interface both LM heads read, and only during a 10-step crossfade into token-level OPD. This keeps the useful part of the latent signal and hands the student to token-level supervision before the collapse sets in. Extensive experiments show that LastOPD improves MATH-500 over token-only OPD by 5.55 and 4.02 points with the 4B and 8B teachers, leads on most held-out datasets, and reaches the final score of token-only OPD in about half the steps. Code is available at https://github.com/Muyiiiii/LastOPD.

An estimated 1 billion people worldwide live with vision impairment, yet current vision-language models (VLMs) produce descriptions too vague for safe navigation by blind and low-vision (BLV) users. Large VLMs can generate high-quality audio-description-compliant narrations but cannot run on mobile devices; small VLMs offer competitive latency but lack spatial detail, directional cues, and hazard awareness for navigational assistance. We present Smol-VL-BLV, a compact VLM for blind and low-vision users that closes this gap using a 500M decoder transformer model and two post-training mechanisms: (1) teacher-student distillation and (2) Group Relative Policy Optimization (GRPO) with a composite BLV reward targeting directional language, metric distances, and hazard detection. Because multi-stage post-training can induce catastrophic forgetting, we add a lightweight finetuning stage after the last stage GRPO finetuning to recover general descriptive quality while preserving BLV-specific spatial grounding. Our best model substantially outperforms the baseline across various benchmarks, including tasks: VQA, BLV captioning, OCR, and latency. Compared with the baseline for relative improvement, it improves the Spatial score gain of 19.3%, and the Social score gain of 14.8%. It also increases OCR-Bench by 101.5%, and raises TextVQA accuracy by 44.2%. These results show that BLV-focused post-training improves both accessibility-specific spatial grounding and general visual-text reasoning. Deployed on a mid-range Android smartphone via Mixed-Precision Quantization, the model remains approx. 450 MB and runs entirely on-device, offline and without network dependency, generating descriptions with latency dependent on host hardware capabilities. Our model, dataset, and code is publicly released at https://smol-vl-blv.github.io/Smol-VL-BLV-website/

Transformer language models process sequences token by token in an autoregressive manner, making growing contexts increasingly expensive. Yet many adjacent token spans are highly predictable or frequently occur as stable units, suggesting that their representations may be compressible. We introduce a method for distilling sequential computation by replacing spans of input tokens with collapsed representations, computed on the fly by a lightweight merge module. This module generates a single surrogate embedding from a sequence of static token embeddings that captures the functional role of the multiple tokens, allowing pretrained models to operate on compressed inputs without architectural changes or re-training. We apply this approach during inference to compress both prompts and intermediate decoding steps, using a rollback mechanism to substitute stored multi-token KV cache entries with their single-step surrogates. Experiments across diverse models show that the merge module can be used to reduce effective sequence length by up to 40% with minimal accuracy degradation across language modeling evaluations and downstream tasks, including question answering, summarization, commonsense reasoning, and long-form mathematical reasoning. Additional lightweight adaptation of the merge module further improves the accuracy-compression trade-off in selected settings. These results demonstrate that sequential token computation in Transformers can be effectively approximated through condensed surrogate representations that approximate the original behavior without model updating.

Hi-OPD addresses a failure mode left uncontrolled by flat open-prompt training: descendant retrieval need not persist under ancestor queries when multi-source remote sensing annotations exhibit inconsistent granularity and missing labels. A detector may localize car and van under atomic prompts yet miss the same instances under vehicle; flat AP does not expose this cross-level inconsistency. We propose Hi-OPD, a hierarchy-aware open-prompt detector, and construct RS153-HierOPD from 175,644 retained training image/tile records and 3.48M boxes mapped to 153 atomic categories with sparse hierarchy and alias relations. Hi-OPD learns ancestor retrieval through hierarchy-safe negative sampling, path multi-positive supervision, and one-way upward consistency, while per-source risk exclusion handles potentially missing labels. ConvVPE converts K-shot support boxes into text-compatible embeddings using detector-native features and the shared contrastive head. On Track A, Hi-OPD obtains 79.7/72.3 AP50 on DIOR/DOTA-v2.0, above the literature-reported OpenRSD results of 76.7/71.8. Under controlled training on the original converted annotations, the full hierarchy recipe raises DOTA-v2.0 parent AP50 from 7.2 to 71.5 and FAIR1M grandparent AP50 from 31.6 to 71.4, while DOTA-v2.0 atomic AP50 changes from 71.4 to 72.3. The text path reaches 99.7% CAR50 (0.3% violation) across the three common sources and 99.9%/0.1% on FAIR1M grandparent relations. On held-out VEDAI, text AP50 is 75.9, 6.2 points above OpenRSD. Joint AP and CAR show that explicit hierarchy training repairs this failure mode while retaining atomic detection and prompt transfer.

Reinforcement learning has become a central component of large language model (LLM) post-training, yet token-level credit lacks a generally accepted mathematical definition, leaving its relationship to commonly used training signals unclear. We formulate three regularity conditions, namely Completeness, Prefix Consistency, and Neutrality, and prove that they uniquely determine token-level credit. This characterization provides a unified basis for explaining phenomena across existing algorithms and guides the development of an improved actor-critic training procedure. Through this lens, an ideal teacher in On-Policy Distillation (OPD) acts as an implicit critic, yielding an expected policy gradient proportional to that induced by token-level credit. Response-level REINFORCE Leave-One-Out (RLOO) signals match the expected policy-gradient contribution of token-level credit despite their coarser granularity. We further establish approximate credit sparsity under bounded outcome rewards and show how intermediate critic errors in Generalized Advantage Estimation (GAE) can become comparable to the underlying credit. These motivate Policy Aligned Critic Training (PACT), which adopts an Actor-then-Critic update order to apply importance sampling correction to critic training and better align the critic with the updated policy. In agentic mathematical reasoning, PACT achieves 72.87% average accuracy across four benchmarks, outperforming GRPO and PPO by 8.80 and 13.16 percentage points, respectively. On SWE-bench Verified, PACT achieves a pass rate of 67.4%, outperforming PPO, GRPO, and SAO by 2.4, 2.0, and 3.8 percentage points, respectively.

Multimodal large language models (MLLMs) often struggle with fine-grained visual perception when processing complete images, as critical evidence may only appear in local regions. On-policy self-distillation (OPD) enables transferring privileged visual knowledge from informative views to full-image policies, but querying the teacher for every rollout introduces substantial supervision costs. In this work, we propose BAS-OPD, a budget-aware selective OPD framework that allocates teacher supervision under limited query budgets. Instead of querying all rollouts, BAS-OPD selects informative samples while maintaining full-batch student generation. We explore random, uncertainty-based, and learned utility-based selection strategies, where the learned selector estimates query value from detached rollout statistics and online utility signals derived from student--teacher agreement and teacher confidence without additional student forward passes. BAS-OPD only changes training-time supervision allocation and preserves single-pass full-image inference. Experiments on fine-grained multimodal perception benchmarks demonstrate that BAS-OPD achieves strong performance while substantially reducing teacher supervision costs, highlighting the effectiveness of selective OPD under constrained budgets.

More privileged information does not always make a better teacher. We study this tension in on-policy self-distillation (OPSD), where a self-teacher scores the student's own rollouts under privileged context, conventionally a complete reference solution that bundles the final answer with one particular reasoning path. Holding the student view and training fixed within each scale, we compare that default against three abstractions compiled offline, a named strategy, a method-independent framing, and a problem category, and against an answer-only control that keeps the destination but removes the path. In the primary runs on competition mathematics, the best intermediate contexts improve the in-domain peak mean over the full solution by 1.4 points at 4B and 1.6 at 8B, while storing an order of magnitude fewer hint tokens. Comparisons across three seeds also show positive mean gains for the framing and category contexts at both scales. Answer-only conditioning remains competitive in the primary runs, within 0.2 points of the full solution at these scales. The preferred context varies with student scale and task. Initial teacher-student KL does not order downstream performance. What a self-teacher should see is therefore not everything it could, but the level of abstraction its student can still act on.

Commercial image editing requires product identity preservation, accurate text rendering, and user appeal alongside general editing quality. We present KwaiMind, an image editing system combining general capabilities with e-commerce specialization. An agent-based data engine maintains approximately 1.8 million high-quality editing pairs. Built on a multimodal diffusion transformer, KwaiMind undergoes continued pre-training and supervised fine-tuning, followed by preference optimization and online reinforcement learning. A general-purpose vision-language judge and specialized rewards for click-through rate (CTR), text rendering, and product consistency guide specialized policies, which are consolidated through on-policy distillation. We introduce Ecom-Bench, covering 11 commercial editing tasks with task-specific visual evaluation and CTR-based ranking. KwaiMind achieves the strongest overall scores among evaluated open-source editors on ImgEdit, GEdit, both language splits of REDEdit, and Ecom-Bench visual quality, and the highest aggregate CTR ranking score among compared systems. Offline, CTR-guided optimization increases the proportion of generated images whose predicted CTR exceeds that of the original product image from 12.16% to 37.41%. In an online A/B experiment, CTR-based selection of product main images yields an approximately 2.44% relative increase in actual CTR. These results demonstrate the value of domain-specific data and reward-driven alignment for commercial image editing.

Quantization-aware distillation (QAD) restores much of the short-form question-answering performance lost to sub-3-bit quantization, yet leaves mathematical and code reasoning substantially impaired. Long generations often degenerate into repetitive loops, exhausting the decoding budget without completing a solution. We trace this gap to quantization-amplified exposure bias: QAD trains on fixed corpus prefixes, while quantization-induced deviations compound along the model's own autoregressive trajectories. To address this mismatch, we introduce an on-policy distillation (OPD) stage that places teacher supervision where the quantized model actually goes. Starting from a QAD checkpoint, the student generates through the quantized forward path used at deployment and receives feedback from a frozen full-precision teacher on its own prefixes, combining dense token-level guidance with task-verifier rewards. Across four models at 2.79 and 1.88 effective bits, OPD raises average BF16 performance retention from 35% to 70% on MATH-500 and from 66% to 91% on HumanEval while preserving short-form performance, with reasoning gains substantially exceeding those of continued teacher-forced QAD in matched-budget comparisons. By coupling QAD's stable low-bit initialization with OPD's on-policy reasoning recovery, our framework provides a comprehensive sub-3-bit solution that preserves broad capabilities while restoring long-form reasoning.

Failure-based textual knowledge distillation aims to discover gaps in a model's knowledge by examining its task errors. The distilled knowledge can be useful for the reasoning of both this model ("source model") and other models. However, this transfer of knowledge may not be stable. We define a rule atom to be a standalone rule injected into a model's textual input at inference time. A rule atom can encode transferable task knowledge or model-specific reasoning patches that can confuse other models. Also, the injected rule atoms can be misapplied to unrelated cases, causing the model to incorrectly flip its answer based on irrelevant information. Building on a pipeline that distills training examples into task-specific cheat sheets that aid model reasoning, we examine when failure-derived rules can improve these cheat sheets. Our early experiment shows rule distillation from a single model's failures underperforms the baseline cheat sheet on non-source model families. This motivates Robust Failure, Conservative Repair (RFCR), a textual distillation procedure that derives rules from failures shared across models, sharpens their application boundaries using boundary cases, and abstains when no useful rule is found. On a 400-item BIG-Bench Hard task set, RFCR improves the baseline cheat sheets from 68.50% to 71.25% (+2.75 pp; 95% CI [+1.25,+4.50]) without performance degradation on previously correct cases. Ablations and cross-model diagnostics support that accuracy gains come from both new knowledge injection and strict rule-application control.

Real-time voice assistants must reason over evolving requests, execute actions, and follow conversational rules. Qwen-Audio-3.1-Realtime brings these requirements together through Think, Act, and Speak and Coordinate. Think combines Core-Cocktail supervised fine-tuning with Multimodality and Multi-Teacher On-Policy Distillation (M$^{2}$-OPD) to transfer language capabilities and develop native audio skills. Act uses self-evolving executable environments and multi-granularity rollouts for Group Relative Policy Optimization (GRPO), teaching the model to use tools, interpret feedback, and complete tasks. Speak and Coordinate aligns whether, when, and how the assistant speaks or acts. We evaluate audio reasoning, multilingual understanding, tool use, conversational behavior, full-duplex interaction, and safety. Compared with Qwen-Audio-3.0-Realtime, 3.1 raises overall task success from 78.4% to 82.0% on our half-duplex speech-to-text adaptation of $τ$-Voice. On speech-to-speech Full-Duplex-Bench v1.5, the response rate to background speech falls from 73.0% to 13.0%. We also present a separate Voice Harness prototype, using Qwen-Audio-3.0-Realtime as its foreground, that extends spoken interaction to persistent tasks through foreground--background coordination and memory.

Low-light image enhancement (LLIE) is an im- portant component of visual sensing systems operating under degraded illumination, including nighttime surveillance, au- tonomous navigation, remote sensing, and inspection in poorly lit industrial environments. Most LLIE methods rely on output- level reconstruction losses that supervise only the final restored image, leaving the intermediate feature recovery process weakly constrained. This paper proposes MirrorDistill, an illumination- aware latent distillation framework that links the low-light and clean domains through feature mirroring. During training, a shared encoder and an exponential-moving-average teacher decoder process the clean reference image to generate clean- domain latent targets. These targets supervise the low-light student at two levels: raw encoder features and standardized multi-scale decoder projections. The alignment is applied layer by layer, while a proposed illumination-aware weighting scheme gives greater emphasis to underexposed regions. The teacher and reference branches are used only during training, so inference requires only the lightweight student encoder-decoder and in- troduces no teacher-side computational cost. Under evaluation on the standard LOL benchmarks, MirrorDistill outperforms the state-of-the-art methods on the real-captured LOL-v2-Real set, while having the lowest compute complexity (GMACs) and while remaining competitive on the LOL-v1 and LOL-v2-Synthetic datasets. Ablation studies further show the contributions of the encoder mirror, decoder mirror, and illumination-aware weighting. Finally, we release our code as open-source for the benefit of future research.

On-policy self-distillation fine-tuning (SDFT) learns new skills from demonstrations while reducing forgetting, but it always distils toward the full demonstration-conditioned teacher. This fixes teacher influence at the full-teacher endpoint, providing no control over how much demonstration information should be transferred at each prediction state. We introduce Information-Proximal SDFT (iSDFT), which instead treats the teacher as a budgeted source of information. At each token, iSDFT selects the distribution closest to the current student that satisfies a prescribed teacher-information constraint, yielding a closed-form exponential target with a locally determined tilt. To control cumulative drift, we further anchor the student to its frozen base policy. Across four heterogeneous LLM backbones and two specialisation tasks, iSDFT improves vanilla SDFT in 7 of 8 model-task settings and matches it in the remaining one. It also provides tighter retention on the original SDFT benchmark suite, with 73% of evaluations remaining within 0.5 points of the base model versus 52% for the strongest baseline, while achieving the largest mean improvement on all ten additional mathematics, coding, and competition-mathematics benchmarks. These results show that controlling how much and when teacher information is introduced improves specialisation while preserving broader capability.

Video world models enable interactive exploration of dynamic environments, yet struggle to respect prior observations over long horizons and across viewpoints. We present WorldCrafter, a video world model that learns a camera-queryable implicit 3D-aware memory for this purpose. The key insight is to let the requested viewpoint shape how multi-view evidence is compressed into the video generator's limited token budget. Trained jointly with the video generator, a memory encoder and pose-conditioned readout module integrate historical observations into a fixed set of target view-specific tokens before denoising, without explicit depth-based correspondences. By combining this memory with recent temporal context and few-step distillation, WorldCrafter enables streaming scene exploration from a single input image or text prompt. Experiments across static and dynamic scenes show substantial gains in long-horizon consistency and camera-control accuracy while preserving visual quality during minute-scale exploration.

Agent harnesses, the external systems that mediate model-environment interaction, can substantially improve agent performance, but their gains remain tied to the harness at deployment. Because the best harness varies across domains, instances, and models, a general-purpose agent must either settle for a suboptimal shared harness or route among an ever-growing set of specialized ones. We therefore study agent harness distillation: using a domain- or instance-optimized harness as training-time guidance and transferring the behaviors it induces into model weights, so that its gains survive under a single fixed target harness. The challenge is that the two harnesses differ in action space and available information, so guidance from the optimized harness cannot serve directly as supervision for the target one. We introduce Harness-Zero, which enables harness distillation through agent-as-harness. Guided by the optimized harness, a harnessing agent corrects student responses before execution in the target harness's action space, turning harness guidance into training demonstrations. Fine-tuning on the resulting trajectories internalizes harness-induced behavior into the model, so the specialized harness can be removed at deployment. Our experiments spanning knowledge work, tool use, and science domains show that: (1) For frontier LLMs using the same evolved harness, agent-as-harness outperforms code-as-harness. (2) With the specialized harness removed at deployment, Harness-Zero improves the base model's macro-average task success from 23.3% to 44.3%, even exceeding the 41.7% it reaches with that harness still attached. (3) Harness-Zero recovers harness-induced behaviors absent from the base model, with 82.3% average recovery across 28 patterns in the three domains.

Building general-purpose agents for industrial deployment requires integrating multiple capabilities, each typically acquired at a distinct stage of training. Yet there is currently no well-established recipe for Agent Continual Learning (ACL), with little understanding of the trade-offs among existing integration paradigms. To address this gap, we introduce ACLArena, a framework for comprehensively studying, analyzing, and evaluating ACL. We first build a sequential training pipeline and conduct an in-depth analysis that explains the mechanisms of forgetting and generalization from two complementary perspectives, the model level and the token level. Guided by these analyses, we systematically compare multi-teacher on-policy distillation, self-distilled fine-tuning, and model merging to assess their ability to recover previously learned capabilities while preserving newly acquired ones. Through extensive experiments, we develop a detailed understanding of how capabilities transfer across stages. Finally, we propose a new ACL recipe that combines offline replay over high-quality trajectories with a routed network of multiple LoRA experts each specialized via RL, substantially improving the agent's ability to learn across multiple domains. Comprehensive experiments on four reasoning and agentic tasks, evaluated under both in-domain and out-of-domain settings, demonstrate the value of our analysis and the effectiveness of our approach.

On-policy distillation (OPD) pays twice for each fresh batch: the student generates trajectories and a stronger teacher scores them. Existing methods improve which trajectories are scored and how the teacher signal is constructed, but usually consume it with one actor update. We introduce CLOOPD, a closed-loop framework separating teacher-signal acquisition from student-side realization. CLOOPD selects an adaptive $α$ waypoint inside a KL envelope, freezes the scored batch and its advantages, re-forwards the student after each actor pass, measures realization, and allocates actor work under a separate token budget. The framework includes deterministic two- and three-pass policies, token-priced CLOOPD-TPMR, and a budget-matched control. Across six 300-step runs on an 8-H20 node, every CLOOPD policy improves the one-pass TOP-D anchor at comparable teacher-token scale: macro accuracy rises from 15.41 to 17.78 with CLOOPD-Fixed2 and 19.36 with CLOOPD-Fixed3. At step 100, CLOOPD-Fixed3 reaches 15.35, nearly matching TOP-D at step 300 while using 67.2% fewer teacher-scored tokens and 28.0% fewer GPU-hours. Earlier 8-A100 ablations show adaptive $α$ eliminates observed trust-envelope violations; a third pass adds headroom. These results position CLOOPD as a framework for budgeting how fully students learn from teacher-scored tokens.

Sparse on-policy distillation (OPD) allocates teacher supervision to a small subset of tokens in student-generated trajectories. However, useful teacher guidance can yield a noisy update when its gradient is estimated from a sampled next token. We study this estimation problem at a fixed prefix in information geometry and propose an information-efficiency ratio (IER) based on a signal-to-noise decomposition. IER characterizes relative gradient estimation error under an optimal scalar baseline. A candidate-set approximation enables token selection based on IER and its combination with existing usefulness scores, while retaining the sampled reverse-KL training objective. On mathematical and medical reasoning tasks, adding IER improves existing selectors in multiple settings, with sparse configurations matching or exceeding full OPD without token selection at small token budgets of 0.1%--1%. These results support accounting for both usefulness and gradient-estimation reliability when allocating sparse supervision. Our code is available at https://github.com/BruceSheng1202/IER-OPD.