每天知道
你的领域
出了什么

Specialized models encode task-oriented behavior, but transferring that behavior to a general language model usually requires training, distillation, or representation alignment. We study whether such ability can instead be transferred directly at the parameter level. We apply two existing training-free heterogeneous merging methods, previously shown to transfer knowledge between general language models, to specialist-to-general transfer, projecting a specialist donor into the recipient's shape and interpolating backbone parameters without gradient updates or semantic alignment. Intersection-Merge (IM) injects a prefix-aligned donor slice matching the recipient shape, while Activate-Prune-Merge (APM) uses forward-pass activation statistics to select which donor dimensions to retain before injection. Across embedding, reranking, reward modeling, and MoE code-specialist transfer, both methods improve the general recipient, showing that simple heterogeneous merging can move capabilities across diverse specialist roles.

Common policy improvement methods, including TRPO, PPO, and GRPO, estimate policy improvement under the behavioral policy's state-visitation distribution rather than the improved policy's own. The substitution makes the objective estimable from the behavioral policy's rollouts but adds a bias growing with policy divergence, hence the trust region or clip, and hence no reuse of a batch far off-policy. We show that under history-injective dynamics, where each state is reached by exactly one history, the dropped state-visitation ratio equals the product of per-step policy ratios along the sampled prefix, on every trajectory and not only in expectation. The ratio is therefore restored exactly, from log-probabilities PPO already computes. Autoregressive generation and canonical-order constructive optimization are both history-injective. The exact correction pays importance-sampling variance that grows with the horizon, so we generalize it to a one-parameter family with PPO ($α{=}0$) and the full correction ($α{=}1$) as endpoints: a single bias--variance knob. A gradient-level analysis of the unclipped surrogate identifies two channels the correction acts through and three conditions under which it carries signal; an enumerable testbed confirms the conditions' predictions. On hard credit-assignment scheduling tasks, a short corrected warmup with aggressive early sample reuse learns faster than PPO and than the same reuse uncorrected; the marginal gain grows with task difficulty ($+0.02$ to $+0.09$ learning-curve AUC), and the early win over PPO tracks the prefix bias that reuse incurs. A correction held throughout, or applied where clipping already contains the reuse bias, is null to harmful.

Cinematic camera motion is a fundamental storytelling tool, defined not only by where the camera is positioned in the scene, but also by how it moves in terms of direction and speed. Recent work on camera trajectory generation and alignment to text relies on pose-centric representations. While in principle a network could derive direction of movement and speed, we find that in practice this might not happen. In fact, in this paper we discover that decomposing the camera trajectory representation from the traditional per-frame poses to direction and speed has surprising benefits across multiple tasks, including trajectory-to-text alignment as well as text-to-trajectory generation. To accurately evaluate the former, we introduce a simple and reliable protocol that overcomes the limitations of prior evaluation baselines. For the latter, building on this representational insight, we propose a novel generative model for camera trajectories, CineGEN, that achieves superior performance across a variety of metrics. We also propose a novel dataset, CineScript, containing movie clips that are enriched with scene descriptions as well as higher-level metadata. This novel data allows us to test models' ability to capture high-level cinematographic information. We show that, despite its simplicity, representing camera trajectories through direction and speed not only helps numerically to achieve better alignment and generation, but also inherently encodes complex directorial intent.

Inverse Reinforcement Learning (IRL) aims to recover a reward function that explains expert demonstrations. Existing IRL methods typically rely on a bi-level optimization procedure that alternates between reward learning and policy optimization, leading to substantial computational burden and training instability. In this work, we introduce a different route that eliminates policy optimization entirely by leveraging diffusion policies. Our key insight is that a diffusion policy encodes the action-gradient structure of the optimal soft Q function, enabling reward learning to be cast as a sequence of value recovery problems, thereby allowing us to bypass reward-policy loops inherent in prior IRL methods. Specifically, our method proceeds in three stages: (I) recovering the optimal soft Q function via action-gradient matching and estimating the corresponding soft value function (LogSumExp of Q values) in a way inspired by Gumbel regression; (II) calibrating these soft values by inferring a state-dependent offset; (III) extracting the reward by enforcing Bellman consistency. This leads to Loop-Free Inverse Reinforcement Learning (LFIRL), a fully offline algorithm that operates in a simple, loop-free, and sequential manner. LFIRL is simple to implement and significantly improves training efficiency while maintaining strong reward recovery performance. Empirically, across Maze, Franka Kitchen, Adroit Hand Pen, and Push-T benchmarks, LFIRL achieves 2-3x speedup over the fastest baselines, while matching or surpassing state-of-the-art methods in reward recovery quality.

We study preference-based bandits with general reward function classes, where a learner sequentially selects pairs of arms and observes binary preference feedback governed by the Bradley--Terry model. This setting naturally arises in applications such as recommender systems, tournament ranking, and learning from human feedback, where relative preferences are easier to elicit than absolute rewards. The observation model inherits the logistic bandit challenge of handling the problem-dependent constant $κ$, which accounts for the non-linearity of the link function and can grow arbitrarily large. Moreover, prior work has predominantly focused on linear or kernelized reward models, precluding the use of richer function classes. To address these limitations, we consider general reward function classes and introduce the locally sensitive eluder dimension, a novel complexity measure tailored to the logistic structure of preference feedback that yields fine-grained regret guarantees without unfavorable dependence on $κ$. Building on this notion, we propose GINOP (Generic INformative OPtimism), an algorithm that constructs log-loss confidence sets and jointly selects arm pairs to balance optimism and informative exploration. We establish a first-order regret bound that, in contrast with what previous results suggest, demonstrates that learning with preference feedback is as statistically efficient as learning from direct reward observation. Finally, we corroborate our theoretical findings with empirical evaluations against competitive baselines.

Deploying vision-and-language navigation (VLN) agents in unseen environments remains challenging because unfamiliar layouts and visual conditions can cause execution to go off track. Rather than relying on continuous human supervision, a practical strategy is to selectively request corrective guidance, recover the ongoing task, and reuse corrective interactions to improve subsequent navigation. We propose Abstention-aware Visual Error Recovery and Training for Vision-and-Language Navigation (AVERT-VLN), a closed-loop framework that uses a plug-in vision-language Monitor for online human-assisted recovery and offline preference learning. The Monitor operates separately from navigation decision generation and assesses instruction-execution consistency from the instruction, visual history, and current observation. To train the Monitor for deviation recognition, we construct LOSTNAV DATASET with 20K counterfactual risk trajectories and rule-based deviation labels. The Monitor is first fine-tuned on 40K normal trajectories to assess instruction progress and then jointly fine-tuned on normal and risk trajectories to recognize semantic deviations. At runtime, Asynchronous Sidecar Monitoring evaluates execution alongside the navigation model. When the controller accepts a LOST verdict, it suspends autonomous execution and requests human guidance for recovery. For offline policy improvement, Trajectory-Anchored Preference Learning converts deviation-associated failures into decision-level preference pairs under shared decision contexts, restricting supervision to the decisions targeted for correction. Under human-assisted evaluation, the full AVERT-VLN system achieves success rates of 76.2% and 66.3% on the val-unseen splits of R2R-CE and RxR-CE, respectively. The same monitoring and human-assisted recovery interface also improves success rates across the three evaluated navigation architectures.

Material generation should produce not only an appearance, but also the rules that construct it. We introduce MatLoom, a compact, layer-oriented language for text-to-material generation with pretrained language models. Each program composes alpha-masked layers whose shared spatial expressions define coverage and physically based rendering (PBR) channels, making dependencies between patterns, color, and relief explicit. A standalone interpreter evaluates the program into material maps, while the source retains named fields and layer parameters for subsequent authoring. Without task-specific fine-tuning, our pipeline uses parser-guided repair and preview-based critique to revise material designs, then searches noise seeds while keeping each candidate's remaining source fixed. On a curated benchmark of 141 prompts evaluated with six backbones, our best-performing configuration achieves higher mean scores than three diffusion baselines on all four flat-layout prompt-alignment metrics. Its initial programs already exceed all three baselines on mean BLIPScore, before critique or seed search. Retained programs have a median length of 21 lines when pooled across backbones. In a blind four-way comparison involving 30 participants and 20 prompts, our renders receive 59.2% of choices, compared with 19.3% for the most-preferred baseline. Compact executable programs thus offer a way to generate prompt-aligned materials while retaining their construction as part of the asset.

Reconstruction errors in multivariate time-series anomaly detection may not reliably distinguish abnormal behavior from benign deviations. Language-derived semantics offer complementary context, but existing multimodal approaches may rely on time-associated paired textual information that is difficult to obtain consistently and is not provided by standard multivariate time-series anomaly detection benchmarks. This setting poses two challenges: (1) conditioning masked reconstruction on window-specific semantics without exposing exact numerical targets or anomaly-specific cues, and (2) using a window-independent concept of normality as a complementary semantic reference rather than an independent anomaly detector. We propose LLM-Enhanced Alignment and Reconstruction with Normality Guidance for Time Series (LEARN-TS), which uses a frozen language model to construct two role-separated semantic representations without requiring temporally paired external text. Window-specific observation semantics encode temporal and cross-variable context without exact numerical values to guide channel-shared patch-masked reconstruction. A fixed, dataset-agnostic normality prompt provides a window-independent semantic reference for aligning normal representations and estimating normality discrepancy. At inference, masking each temporal patch once yields timestamp-level reconstruction evidence, conditionally modulated by discrepancy from a separate unmasked view. Across four benchmarks, LEARN-TS achieves the highest mean performance in 13 of 16 dataset-metric comparisons. Controlled ablations examine observation conditioning, joint normality alignment and scoring, and reference content, showing dataset-dependent ranking benefits and modest average gains from semantic over random references.

Standard language model RL algorithms credit every token of a long rollout with the same advantage determined by the terminal reward. Actor-critic methods can provide finer-grained credit assignment, but learned critics are generally considered too inaccurate to trust when training LLMs with RL. In recent works, even when a critic is present, it is used only for baseline estimation, so every trajectory must be rolled out to its terminal reward. We introduce Actor-Critic with Action Chunking (AC2) that removes the need to roll every trajectory to completion. AC2 instead assigns credit to action chunks: short continuations of prefixes of past trajectories. A learned critic scores the state reached at the end of each action chunk, allowing the policy to update without observing a terminal reward. We make critic-based credit assignment reliable through three design choices. First, we introduce local readiness which uses critic-based updates on a problem only when the critic is sufficiently accurate on that particular problem. Second, when available, we provide the critic with a reference solution from a previous successful rollout. Third, we assign credit over action chunks of 10k tokens rather than individual tokens, giving the critic a more meaningful portion of the trajectory to evaluate. We train Qwen3-4B on FineProofs-RL using AC2 and evaluate on IMO-ProofBench. AC2 exceeds GRPO's peak validation score of 18.5% using 2.5x fewer decoding FLOPs. This gain comes from two sources, (1) AC2 requires 25% fewer training steps to reach this score, and (2) each step generates fewer tokens because the policy does not need to continue every trajectory to completion. Conceptually, we demonstrate that we can remove the need to roll out every trajectory to completion, opening up a large previously unexplored design space for LLM RL algorithms.

Group Relative Policy Optimization (GRPO) has emerged as a memory-efficient reinforcement fine-tuning (RFT) technique for reasoning-intensive tasks. How- ever, GRPO training dynamics on small language models (SLMs) remain poorly understood, limiting its reliable adoption and reproducibility in open and resource- constrained environments. In this work, we present a systematic study of GRPO fine-tuning for SLMs ranging from 1.5B to 7B parameters under a practical single- node 8xA100 compute budget. Our study spans multiple model families and reasoning domains, including mathematics, coding, and multiple-choice question answering (MCQ) in science. Across these settings, we analyze how group size affects policy convergence, training stability, and downstream benchmark per- formance. We further characterize tensor-level update dynamics during GRPO training and investigate whether the choice of LoRA target modules and layers can improve the performance of GRPO-tuned models. While our initial GRPO-tuned models outperform their base counterparts on approximately 80% of mathematical benchmark evaluations, they demonstrate limited capability on MCQ and code reasoning tasks. Guided by our mechanistic evaluations, we refined our LoRA and reward-shaping configurations to improve performance in latter domains. These findings provide practical guidance for GRPO training for SLMs.

Multimodal encoders such as CLIP underlie many downstream systems, but their web-scale training data embed harmful associations that safety alignment must suppress without unnecessarily changing benign representations. Because ethical and practical constraints prevent collecting real unsafe content at scale, existing datasets pair safe real samples with generated counterparts, but label every generated sample unsafe, even when one modality is individually safe. To address this, we introduce ShieldCLIP, the first framework to condition safety alignment on the observed safety state of each modality rather than the origin of a sample, preserving safe content while redirecting only what is unsafe. We also introduce ViSUv2, a 195k-quadruplet dataset with independent per-modality safety labels across 578 concepts and 28 categories. Using these labels, ShieldCLIP defines a four-way conditional objective beyond pair-level supervision: safe content is anchored, unsafe modalities are redirected to their safe counterparts, mixed pairs update only the unsafe branch, and coherence is enforced when both are unsafe. We evaluate ShieldCLIP on cross-modal retrieval, text-to-image generation with Stable Diffusion v1.4 and SDXL, and image-to-text generation with LLaVA. Across these settings, ShieldCLIP consistently reduces harmful outputs over prior safety-aligned encoders and strong mitigation baselines, while preserving the utility of the original embedding space. Extensive ablation studies further show that both modality-specific supervision and the selective alignment objective contribute to these gains. Source code, trained models, and ViSUv2 (under a controlled-access protocol) will be made publicly available at https://aimagelab.github.io/ShieldCLIP/.

Reinforcement learning with human feedback (RLHF) learns from human comparisons, which can be corrupted or deliberately manipulated. This paper studies online risk-sensitive RLHF with static conditional value-at-risk (CVaR) under adversarial preference-label flips. We consider additive linear rewards and a fixed-reference protocol with one comparison per episode and at most $C$ flipped labels over $K$ episodes. We propose weighted streamed-preference CVaR RLHF (WSP-CVaR-RLHF), which combines uncertainty-weighted reward estimation with optimistic augmented-state CVaR planning. For known transitions and normalized rewards, we establish the regret bound $\widetilde{O}\left(\frac{d}κ\sqrt{\frac{K}α}+\frac{dC}{κα}\right)$ up to lower-order terms, where $d$ is the reward-feature dimension, $α$ is the CVaR level, and $κ$ characterizes the preference link. The bound separates the clean statistical cost from the penalty caused by corrupted feedback. We further extend the analysis to unknown tabular transitions, where the trajectory distribution entering the CVaR objective must be learned together with the reward. We address the resulting coupled uncertainty using rectangular transition confidence sets, joint optimistic planning, and a history-level CVaR simulation argument. Experiments under four adversarial attacks demonstrate that WSP-CVaR-RLHF consistently reduces cumulative regret relative to its unweighted robust counterpart while preserving confidence-set coverage.

Legal language models require reward signals that capture not only answer correctness but also the multidimensional quality of legal responses. Existing reward methods, however, often rely on coarse-grained holistic judgments, providing limited domain specificity and interpretability. We introduce LexReward, a taxonomy-driven framework for legal reward modeling. LexReward characterizes legal response quality along three complementary dimensions: Style, covering lexical and syntactic quality; Element, assessing legal subjects, facts, statutes, and decisions; and Chain, evaluating the order, completeness, correctness, and non-redundancy of legal reasoning. For each dimension, we develop rubrics that specify evaluation criteria and quality levels. The resulting rewards are used to construct pairwise preference data for Direct Preference Optimization (DPO) and reward-model training. Experiments show that the rubric-based rewards reliably distinguish legal responses of different quality and that DPO training on the preference data improves performance across all three dimensions. The learned reward models, LexRM, also support effective downstream optimization: each dimension-specific reward model improves policy performance in its corresponding dimension through reinforcement learning, without requiring reference answers at reward time. Dimension-wise analyses further support the effectiveness of the proposed taxonomy and reward construction.

Training data attribution (TDA) estimates the contribution of individual training examples to model outputs. Most scalable TDA methods rely on per-example gradients, whose computation and use at LLM scale pose challenges in efficiency, compatibility, and extensibility. We introduce dattri-LLM, a TDA library that makes gradient-based attribution more practical at scale. For efficiency, dattri-LLM uses compact gradient representations and dynamically routes gradient operations based on a cost model. For compatibility, its capture mechanism collects per-example gradients from existing training loops that call backward(), without requiring changes to the loop or its configuration. This includes distributed training with DDP and FSDP and pipelines built with HuggingFace Transformers, TRL, and OLMo. For extensibility, dattri-LLM exposes reusable gradient operations and training-time callbacks for implementing attribution methods and applications. These interfaces support a variety of attribution methods, including gradient similarity, curvature-based influence, and trajectory-based methods, as well as applications that act on gradients during training, such as online data selection. On the same hardware and workload, dattri-LLM achieves 3.2x the throughput of the fastest competing library on average, scales multiple attribution methods to 110B-parameter models across four H200 GPUs, and offers superior attribution fidelity-cost trade-offs across a range of models with different model families and scales. The source code of dattri-LLM is available at https://github.com/TRAIS-Lab/dattri-llm.

As model sizes continue to scale, distributed training has become inevitable. Automatic parallelization techniques can derive efficient training parallelism strategies at low cost while achieving superior performance. The difficulty of this problem is jointly determined by the complexity of the model and the underlying compute cluster. Meanwhile, mixture-of-experts (MoE) models are increasingly emerging as the dominant architecture and the rapid evolution of accelerator hardware has made cluster heterogeneity commonplace, posing substantial challenges to automatic parallelization. However, existing approaches typically target either MoE architectures or heterogeneous clusters, failing to generalize to scenarios where both challenges coexist. To this end, we present HAPMoE, a heterogeneity-aware automatic parallelism planner for MoE training. HAPMoE builds a lightweight MoE-aware cost model and efficiently searches a six-dimensional parallel space, producing parallel plans directly deployable on Megatron-LM. Experiments show that HAPMoE improves end-to-end training throughput by up to 3.2$\times$ over baselines across heterogeneous clusters. Its non-uniform pipeline partitioning yields an additional up to 78% gains, and its pruning-enhanced dynamic programming algorithm completes the search within 1 minute, demonstrating high efficiency and practical value in complex hardware environments.

Policy mirror descent (PMD) enjoys fast convergence in regularized Markov decision processes (MDPs), but existing guarantees often rely on exact or increasingly accurate policy evaluation. We analyze PMD coupled with a persistent critic advanced by one temporal-difference (TD) update. For finite discounted MDPs, we establish global linear convergence in value for exact coordinate-wise Bellman updates, with any positive constant actor stepsize and arbitrary finite critic initialization. The proof combines a resolvent-based auxiliary distribution with a decaying Bellman-violation correction and a potential weighted by inverse coordinate weights. We then study stochastic TD-PMD with general strongly convex mirror maps under a single off-policy Markov trajectory. With suitably chosen constant stepsizes and a finite-batch TD update, the method achieves an expected value gap of $ε$ after $\widetilde{O}(1/((1-γ)^5 \widetildeσ_b ε))$ transitions. The stochastic analysis relies on the trajectory-wise Lipschitz continuity of the regularizer, derived from uniform bounds on vertex Bregman divergences, together with a visitation-weighted resolvent estimate for signed critic-error propagation that yields an inverse-linear dependence on behavior coverage $\widetildeσ_b$. In contrast to many prior guarantees for regularized policy optimization, our sample-complexity guarantee holds without trajectory resets, generative-model access, or nested policy-evaluation loops. Numerical results are consistent with the theoretical convergence analysis.

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.

Generative image compression produces visually plausible reconstructions at low bitrates, yet their behavior as the rate approaches zero remains largely unexplored. When pushed below normal operating rates, representative codecs undergo semantic collapse: rather than gracefully losing source-specific detail, they produce malformed or unrecognizable content. Our analysis identifies two factors. As the bitrate decreases, reconstruction losses increasingly conflict with semantic objectives on gradients and visual results, while pixel-space and reconstruction-oriented VAE diffusion models become less efficient on semantic preservation. Guided by these findings, we introduce RAE-CoD, a compression-oriented diffusion (CoD) built in a representation autoencoder (RAE) space with direct alignment between compressed and source representations, preserving recognizable, naturally structured content for a $256\times256$ image with as few as 16 bits. We evaluate this framework using five vision foundation models (VFM) and a blinded vision-language model protocol. On MSCOCO-30K, RAE-CoD stands out from all evaluation. At 0.001-0.008 bpp, it reduces relative VFM feature MSE and Fréchet Distance ratio by at least 25.7% and 69.1% over the best competitors. Meanwhile, semantic recognizability and quality of the reconstructions remain nearly constant while source consistency falls smoothly, replacing abrupt semantic collapse with a graceful transition toward unconditional generation. Code will be released at https://github.com/LuizScarlet/RAE-CoD.

Sampling several responses and voting over their answers can improve a language model's accuracy, but repeated answers limit the benefit of additional samples. Raising temperature increases diversity at a potential cost to per-sample accuracy. We introduce Self-Repulsion (SR), a sampler for masked diffusion language models that uses peer commitments to diversify the pool. At each penalized denoising step, each path lowers a token's logit according to how many peers have committed that token at the same position. Paths share a batched forward pass and then commit in sequence, so later paths observe choices made earlier in the same step. This coupling requires no training or additional forward or backward pass and can produce distinct paths even at temperature zero. When all paths commit a position together from identical logits, the update exactly maximizes total logit minus a convex duplication cost. On LLaDA-8B-Instruct with ten paths and 128 denoising steps, deterministic SR reaches 80.38% plurality accuracy on GSM8K, compared with 70.17% for the unpenalized greedy decoder. At temperature 0.6 and matched model-evaluation budgets, the count penalty improves over self-consistency by 2.06 percentage points in blocks of 32 and 14.50 under pure diffusion. Experiments on GSM8K, MATH and TruthfulQA show that voting gains arise mainly from higher coverage of correct answers, with gains that vary by benchmark and decoding regime.

Terminal agents rely on self-verification to assess and correct their solutions as they solve tasks through interaction with command-line environments. Yet how trustworthy such self-verification is remains poorly understood. To investigate this question systematically, we introduce a diagnostic framework that identifies the first complete solution in each trajectory, determines whether it is objectively correct, and uses this ground truth to quantify the agent's subsequent verification and recovery behavior. Applying it to ten terminal agents on TerminalBench2.1, we find that verification is nearly universal after a complete candidate is formed, yet only 61.43% of incorrect candidates are detected and only 49.36% of detected errors are successfully repaired. These results show that the main weakness in self-verification lies not in initiating verification, but in detecting and repairing errors. Motivated by these findings, we propose Student-Conditioned Verification Distillation (SCVD), which lets the student first produce a candidate solution and distills a stronger teacher's subsequent verification and recovery from the same interaction context. Across three Qwen3.5 backbones, SCVD improves \textsc{Pass@1} on TerminalBench2.1 by 9.74--16.85 percentage points over the corresponding base models and by 4.49--8.61 points over the standard full-trajectory distillation, while avoiding the pronounced out-of-distribution degradation of full-trajectory distillation on SWE-bench Verified.

Passenger access doors are safety-critical subsystems in railway vehicles, yet detecting abnormal door behavior in real operation is challenging because faults are rare, diverse, and often unlabeled. This paper addresses railway door condition monitoring as a cycle-level unsupervised anomaly detection problem, where each complete opening-dwell-closing cycle is treated as a single monitoring unit. We propose the Temporal Cycle-Aware Attention Autoencoder with Cross-Signal Consistency (TCAA-CS), trained exclusively on nominal cycles. It combines a dual-stream encoder that processes continuous physical measurements (position, current, voltage) and binary logical states (door-closed, door-locked) through separate 1D-CNN branches, an LSTM encoder with temporal attention pooling, and a triple hybrid anomaly score fusing reconstruction error, latent-space deviation, and phase-aware cross-signal consistency. The consistency term helps identify cases where individual signals appear plausible but their inter-signal relationships become physically or logically inconsistent. On real industrial data from a passenger train in commercial service, TCAA-CS achieves 93.8% recall, 97.3% precision, and a 0.5% false-alarm rate, outperforming representative unsupervised baselines. System-level evaluation on an NVIDIA Jetson AGX Xavier supports the feasibility of real-time onboard deployment.

Despite the widespread use of Low-Rank Adaptation (LoRA), little is known about its dynamics in continual learning and the mechanisms by which low-rank updates affect catastrophic forgetting. We provide an asymptotically exact dynamical characterization of LoRA in a solvable two-task teacher-student model. In the high-dimensional online-learning limit, we derive a closed system of ordinary differential equations for a finite set of macroscopic order parameters, yielding exact expressions for the generalization errors throughout both the initial Task 1 learning phase and the subsequent LoRA fine-tuning on Task 2. The theory quantitatively matches finite-dimensional simulations and exposes two characteristic effects of LoRA: low-rank adaptation reduces interference with features learned on the first task, but its initialization slows adaptation to the second task. Building on this mechanistic picture, we analyze a state-dependent masking strategy that freezes hidden units carrying the strongest first-task representations and restricts adaptation to the complementary subspace. This structural partitioning markedly reduces forgetting, while preserving plasticity on the new task. Our framework further clarifies the role of adapter rank: transfer improves only up to the intrinsic dimensionality of the target task and saturates beyond it, while forgetting continues to grow with rank. These results provide a dynamical and geometric account of how low-rank adaptation organizes information across sequential tasks and are qualitatively reproduced on a sequential MNIST benchmark.

Few-Shot Class-Incremental Learning (FSCIL) addresses the challenge of learning new classes from very limited samples while retaining knowledge of previously learned ones. Although parameter-efficient fine-tuning methods with pre-trained models show promise for class-incremental learning, strict gradient-based constraints can be unreliable under severe data scarcity, while multi-expert approaches can impose substantial inference-time costs. We propose TALON (Task-Adaptive LoRA-Teachers with Ensemble Knowledge Transfer), an inference-efficient FSCIL framework. TALON dynamically allocates an independent LoRA-Teacher to each incremental task for task-specific representation learning, then distills multiple frozen teachers into a unified LoRA-Student through Ensemble Knowledge Transfer, eliminating runtime module selection or generation. A semantic-guided distillation strategy weights teacher contributions by feature-space similarity to mitigate catastrophic forgetting and overfitting. Across three class-order runs, TALON achieves comparable or better mean average accuracy across four FSCIL benchmarks, obtaining 86.68 +/- 1.22% on CUB200, 90.39 +/- 0.27% on CIFAR100, 78.38 +/- 0.94% on ImageNet-R, and 96.34 +/- 0.33% on miniImageNet. TALON uses up to 33x fewer deployment parameters and reduces average inference time per task to 26.7 s, a 41.70% reduction relative to ASP.

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.

We study training LLM judges from natural language feedback, especially for subjective tasks where the verdict depends strongly on which evaluation criteria the judge invokes and how it weighs them. The dominant approach, outcome-supervised RL (e.g., GRPO), credits every token in the rollout with a single scalar determined only by the accuracy of the final verdict, providing no separate credit at the criterion-choice tokens and ignoring the rich language feedback (e.g., preference rationales) that naturally accompanies preference labels. Self-Distillation (SD) is one natural way to use this language feedback: the same model, conditioned on this feedback, acts as a teacher providing dense, position-level supervision. However, not all positions carry equally useful signal. Using the per-position entropy shift between teacher and student, we identify two regimes: context sharpening, where the teacher concentrates probability on a particular feedback-aligned criterion expression, and context spreading, where the teacher distributes probability across multiple feedback-aligned alternatives. We interpret these patterns as follows: sharpening encourages memorization of a particular criterion expression, whereas spreading promotes semantic understanding by preserving these alternatives. Motivated by this asymmetry, we introduce position masking based on the entropy shift that retains the lower tail of the entropy-shift distribution. Experiments show that masking higher-entropy-shift positions improves out-of-distribution generalization over naive SD. The resulting self-distilled judges outperform judges trained with outcome-supervised RL by 2-9 percentage points on the evaluated subjective subcategories, while remaining competitive on objective ones.

On-policy self-distillation (OPSD) trains a student on its own generated responses using feedback from the same model conditioned on privileged information. On mathematical reasoning, the original OPSD study finds that stylistic tokens can dominate the training signal over math-related tokens, and that pointwise clipping of the forward KL objective stabilizes training. Pointwise clipping caps each vocabulary-wise forward KL term at a fixed threshold before summing over the vocabulary. Follow-up studies have adopted this clipping, but its effect on training has not been directly examined. In matched training runs differing only in whether clipping is applied, we observe that clipped runs produce substantially more repetitions that persist to the end of the response than their unclipped counterparts. We trace this failure to the clipped objective. We prove that the clipped objective can fail to correct the student toward the teacher and can instead push clipped and unclipped token probabilities away from its teacher. Our training runs agree with this analysis: inside repetitions, the clipped student places less probability than its teacher on leaving the repetition, and more on continuing it, whereas the unclipped runs stay close to their teachers.

On-policy self-distillation (OPSD) has attracted growing interest as a promising approach to improve the reasoning ability of language models. Without external rewards nor a separate stronger teacher, the self-teacher with privileged information could provide dense signals on student's trajectories. However, its behavior in language reasoning remains unclear, with reported outcomes ranging from modest gains to behavioral collapse. In this work, we diagnose OPSD for mathematical reasoning across models spanning 0.6B--8B parameters. We conduct controlled experiments and token-level analyses to fully delve into OPSD. We point out that teacher's signal is shaped by reasoning-mode alignment and the complete teacher prefix, rather than by privileged semantics alone. OPSD improves reasoning only in narrow compatibility regimes. Otherwise, it produces ineffective length growth, stable degradation, or behavioral collapse. Token-level analysis shows that teacher's signal is not stable and does not predict downstream performance. Based on these results, we argue that OPSD is a sensitive algorithm rather than a generally reliable reasoning-improvement post-training method.

On-policy distillation (OPD) improves reasoning by providing token-level supervision from a teacher on a student's own trajectories. Existing methods primarily focus on enhancing this teacher-side guidance (e.g., by enriching teacher inputs and refining teacher feedback), yet we find that limited student perception is another critical bottleneck in multimodal OPD. By providing oracle visual facts, the performance of OPD-trained students can still be substantially improved for both weak and strong teachers. To address this bottleneck, we propose S-OPD, a simple multimodal on-policy distillation framework that explicitly strengthens student perceptual learning through two objectives. Specifically, Teacher-calibrated Policy Contrast separates student policies under original and masked images with teacher-based token-level gating, strengthening the student's reliance on visual evidence during reasoning. Policy Agreement aligns student policies under original and noise-perturbed images, further improving perceptual robustness to visual noise. Notably, our method can be seamlessly plugged into existing OPD frameworks, requiring no additional data annotations, model parameters or inference operations. Extensive experiments on eight benchmarks across student scales and distillation paradigms demonstrate consistent performance improvements, with gains of up to 4.25 points on LogicVista. When combined with existing teacher-side supervision methods, our method can yield further gains. Code is available at https://github.com/Sirilaw/S-OPD.