Reinforcement Learning (RL) has traditionally focused on training specialized agents to optimize predefined reward functions within narrowly defined environments. However, the advent of powerful Large Language Models (LLMs) and increasingly complex, open-ended tasks has catalyzed a paradigm shift towards agentic paradigms within RL. This emerging framework extends beyond traditional RL by emphasizing the development of autonomous agents capable of goal-setting, long-term planning, dynamic strategy adaptation, and interactive reasoning in uncertain, real-world environments. Unlike conventional approaches that rely heavily on static objectives and episodic interactions, LLM-based Agentic RL incorporates cognitive-like capabilities such as meta-reasoning, self-reflection, and multi-step decision-making directly into the learning loop. In this paper, we provide a deep insight for looking the conceptual foundations, methodological innovations, and effective designs underlying this trend. Furthermore, we identify critical challenges and outline promising future directions for building LLM-based Agentic RL.

Diffusion large language models generate text through multi-step denoising, where hallucination signals may emerge throughout the trajectory rather than only in the final output. Existing detectors mainly rely on output uncertainty or coarse trace statistics, which often fail to capture the richer hidden dynamics of D-LLMs. We propose HIVE, a hidden-evidence verification framework that extracts compressed hidden evidence from denoising trajectories, selects informative step-layer evidence, and conditions a verifier language model on the selected evidence through prefix embeddings. HIVE produces both a continuous hallucination score from verifier decision logits and structured verification outputs, including hallucination types, evidence pairs, and short rationales. Across two D-LLMs and three QA benchmarks, HIVE consistently outperforms eight strong baselines and achieves up to 0.9236 AUROC and 0.9537 AUPRC. Ablation studies further confirm the importance of hidden-evidence conditioning, learned evidence selection, two-stream evidence representation, and step-layer embeddings. These results suggest that selected hidden evidence from denoising trajectories provides a stronger and more usable hallucination signal than output-only uncertainty or coarse trace statistics.

This paper presents a comprehensive review of the NTIRE 2026 Low Light Image Enhancement Challenge, highlighting the proposed solutions and final results. The objective of this challenge is to identify effective networks capable of producing clearer and visually compelling images in diverse and challenging conditions by learning representative visual cues with the purpose of restoring information loss due to low-contrast and noisy images. A total of 195 participants registered for the first track and 153 for the second track of the competition, and 22 teams ultimately submitted valid entries. This paper thoroughly evaluates the state-of-the-art advances in (joint denoising and) low-light image enhancement, showcasing the significant progress in the field, while leveraging samples of our novel dataset.

Cross-dose denoising for low-dose positron emission tomography (LDPET) has been proposed to address the limited generalization of models trained at a single noise level. However, neural networks trained on a specific dose level often fail to generalize to other dose conditions due to variations in noise magnitude and statistical properties. Conventional "one-size-for-all" models attempt to mitigate this variability but tend to learn averaged representations across noise levels, resulting in degraded performance. In this work, we analyze this limitation and show that standard training formulations implicitly optimize an expectation over heterogeneous noise distributions, causing the network to learn an averaged denoising mapping that cannot accurately model dose-specific noise characteristics. We propose a unified residual noise learning framework that estimates noise directly from low-dose PET images rather than predicting full-dose images. Experiments on large-scale multi-dose PET datasets from two medical centers demonstrate that the proposed method outperforms the "one-size-for-all" model, individual dose-specific U-Net models, and dose-conditioned approaches, achieving improved denoising performance. These results indicate that residual noise learning effectively mitigates the averaging effect and enhances generalization for cross-dose PET denoising.

We propose a framework for vision-based human pose estimation and motion prediction that gives conformal prediction guarantees for certifiably safe human-robot collaboration. Our framework combines aleatoric uncertainty estimation with OOD detection for high probabilistic confidence. To integrate our pipeline in certifiable safety frameworks, we propose conformal prediction sets for human motion predictions with high, valid confidence. We evaluate our pipeline on recorded human motion data and a real-world human-robot collaboration setting.

We present DMax, a new paradigm for efficient diffusion language models (dLLMs). It mitigates error accumulation in parallel decoding, enabling aggressive decoding parallelism while preserving generation quality. Unlike conventional masked dLLMs that decode through a binary mask-to-token transition, DMax reformulates decoding as a progressive self-refinement from mask embeddings to token embeddings. At the core of our approach is On-Policy Uniform Training, a novel training strategy that efficiently unifies masked and uniform dLLMs, equipping the model to recover clean tokens from both masked inputs and its own erroneous predictions. Building on this foundation, we further propose Soft Parallel Decoding. We represent each intermediate decoding state as an interpolation between the predicted token embedding and the mask embedding, enabling iterative self-revising in embedding space. Extensive experiments across a variety of benchmarks demonstrate the effectiveness of DMax. Compared with the original LLaDA-2.0-mini, our method improves TPF on GSM8K from 2.04 to 5.47 while preserving accuracy. On MBPP, it increases TPF from 2.71 to 5.86 while maintaining comparable performance. On two H200 GPUs, our model achieves an average of 1,338 TPS at batch size 1. Code is available at: https://github.com/czg1225/DMax

Reinforcement learning (RL) is a core approach for robot control when expert demonstrations are unavailable. On-policy methods such as Proximal Policy Optimization (PPO) are widely used for their stability, but their reliance on narrowly distributed on-policy data limits accurate policy evaluation in high-dimensional state and action spaces. Off-policy methods can overcome this limitation by learning from a broader state-action distribution, yet suffer from slow convergence and instability, as fitting a value function over diverse data requires many gradient updates, causing critic errors to accumulate through bootstrapping. We present FlashSAC, a fast and stable off-policy RL algorithm built on Soft Actor-Critic. Motivated by scaling laws observed in supervised learning, FlashSAC sharply reduces gradient updates while compensating with larger models and higher data throughput. To maintain stability at increased scale, FlashSAC explicitly bounds weight, feature, and gradient norms, curbing critic error accumulation. Across over 60 tasks in 10 simulators, FlashSAC consistently outperforms PPO and strong off-policy baselines in both final performance and training efficiency, with the largest gains on high-dimensional tasks such as dexterous manipulation. In sim-to-real humanoid locomotion, FlashSAC reduces training time from hours to minutes, demonstrating the promise of off-policy RL for sim-to-real transfer.

In robot control, planning, and learning, there is a need for rigid-body dynamics libraries that are highly performant, easy to use, and compatible with CPUs and accelerators. While existing libraries often excel at either low-latency CPU execution or high-throughput GPU workloads, few provide a unified framework that targets multiple architectures without compromising performance or ease-of-use. To address this, we introduce frax, a JAX-based library for robot kinematics and dynamics, providing a high-performance, pure-Python interface across CPU, GPU, and TPU. Via a fully-vectorized approach to robot dynamics, frax enables efficient real-time control and parallelization, while supporting automatic differentiation for optimization-based methods. On CPU, frax achieves low-microsecond computation times suitable for kilohertz control rates, outperforming common libraries in Python and approaching optimized C++ implementations. On GPU, the same code scales to thousands of instances, reaching upwards of 100 million dynamics evaluations per second. We validate performance on a Franka Panda manipulator and a Unitree G1 humanoid, and release frax as an open-source library.

Agent skills, structured packages of procedural knowledge and executable resources that agents dynamically load at inference time, have become a reliable mechanism for augmenting LLM agents. Yet inference-time skill augmentation is fundamentally limited: retrieval noise introduces irrelevant guidance, injected skill content imposes substantial token overhead, and the model never truly acquires the knowledge it merely follows. We ask whether skills can instead be internalized into model parameters, enabling zero-shot autonomous behavior without any runtime skill retrieval. We introduce SKILL0, an in-context reinforcement learning framework designed for skill internalization. SKILL0 introduces a training-time curriculum that begins with full skill context and progressively withdraws it. Skills are grouped offline by category and rendered with interaction history into a compact visual context, teaching he model tool invocation and multi-turn task completion. A Dynamic Curriculum then evaluates each skill file's on-policy helpfulness, retaining only those from which the current policy still benefits within a linearly decaying budget, until the agent operates in a fully zero-shot setting. Extensive agentic experiments demonstrate that SKILL0 achieves substantial improvements over the standard RL baseline (+9.7\% for ALFWorld, +6.6\% for Search-QA, and+10.1\% for WebShop), while maintaining a highly efficient context of fewer than 0.5k tokens per step. Our code is available at https://github.com/ZJU-REAL/SkillZero.

We introduce the Mandarin-English Language Interview (MELI) Corpus, an open-source resource of 29.8 hours of speech from 51 Mandarin-English bilingual speakers. MELI combines matched sessions in Mandarin and English with two speaking styles: read sentences and spontaneous interviews about language varieties, standardness, and learning experiences. Audio was recorded at 44.1 kHz (16-bit, stereo). Interviews were fully transcribed, force-aligned at word and phone levels, and anonymized. Descriptively, the Mandarin component totals ~14.7 hours (mean duration 17.3 minutes) and the English component ~15.1 hours (mean duration 17.8 minutes). We report token/type statistics for each language and document code-switching patterns (frequent in Mandarin sessions; more limited in English sessions). The corpus design supports within-/cross-speaker, within/cross-language acoustic comparison and links acoustics to speakers' stated language attitudes, enabling both quantitative and qualitative analyses. The MELI Corpus will be released with transcriptions, alignments, metadata, scans of labelled maps and documentation under a CC BY-NC 4.0 license.

AI agents are becoming active decision-makers on the Internet. As they make decisions in the same environments as humans, the environments themselves can change to influence them. We call this $\textit{mecha-nudging}$: changes to how choices are presented that systematically influence AI agents without materially degrading the decision environment for humans. To measure this phenomenon, we combine two frameworks -- Bayesian persuasion from economics and $\mathcal{V}$-usable information from computer science -- to get a common unit (bits) for quantifying how environments change across a wide range of interventions, contexts, and models. We apply this framework to over six million Etsy listings and find that, after ChatGPT's release, listings contain significantly more machine-usable information for predicting agent curation decisions, increasing by 0.143 bits out of a maximum possible increase of 0.355. This shift is robust across prompts, token choices, labeling models, and fine-tuning architectures; absent in a regulated-text placebo; and far larger than the effect of generic LLM rewriting. In contrast, a human study finds little to no change in human-usable information. Our results provide the first large-scale evidence that systematic mecha-nudging is already occurring in the wild, but going unnoticed.

Geometric data augmentation is widely used in segmentation workflows, but polygon annotations are often assumed to remain valid after transformation. This assumption can fail in structured domains such as architectural floorplan analysis, where a region may contain an interior void encoded as part of a single ordered polygon chain. Cropping or clipping can remove bridge vertices in this chain, causing one semantic region to split into disconnected components. We propose a lightweight topology-preserving augmentation strategy that repairs missing adjacency relations in index space while preserving the original vertex order. The method adds minimal overhead and can be integrated into existing preprocessing workflows. Experiments show that the proposed approach achieves near-perfect Cyclic Adjacency Preservation (CAP) across common geometric transformations and improves annotation consistency in polygon-based segmentation.

Electroencephalogram (EEG) classification plays a key role in medical diagnosis and brain-computer interfaces, but remains challenging due to low signal-to-noise ratios and high inter-subject variability. As a result, many existing approaches rely on subject-specific models, which fail to exploit shared structure in neural signals and do not generalize to unseen subjects. To address these limitations, we propose LAtte, a framework that combines Lorentz attention with a hyperbolic InceptionTime-based encoder to improve cross-subject generalization in EEG classification. The model explicitly decomposes EEG signals into a learned baseline component and task-relevant deviations, enabling more structured representation learning. To further improve robustness and adaptability, we incorporate subject-specific low-rank adaptation (LoRA) modules at both encoder and decoder levels, augmented with a Lorentz boost-based LoRA mechanism and hyperbolic projection layers to reduce overfitting in geometric representations. We evaluate LAtte with and without finetuning in three settings: subject-specific, subject-conditional, and leave-one-subject-out (LOSO) on five established EEG datasets, achieving a consistent improvement in performance over current state-of-the-art methods for smaller datasets and maintaining performance for larger datasets.

Cross-view UAV geolocalization is fundamentally a challenging large-scale image retrieval task, aiming to determine the geographic coordinates of Unmanned Aerial Vehicle (UAV) queries by matching them against an extensive geo-tagged satellite image database. Most existing methods learn separate feature representations for each view and determine the final prediction using naive heuristics to assess feature similarity, thereby neglecting to model the crucial cross-view relationships. In this paper, we propose SkyLink, a novel plug-and-play ranking framework that pioneers joint relational modeling of inter-view relationships to enhance cross-view UAV geolocalization. SkyLink leverages a Large Vision-Language Model (LVLM) to model the intricate visual-semantic relationships between UAV and satellite views, facilitating effective cross-view matching. To further refine the learning process, we introduce a relational-aware loss. It leverages soft labels to provide a more nuanced supervision signal, mitigating the harsh penalty on near-positive pairs. This approach enhances both training stability and the model's discriminative capacity. Extensive experiments conducted across multiple base retrieval architectures and benchmark datasets demonstrate that SkyLink significantly boosts the ranking effectiveness of existing models, consistently achieving superior performance in various challenging scenarios.

Drifting models train one-step generators by optimizing a kernel-induced mean-shift discrepancy between the data and model distributions, with Laplace kernels used by default in practice. At each point, this discrepancy compares the kernel-weighted displacement toward nearby data samples with the corresponding displacement toward nearby model samples, thereby defining a transport direction for generated samples. In this paper, we show that drifting is more closely connected to score-based generative modeling than it may first appear, establishing a precise link to the score-matching principle underlying diffusion models. For Gaussian kernels, the population mean-shift field exactly equals the difference between the scores (i.e., the gradient-log-densities) of the Gaussian-smoothed data and model distributions. This identity follows from Tweedie's formula, which links the score of a Gaussian-smoothed density to its conditional mean, and implies that Gaussian-kernel drifting is exactly a score-matching objective on smoothed distributions. More generally, we derive an exact decomposition for radial kernels in which mean shift equals a score-based field plus a residual term. For the practical Laplace kernel, we further show theoretically and empirically that this residual is negligible in high dimension, implying that the transport field used in practice is nearly score-based. Our results reveal a structural connection to diffusion models: both methods use score-mismatch transport directions, but drifting realizes the score nonparametrically through kernel-based estimates, whereas diffusion models learn it parametrically with neural networks.

We present Whole-Body Mobile Manipulation Interface (HoMMI), a data collection and policy learning framework that learns whole-body mobile manipulation directly from robot-free human demonstrations. We augment UMI interfaces with egocentric sensing to capture the global context required for mobile manipulation, enabling portable, robot-free, and scalable data collection. However, naively incorporating egocentric sensing introduces a larger human-to-robot embodiment gap in both observation and action spaces, making policy transfer difficult. We explicitly bridge this gap with a cross-embodiment hand-eye policy design, including an embodiment agnostic visual representation; a relaxed head action representation; and a whole-body controller that realizes hand-eye trajectories through coordinated whole-body motion under robot-specific physical constraints. Together, these enable long-horizon mobile manipulation tasks requiring bimanual and whole-body coordination, navigation, and active perception. Results are best viewed on: https://hommi-robot.github.io

Deployed reinforcement learning systems lack a principled runtime reliability theory. We close this gap by introducing Bipredictability, P, a closed form information theoretic metric that quantifies how efficiently a closed loop interaction between agent and environment converts uncertainty into shared predictability. P admits a provable classical bound P equal, smaller than 0.5, derived from Shannon entropy subadditivity, and responsive agency necessarily suppresses P below this ceiling, a structural prediction we term the informational cost of agency. Across 21 trained continuous control agents, we confirm this prediction empirically at P = 0.33 plus minus 0.02. The same suppression signature reproduces in language model dialogue, convolutional vision systems, and classical mechanical baselines, indicating that P captures a substrate independent property of agentic interaction rather than an algorithm specific artifact. The Information Digital Twin, IDT, a model agnostic architecture that computes P from the external interaction stream, detects 89.3% of coupling degradations against 44.0% for reward based monitoring, with 4.4 times lower latency. P provides the missing measurement layer for runtime reliability and closed loop self regulation in deployed autonomous systems.

Angle-encoded variational quantum circuits admit a truncated Fourier series representation of their output, but approximating functions with maximum frequency $ω_{\max}$ using fixed unary encoding requires $\mathcal{O}(ω_{\max})$ encoding gates. Trainable-frequency (TF) circuits promise a reduction by learning the data-encoding prefactors alongside the ansatz parameters, adapting the accessible frequency spectrum to the target during training. We identify a practical barrier that prevents this promise from being realized: the prefactor gradient is suppressed by the spectral gap between the circuit's accessible frequencies and the target spectrum, independently of the ansatz parameters, confining gradient-driven prefactor movement to a narrow neighborhood of initialization. We propose \emph{ternary grid initialization} -- setting prefactors to $\{1, 3, 9, \ldots, 3^{k-1}\}$ -- which resolves this limitation by ensuring every target frequency within $[-ω_{\max}, ω_{\max}]$ lies within $\tfrac{1}{2}$ unit of a grid point at initialization, removing the spectral gap suppression by construction. On a synthetic benchmark with target frequencies shifted well beyond the standard initialization range, ternary initialization achieves median $R^2 = 0.997$ versus $0.18$ for unary initialization, with $100\%$ of runs achieving $R^2 > 0.95$ against $0\%$. CMA-ES with $20\times$ the evaluation budget reaches only $25\%$ success, confirming the limitation is a property of the optimization landscape rather than of gradient-based optimization specifically. Real-world validation on two benchmark datasets demonstrates consistent advantages over both fixed and trainable unary baselines.

Vision-language models (VLMs) can read text from images, but where does this optical character recognition (OCR) information enter the language processing stream? We investigate the OCR routing mechanism across three architecture families (Qwen3-VL, Phi-4, InternVL3.5) using causal interventions. By computing activation differences between original images and text-inpainted versions, we identify architecture-specific OCR bottlenecks whose dominant location depends on the vision-language integration strategy: DeepStack models (Qwen) show peak sensitivity at mid-depth (about 50%) for scene text, while single-stage projection models (Phi-4, InternVL) peak at early layers (6-25%), though the exact layer of maximum effect varies across datasets. The OCR signal is remarkably low-dimensional: PC1 captures up to 72.9% of variance. Crucially, principal component analysis (PCA) directions learned on one dataset transfer to others, demonstrating shared text-processing pathways. Surprisingly, in models with modular OCR circuits (notably Qwen3-VL-4B), OCR removal can improve counting performance (up to +6.9 percentage points), suggesting OCR interferes with other visual processing in sufficiently modular architectures.

Keypoint-based matching is a fundamental component of modern 3D vision systems, such as Structure-from-Motion (SfM) and SLAM. Most existing learning-based methods are trained on image pairs, a paradigm that fails to explicitly optimize for the long-term trackability of keypoints across sequences under challenging viewpoint and illumination changes. In this paper, we reframe keypoint detection as a sequential decision-making problem. We introduce TraqPoint, a novel, end-to-end Reinforcement Learning (RL) framework designed to optimize the \textbf{Tra}ck-\textbf{q}uality (Traq) of keypoints directly on image sequences. Our core innovation is a track-aware reward mechanism that jointly encourages the consistency and distinctiveness of keypoints across multiple views, guided by a policy gradient method. Extensive evaluations on sparse matching benchmarks, including relative pose estimation and 3D reconstruction, demonstrate that TraqPoint significantly outperforms some state-of-the-art (SOTA) keypoint detection and description methods.The code will be available at https://github.com/xiaomi-research/traqpoint.

Knowledge editing in Large Language Models (LLMs) aims to update the model's prediction for a specific query to a desired target while preserving its behavior on all other inputs. This process typically involves two stages: identifying the layer to edit and performing the parameter update. Intuitively, different queries may localize knowledge at different depths of the model, resulting in different sample-wise editing performance for a fixed editing layer. In this work, we hypothesize the existence of fixed golden layers that can achieve near-optimal editing performance similar to sample-wise optimal layers. To validate this hypothesis, we provide empirical evidence by comparing golden layers against ground-truth sample-wise optimal layers. Furthermore, we show that golden layers can be reliably identified using a proxy dataset and generalize effectively to unseen test set queries across datasets. Finally, we propose a novel method, namely Layer Gradient Analysis (LGA) that estimates golden layers efficiently via gradient-attribution, avoiding extensive trial-and-error across multiple editing runs. Extensive experiments on several benchmark datasets demonstrate the effectiveness and robustness of our LGA approach across different LLM types and various knowledge editing methods.

Recent years have witnessed tremendous progress in enabling LLMs to solve complex reasoning tasks such as math and coding. As we start to apply LLMs to harder tasks that they may not be able to solve in one shot, it is worth paying attention to their ability to construct intermediate stepping stones that prepare them to better solve the tasks. Examples of stepping stones include simplifications, alternative framings, or subproblems. We study properties and benefits of stepping stones in the context of modern reasoning LLMs via ARQ (Asking the Right Questions), a simple framework that introduces a question generator to the default reasoning pipeline. We first show that good stepping stone questions exist and are transferrable, meaning that good questions can be generated, and they substantially help LLMs of various capabilities in solving the target tasks. We next frame stepping stone generation as a post-training task and show that we can fine-tune LLMs to generate more useful stepping stones by SFT and RL on synthetic data.

Linear attention transformers have become a strong alternative to softmax attention due to their efficiency. However, linear attention tends to be less expressive and results in reduced accuracy compared to softmax attention. To bridge the accuracy gap between softmax attention and linear attention, we manipulate Mamba-2, a very strong linear attention variant. We first simplify Mamba-2 down to its most fundamental and important components, evaluating which specific choices make it most accurate. From this simplified Mamba variant (Mamba-2S), we improve the A-mask and increase the order of the hidden state, resulting in a method, which we call 2Mamba, that is nearly as accurate as softmax attention, yet much more memory efficient for long context lengths. We also investigate elements to Mamba-2 that help surpass softmax attention accuracy. Code is provided for all our experiments.

Embedding tables are critical components of large-scale recommendation systems, facilitating the efficient mapping of high-cardinality categorical features into dense vector representations. However, as the volume of unique IDs expands, traditional hash-based indexing methods suffer from collisions that degrade model performance and personalization quality. We present Multi-Probe Zero Collision Hash (MPZCH), a novel indexing mechanism based on linear probing that effectively mitigates embedding collisions. With reasonable table sizing, it often eliminates these collisions entirely while maintaining production-scale efficiency. MPZCH utilizes auxiliary tensors and high-performance CUDA kernels to implement configurable probing and active eviction policies. By retiring obsolete IDs and resetting reassigned slots, MPZCH prevents the stale embedding inheritance typical of hash-based methods, ensuring new features learn effectively from scratch. Despite its collision-mitigation overhead, the system maintains training QPS and inference latency comparable to existing methods. Rigorous online experiments demonstrate that MPZCH achieves zero collisions for user embeddings and significantly improves item embedding freshness and quality. The solution has been released within the open-source TorchRec library for the broader community.

We study reward-free and reward-agnostic exploration in episodic finite-horizon Markov decision processes (MDPs), where an agent explores an unknown environment without observing external rewards. Reward-free exploration aims to enable $ε$-optimal policies for any reward revealed after exploration, while reward-agnostic exploration targets $ε$-optimality for rewards drawn from a small finite class. In the reward-agnostic setting, Li, Yan, Chen, and Fan achieve minimax sample complexity, but only for restrictively small accuracy parameter $ε$. We propose a new algorithm that significantly relaxes the requirement on $ε$. Our approach is novel and of technical interest by itself. Our algorithm employs an online learning procedure with carefully designed rewards to construct an exploration policy, which is used to gather data sufficient for accurate dynamics estimation and subsequent computation of an $ε$-optimal policy once the reward is revealed. Finally, we establish a tight lower bound for reward-free exploration, closing the gap between known upper and lower bounds.

We present the first regret bound for classical online Q-learning in infinite-horizon discounted Markov decision processes (MDPs), without relying on optimism or bonus terms. We first analyze Boltzmann Q-learning with decaying temperature and show that its regret depends critically on the suboptimality gap of the MDP: for sufficiently large gaps, the regret is sublinear, while for small gaps it deteriorates and can approach linear growth. To address this limitation, we study a Smoothed $ε_n$-Greedy exploration scheme that combines $ε_n$-greedy and Boltzmann exploration, for which we prove a gap-robust regret bound of near-$\tilde{O}(N^{9/10})$. We also obtain sample complexity guarantees, with both regret and sample complexity bounds holding with high probability. To analyze these algorithms, we develop a high-probability concentration bound for contractive Markovian stochastic approximation with iterate- and time-dependent transition dynamics. This bound may be of independent interest as the contraction factor in our framework is allowed to converge to one asymptotically.

Large-scale deep learning models are well-suited for compression. Across a variety of tasks, methods like pruning, quantization, and knowledge distillation have been used to achieve massive reductions in model parameters with only marginal performance drops. This raises the central question: *Why are deep neural networks suited for compression?* In this work, we take up the perspective of algorithmic complexity to explain this behavior. We hypothesize that the parameters of trained models have more structure and, hence, exhibit lower algorithmic complexity compared to the weights at (random) initialization. Furthermore, model compression methods harness this reduced algorithmic complexity to compress models. Although an unconstrained parameterization of model weights, $\mathbf{w} \in \mathbb{R}^n$, can represent arbitrary weight assignments, the solutions found during training exhibit repeatability and structure, making them simpler to implement than a trivial program. To this end, we formalize the Kolmogorov complexity of $\mathbf{w}$ by $\mathcal{K}(\mathbf{w})$. We introduce a constrained parameterization $\widehat{\mathbf{w}}$ that partitions parameters into blocks of size $s$ and restricts each block to be selected from a set of $k$ reusable motifs, specified by a reuse pattern (or mosaic). The resulting method, $\mathit{Mosaic\text{-}of\text{-}Motifs}$ (MoMos), provides a theoretically justified parameterization that biases optimization toward algorithmically simpler solutions. Empirical evidence from multiple experiments shows that MoMos consistently lowers the algorithmic complexity of neural networks during training while preserving the performance of unconstrained models. These results suggest that parameter compressibility is not only observed after training, but can be induced from the optimization domain.

Diffusion large language models (DLLMs) have emerged as powerful generative models with the promise of fast text generation through parallel decoding. However, realizing this potential in practice remains challenging: reducing the number of decoding steps, typically causes a substantial degradation in output quality due to token factorization error. To alleviate this, we propose a self-distillation framework that trains a few-step student to match the generative trajectory of a full-step teacher. We theoretically and empirically show that trajectory-level supervision mitigates this factorization error, thereby enabling effective few-step decoding. We further incorporate Direct Discriminative Optimization (DDO), a reverse-KL objective that encourages mode-seeking toward the teacher's modes, yielding stronger performance on challenging reasoning tasks. Across reasoning and code-generation benchmarks, our method substantially narrows the gap between few-step and full-step decoding. The source code is available at https://github.com/Tyrion58/T3D.

Existing forgery detection methods are often limited to uni-modal or bi-modal settings, failing to handle the interleaved text, images, and videos prevalent in real-world misinformation. To bridge this gap, this paper targets to develop a unified framework for omnibus vision-language forgery detection and grounding. In this unified setting, the {interplay} between diverse modalities and the dual requirements of simultaneous detection and localization pose a critical ``difficulty bias`` problem: the simpler veracity classification task tends to dominate the gradients, leading to suboptimal performance in fine-grained grounding during multi-task optimization. To address this challenge, we propose \textbf{OmniVL-Guard}, a balanced reinforcement learning framework for omnibus vision-language forgery detection and grounding. Particularly, OmniVL-Guard comprises two core designs: Self-Evolving CoT Generatio and Adaptive Reward Scaling Policy Optimization (ARSPO). {Self-Evolving CoT Generation} synthesizes high-quality reasoning paths, effectively overcoming the cold-start challenge. Building upon this, {Adaptive Reward Scaling Policy Optimization (ARSPO)} dynamically modulates reward scales and task weights, ensuring a balanced joint optimization. Extensive experiments demonstrate that OmniVL-Guard significantly outperforms state-of-the-art methods and exhibits zero-shot robust generalization across out-of-domain scenarios. The dataset and code are publicly available at https://github.com/shen8424/OmniVL-Guard.

Transformers update token representations through multi-head attention and residual connections as $X \leftarrow X + \sum_{i} P^{(i)}XW_{V_i}W_{o_i}$, where $P^{(i)}$ is the softmax attention matrix in head $i$. We propose replacing a subset of $P^{(i)}$'s with the Laplacian $I - P^{(i)}$, giving $X \leftarrow X + \sum_{i \in \mathcal{A}} P^{(i)}XW_{V_i}W_{o_i} + \sum_{i \in \mathcal{L}} (I - P^{(i)})XW_{V_i}W_{o_i}$. Our proposal has two motivations. First, it allows attention heads to update the mean of token representations, while Laplacian heads can directly control within-sequence variance. Second, if tokens are viewed as nodes in a graph with edge weights $P^{(i)}$, then $I - P^{(i)}$ is the corresponding graph Laplacian, and the update can be interpreted as one step of heat diffusion on the graph. We show that this simple modification improves performance across supervised learning, language modeling, and self-supervised learning tasks. To investigate why, we examine the token representations learned with and without Laplacian heads. In supervised learning, Laplacian heads collapse token representations within the same sequence and align the sequence means with the geometry of Neural Collapse. In language modeling, they increase the separability of token representations that share the same next-token prediction. In self-supervised learning, they produce token representations whose principal components are better suited for segmentation. Across modalities, they also lead to faster-decaying spectra, indicating stronger token smoothing. Overall, our findings challenge the prevailing view that token oversmoothing is inherently harmful, showing instead that certain forms of smoothing can be beneficial.