Regularization

Multi-Agent Proximal Policy Optimization (MAPPO) is a variant of the Proximal Policy Optimization (PPO) algorithm, specifically tailored for multi-agent reinforcement learning (MARL). MAPPO optimizes cooperative multi-agent settings by employing a centralized critic with decentralized actors. However, in case of multi-dimensional environment, MAPPO can not extract optimal policy due to non-stationary agent observation. To overcome this problem, we introduce a novel approach, Entropy Regularization-based Proximal Policy Optimization (ERPPO). For the policy optimization, we first define the object detection ambiguity under multi-dimensional observation environment. Distributional Spatiotemporal Ambiguity (DSA) learner is trained to estimate object detection uncertainty in non-stationary constraints. Then, we enhance PPO with a novel Entropy Regularization term. This regularization dynamically adjusts the policy update by applying a stronger (L1) regularization in high-ambiguity observation to encourage significant exploratory actions and a weaker (L2) regularization in low-ambiguity observation to stabilize the proximal policy optimization. This approach is designed to enhance the probability of successful object localization in time-critical operations by reducing detection failures and optimizing search policy. Experiments on a testbed with AirSim-based maritime searching scenarios show that the proposed ERPPO improves accuracy performance. Our proposed method improves higher gradient than MAPPO. Qualitative results confirm that ERPPO effectiveness in terms of suppressing false detection in visually uncertain conditions.

Classical communication systems fail not only through random noise but also when transmitter and receiver use incompatible operational codebooks. Variational autoencoders (VAEs) train an encoder $q_φ$ and decoder $p_θ$ jointly, and practitioners treat the resulting latent space as a discrete code -- for clustering, conditional generation, and mechanistic interpretability. Yet standard VAE diagnostics -- ELBO, active units, mutual information, and code histograms -- certify only whether this code is used, never whether the decoder reads each latent under the encoder's code. We close this gap with the neural codebook channel $K_{e\to d}(j\mid i)$, a coupled encoder-decoder diagnostic whose off-diagonal mass is bounded by an architecture-free Bernoulli-KL certificate $d_{\mathrm{bin}}(1-\mathcal{A} \,\|\, \barη_p) \le \barΔ$ controlled by the variational gap. The certificate is the operational specialization of the classical KL chain rule under disintegration to the encoder-decoder disagreement event, complemented by a constructive marginal-impossibility result: no combination of marginal histograms, entropies, active-code counts, or mutual information determines $K_{e\to d}$. We audit the certificate on four sklearn datasets (finite-grid exact, 5/5 seeds, 20/20 pairs satisfy the bound), a 2D model where the bound is non-vacuous at $2.71\times$ the observed disagreement and the four-term identity closes within $10^{-4}$, MNIST under importance-sampling control, and a VQ-VAE attaining the predicted limit $\hat{\mathcal{A}}=1.000$. The package $(K_{e\to d}, \mathcal{A}, R_{\mathrm{eff}}, R, \mathrm{AU})$ is an audit-ready reporting unit. More broadly, the framework makes mismatched decoding -- a failure mode classical communication theory named decades ago -- visible inside a single deep generative model.

We study the problem of learning Nash equilibria in offline two-player zero-sum Markov games. While existing approaches often rely on explicit pessimism to address distribution shift, we show that KL regularization alone suffices to stabilize learning and guarantee convergence. We first introduce Regularized Offline Sequential Equilibrium (ROSE), a theoretical framework that achieves a fast $\widetilde{\mathcal{O}}(1/n)$ convergence rate under \textit{unilateral concentrability}, improving over the standard $\widetilde{\mathcal{O}}(1/\sqrt{n})$ rates in unregularized settings. We then propose Sequential Offline Self-play Mirror Descent (SOS-MD), a practical model-free algorithm based on least-squares value estimation and iterative self-play updates. We prove that the last iterate of SOS-MD attains the same $\widetilde{\mathcal{O}}(1/n)$ statistical rate up to a vanishing optimization error of order $\widetilde{\mathcal{O}}(1/\sqrt{T})$ in the number of self-play iterations $T$.

Leaderboards rank frontier models on independent axes but do not reveal whether capabilities reinforce or trade off across releases -- and at the frontier, this interaction is the more informative signal. We decompose paired SWE-bench and GPQA Diamond scores into a population coupling trend and per-release residual ($h$-field) that diagnoses capability emphasis from two public benchmark scores. Across 34 models from 10 labs (2024--2026), capabilities cooperate ($r = +0.72$, $p < 10^{-6}$), but cooperation varies systematically: per-lab coupling slopes span $5\times$ (Google $1.15$ vs. DeepSeek $0.23$), and labs pivot -- DeepSeek reversed from reasoning-rich to coding-first ($Δh = 15.9$~pp); Anthropic oscillates between coding excursions and recovery. The population regression serves as an isocline phase boundary: the same $\sqrt{(a/b)\cdot B_1}$ classifier that identifies the base-scale coupling transition [Amin, 2026] classifies frontier models and already detects mixed-phase behavior at the next transition (two models below the GPQA--IFEval isocline). The $h$-field is not just diagnostic -- it tells you what to change. Pretraining establishes coupling at $0.871$ while RLHF adds $0.081$ [Amin, 2026]: pretraining-level shifts are permanent (DeepSeek's four-release reversal persists), post-training shifts are reversible (Anthropic's three coding excursions each recover within one release), and inference compute alone shifts $h$ by $+7.8$~pp without retraining. Knowing which component dominates determines whether to retrain or wait. We provide a three-step diagnostic (locate, classify, predict), a per-lab measurement-priority table, and seven falsifiable predictions with timestamped criteria. Five post-cutoff releases fall within the 95% prediction interval. Code, data, and an interactive dashboard: https://zehenlabs.com/cape/.

Financial transaction fraud prevention faces challenges such as complex relationship structures, concealed behavioral patterns, and dynamically changing data distribution. Discrimination models relying solely on independent sample features are insufficient to fully characterize the risks of group collaboration and chain transfers within transaction networks. This paper proposes a graph neural network representation learning and risk discrimination framework for financial transaction fraud prevention. It integrates transaction records and identity information into node attributes and constructs a transaction graph based on shared attributes and interaction consistency to explicitly model inter-transaction relationships. In model design, a multi-layer message passing mechanism is employed to aggregate neighborhood information, learn node embedding representations containing structural context semantics, and output transaction-level fraud probability and risk scores through a lightweight risk discrimination head. A weighted supervision objective is introduced to mitigate training bias caused by class imbalance, and structural consistency regularization constraints are combined to suppress the impact of noisy edges on representation drift, thereby improving the stability and usability of risk characterization. Experiments are conducted on a publicly available financial transaction dataset, comparing various methods in the same direction and comprehensively evaluating them under a unified evaluation protocol. The results show that the proposed method outperforms other methods in risk ranking and probability calibration quality, validating the effectiveness of graph structure modeling and representation learning collaboration in financial transaction fraud prevention.

INTRODUCTION | Fully supervised 3D segmentation of high-resolution ex vivo MRI is limited by the prohibitive cost of volumetric annotation, forcing reliance on sparse 2D slices. Weakly supervised Sparse-to-Dense frameworks bridge this gap, but guidelines remain ambiguous regarding human-centric visual enhancements and transferring optimization strategies across dimensions. We analyze divergent regularization needs for multi-class segmentation of high-resolution ex vivo spinal cord MRI. METHODS | We used 9.4T MRI of multiple sclerosis spinal cords (>104,000 slices) with sparse annotations (428 slices). A 2D Teacher trained on sparse slices generated dense pseudo-labels to train a 3D Student. We systematically evaluated the impact of human-centric preprocessing, spatial augmentation, and soft-label regularization on both architectures. RESULTS | We identified a critical divergence in training dynamics. The 2D Teacher required strong spatial augmentation and soft-labeling to overcome data scarcity, improving White Matter Lesion Dice scores by >11 points. However, propagating these techniques to the 3D Student degraded its performance. Furthermore, human-centric preprocessing (e.g., CLAHE) disrupted global statistical cues, dropping Gray Matter Lesion Dice scores by ~25 points. DISCUSSION | Our study highlights a perception divergence (human-centric contrast enhancement harms machine models) and a regularization conflict across dimensions. 3D architectures trained on dense pseudo-labels exhibit fundamentally different optimization landscapes than 2D counterparts and require distinct, conservative regularization. Code and models: https://github.com/ivadomed/model_seg_sc-gm-lesion_human_ms_exvivo_t2star.

Reinforcement learning is structurally harder than supervised learning because the policy changes the data distribution it learns from. The resulting fragility is especially visible in large-model training, where the training and rollout systems differ in numerical precision, sampling, and other implementation details. Existing methods manage this fragility by adding hyper-parameters to the training objective, which makes the algorithm more sensitive to its configuration and requires retuning whenever the task, model scale, or distribution mismatch changes. This fragility traces to two concerns that current objectives entangle through hyper-parameters set before training begins: a trust-region concern, that updates should not move the policy too far from its current value, and an off-policy concern, that data from older or different behavior policies should influence the update only to the extent that it remains reliable. Neither concern is a constant to set in advance, and their severity is reflected in the policy-ratio distribution of the current batch. We present a simple yet effective batch-adaptive objective that replaces fixed clipping with the normalized effective sample size of the policy ratios. The same statistic caps the score-function weight and sets the strength of an off-policy regularizer, so the update stays close to the usual on-policy score-function update when ratios are nearly uniform, and tightens automatically when stale or mismatched data cause ratio concentration, while retaining a nonzero learning signal on high-ratio tokens. Experiments across a wide range of settings show that our method matches or exceeds tuned baselines, introducing no new objective hyper-parameters and removing several existing ones. The code is available at https://github.com/FeynRL-project/FeynRL.

We study the fundamental and timely problem of learning long sequences in autoregressive modeling and next-token prediction under model misspecification, measured by the joint Kullback--Leibler (KL) divergence. Our goal is to characterize how the sequence horizon $H$ affects both approximation and estimation errors in this joint-distribution, sequence-level regime. By establishing matching upper and lower bounds, we provide, to our knowledge, the first complete characterization of long-horizon error behavior under the natural joint KL objective, with improved rates and optimality justification relative to existing work. On the approximation side, we show that joint KL admits a horizon-free approximation factor, in sharp contrast to Hellinger-based analyses that exhibit an $Ω(H)$ dependence for computationally efficient methods; this isolates the choice of divergence as the source of approximation amplification. On the estimation side, we prove a fundamental information-theoretic lower bound of order $Ω(H)$ that holds for both decomposable policy classes and fully shared policies, matching the $\widetilde O(H)$ upper bounds achieved by computationally efficient algorithms. Our analysis clarifies the landscape of recent autoregressive learning results by aligning the log-loss training objective, the sequence-level evaluation metric, and the approximation metric {\color{black}through a sharp joint-KL oracle theory}. We further show that these joint-KL guarantees imply policy learning regret bounds at rates matching prior imitation learning literature.

Catastrophic forgetting remains the central obstacle in continual learning (CL): parameters shared across tasks interfere with one another, and existing regularization methods such as EWC and SI apply uniform penalties without awareness of which input region a parameter serves. We propose KAN-CL, a continual learning framework that exploits the compact-support spline parameterization of Kolmogorov-Arnold Networks (KANs) to perform importance-weighted anchoring at per-knot granularity. Deployed as a classification head on a convolutional backbone with standard EWC regularization on the backbone (bbEWC) KAN-CL achieves forgetting reductions of 88% and 93% over a head-only KAN baseline on Split-CIFAR-10/5T and Split-CIFAR-100/10T respectively, while matching or exceeding the accuracy of all baselines on both benchmarks. We further provide a Neural Tangent Kernel (NTK) analysis showing that KAN's spline locality induces a structural rank deficit in the cross-task NTK, yielding a forgetting bound that holds even in the feature-learning regime. These results establish that combining an architecture with natural parameter locality (KAN head) with a complementary backbone regularizer (bbEWC) yields a compositional and principled approach to catastrophic forgetting.

When people share the same documents and observations yet reach different conclusions, the disagreement often shifts into a judgment that the other party is cognitively defective, irrational, or acting in bad faith. This paper argues that such divergence is better described as a form of non-identifiability inherent in inference and learning, rather than as a defect of the other party. We organize the phenomenon into two levels: (i) $θ$-level non-identifiability, where conclusions diverge under the same world model $W$ because inference settings differ; and (ii) $W$-level non-identifiability, where repeated use of an inference setting $θ$ biases data exposure and update rules, causing the learned world model $W$ itself to diverge. We introduce an inference profile $θ= (R, E, S, D)$, consisting of Reference, Exploration, Stabilization, and Horizon, and show how outputs can split even for the same observation $o$ and the same $W$. We further explain why disagreements tend to project onto a small number of bases -- abstract versus concrete, externalizability, and order versus freedom -- as a consequence of general constraints on learning systems: computational, observational, and coordination constraints. Finally, we relate the framework to deep representation learning, including representation hierarchy, latent-state estimation, and regularization-exploration trade-offs, and illustrate the framework through a case study on AI regulation debates.

Artificial intelligence systems typically rely on large, centrally collected datasets, a premise that does not hold in many real-world domains such as healthcare and public institutions. In these settings, data sharing is often constrained by storage, privacy, or resource limitations. For example, small wearable devices may lack the bandwidth or energy capacity needed to store and transmit high-resolution data, leading to aggregation during data collection and thus a loss of information. As a result, datasets collected from different sources may consist of a mixture of high- and low-resolution samples. Despite the prevalence of this setting, it remains unclear how informative low-resolution data is when models are ultimately evaluated on high-resolution inputs. We provide a theoretical analysis based on the Kullback-Leibler divergence that characterises how the influence of a datapoint changes with resolution, and derive bounds that relate the relative contribution of high- and low-resolution observations to the information lost under downsampling. To support this analysis, we empirically demonstrate, using both a vision transformer and a convolutional neural network, that adding low-resolution data to the training set consistently improves performance when high-resolution data is scarce.

Dropout-based sparse-view 3D Gaussian Splatting (3DGS) methods alleviate overfitting by randomly suppressing Gaussian primitives during training. Existing methods mainly focus on designing increasingly sophisticated dropout strategies, while they overlook the resulting inconsistencies among different dropped Gaussian subsets. This oversight often leads to unstable reconstruction and suboptimal Gaussian representation learning.In this paper, we revisit dropout-based sparse-view 3DGS from a consistency regularization perspective and propose PairDropGS, a Paired Dropout-induced Consistency Regularization framework for sparse-view Gaussian splatting. Specifically, PairDropGS first constructs a pair of the dropped Gaussian subsets from a shared Gaussian field and designs a low-frequency consistency regularization to constrain their low-frequency rendered structures. This design encourages the shared Gaussian field to preserve stable scene layout and coarse geometry under different random dropouts, while avoiding excessive constraints on ambiguous high-frequency details. Moreover, we introduce a progressive consistency scheduling strategy to gradually strengthen the consistency regularization during training for stability and robustness of reconstruction. Extensive experiments on widely-used sparse-view benchmarks demonstrate that PairDropGS achieves superior training stability, significantly outperforms existing dropout-based 3DGS methods in reconstruction quality, while exhibiting the simplicity and plug-and-play nature for improving dropout-based optimization.

Self-supervised learning (SSL) is recognized as an essential tool for building foundation models for Artificial Intelligence applications. The advances in SSL have been made thanks to vigorous arguments about the principles of SSL and through extensive empirical research. The aim of this paper is to contribute to the development of the underpinning theory of SSL, focusing on the deep clustering approach. By analogy to supervised learning, we formulate SSL as K-L divergence optimization. The mode collapse is prevented by imposing an optimisation constraint on the teacher distribution. This leads to normalization using inverse cluster priors. We show that using Jensen inequality this normalization simplifies to the popular batch centering procedure. Distillation and centering are common {heuristics-based} practices in SSL, {but our work underpins them theoretically.} The theoretical model developed not only supports specific existing successful SSL methods, but also suggests directions for future investigations.

Multimodal Large Language Models (MLLMs) have shown strong performance in multi-image cross-modal retrieval, yet suffer from severe position bias, where predictions are dominated by input order rather than semantic relevance. Through empirical analysis, we identify a phenomenon termed Logit-Attention Divergence, in which output logits are heavily biased while internal attention maps remain well-aligned with relevant visual evidence. This observation reveals a fundamental limitation of existing logit-level calibration methods such as PriDe. Based on this insight, we propose a training-free, attention-guided debiasing framework that leverages intrinsic attention signals for instance-level correction at inference time, requiring only a minimal calibration set with negligible computational overhead. Experiments on MS-COCO-based benchmarks show that our method substantially improves permutation invariance and achieves state-of-the-art performance, enhancing accuracy by over 40% compared to baselines. Code is available at https://github.com/brightXian/LAD.

Reinforcement learning with verifiable rewards (RLVR) has become an effective paradigm for improving the reasoning ability of large language models. However, widely used RLVR algorithms, such as GRPO, often suffer from entropy collapse, leading to premature determinism and unstable optimization. Existing remedies, including entropy regularization and ratio-based clipping heuristics, either control entropy in a coarse-grained manner or rely on approximate on-policy training. In this paper, we revisit entropy collapse from a token-level entropy flow perspective. Our analysis reveals that entropy-decreasing tokens consistently outweigh entropy-increasing ones, resulting in a severely imbalanced entropy flow. This perspective provides a unified explanation of entropy collapse in existing RLVR algorithms and highlights the importance of balancing entropy dynamics. Motivated by this analysis, we propose On-Policy Entropy Flow Optimization (OPEFO), an adaptive entropy flow balancing mechanism that rescales entropy-increasing and entropy-decreasing updates according to their contributions to entropy change, while remaining strict on-policy. Experiments on six mathematical reasoning benchmarks demonstrate that OPEFO improves training stability and final performance. We will release the code and models upon publication.

Reinforcement Learning with Verifiable Rewards (RLVR) has become the standard paradigm for LLM mathematical reasoning, with Group Relative Policy Optimization (GRPO) serving as the dominant algorithm. We identify two overlooked inefficiencies inherent in GRPO. First, a fixed KL coefficient overly restricts policy exploration at moments when the model needs to diverge significantly from the reference policy. Second, uniform question sampling overlooks that moderately difficult problems produce the most informative gradient signals. We propose FG-ExPO, short for Frontier-Guided Exploration-Prioritized Policy Optimization, which integrates two lightweight components. Accuracy-Conditioned KL Scaling (AKL) adjusts the KL penalty strength through a smooth nonlinear function of batch average accuracy, loosening the constraint when the model performs poorly and strengthening it when the model achieves satisfactory results. Gaussian Curriculum Sampling (GCS) assigns sampling weights to questions following a Gaussian distribution centered at a moderate accuracy level around 0.5, focusing model training on its learning frontier. We conduct evaluations on DeepSeek-R1-Distill-Qwen-1.5B and Qwen3-8B-Base across six mainstream mathematical reasoning benchmarks. Experimental results demonstrate that FG-ExPO consistently outperforms vanilla GRPO. It delivers an absolute improvement of 13.34 on the AIME 2025 pass@32 metric, rising from 63.33 percent to 76.67 percent, and obtains an average pass@32 gain of 2.66 on the 8B model. The substantially larger performance gains observed on pass@32 compared to pass@1 verify that FG-ExPO enlarges the model's effective exploration space under a fixed inference budget.

Inference-time reward alignment asks how to turn a pre-trained diffusion model with base law $p$ into a sampler that favors a reward $r$ while remaining close to $p$. Since there is no canonical distributional distance for this closeness constraint, different choices lead to different "reward-aligned" laws and, just as importantly, different algorithmic problems. We develop a primitive-based approach to reward alignment: rather than assuming arbitrary reward-aligned laws can be sampled, we ask which simple algorithmic primitives suffice to implement alignment for non-trivial reward classes. If closeness is measured in KL distance, the target law is $q(x) \propto p(x) \exp(λ^{-1}r(x))$. For this setting, we show that linear exponential tilts of the form $q(x)\propto p(x)\exp(\langle θ, x \rangle)$ -- which according to recent work [MRR26] can be efficiently sampled from -- are a sufficient primitive for aligning to a very broad class of convex low-dimensional rewards. If closeness is measured in Wasserstein distance, the corresponding primitive is a proximal transport oracle: given $x$, solve $\mbox{argmax}_y \{r(y)- λc(x,y)\}$. This oracle can be efficiently implemented for concave or low-dimensional Lipschitz rewards $r(x)=f(Ax)$. Together, these results illustrate that the choice of distribution distance for alignment affects the computational primitive and the tractable reward class.

On-policy distillation (OPD) and on-policy self-distillation (OPSD) have emerged as promising post-training methods for large language models, offering dense token-level supervision on trajectories sampled from the model's own policy. However, existing results on their effectiveness remain mixed: while OP(S)D has shown promise in system prompt and knowledge internalization, recent studies also report instability and degradation. In this work, we present a comprehensive empirical study of when OPD and OPSD work, when they fail, and why. We find that OPD on mathematical reasoning is highly sensitive to teacher choice and loss formulation, whereas OPSD fails in our tested settings due to test-time absence of instance-specific privileged information (PI). In contrast, OPSD is effective when PI represents a shared latent rule, such as a system prompt or alignment preference. We identify three failure mechanisms: (1) distribution mismatch between teacher and student caused by conditioning on student-generated prefixes, (2) optimization instability from biased TopK reverse-KL gradients, and (3) an OPSD-specific limitation where the student learns a PI-free policy that aggregates PI-conditioned teachers, which is insufficient when PI is instance-specific. We further show that stop-gradient TopK objectives, RLVR-adapted teachers, and SFT-stabilized students mitigate these failures.

This paper proposes a paradigm shift linking machine unlearning directly to the structure of the data distributions rather than a mere update of the neural network parameters. We show that inferring these distributions with precision enables distilling the exact unlearning signal induced by the modeling. Theoretical bounds on the Kullback-Leibler divergence from the ideal retrained model to our unlearned model, under verifiable admissibility criterion, reveal the soundness of our framework. This method is experimentally validated over three forgetting scenarios as reaching the closest classifier to the ideal retrained model when compared to competitors.

We propose a closed-form spectral framework for relative log-density estimation in linearly parameterized probabilistic models, including unnormalized and conditional models. This is achieved by representing the Kullback-Leibler (KL) divergence as an integral of weighted chi-squared divergences, converting KL estimation into a family of least-squares problems. We derive an explicit spectral formula based only on first- and second-order feature moments, yielding closed-form estimators of both divergences and log-density potentials for fixed features. The framework extends to a broad class of f-divergences and can be combined with kernelization or feature learning with neural networks. We prove convergence guarantees for the resulting estimators and empirically compare them on synthetic data with optimization-based variational formulations, including logistic and softmax regression for normalized conditional models.

Trellis-coded quantization sets the current 2-bit post-training frontier for LLMs (QTIP), but pushing below the PTQ ceiling requires quantization-aware training, and QAT on a trellis is obstructed by the non-differentiable Viterbi argmax. We introduce BCJR-QAT, a relaxation that replaces the argmax with the BCJR forward-backward sum-product algorithm at temperature $T$, producing a soft codeword equal to the Boltzmann expectation over trellis paths, exactly differentiable, recovering the hard QTIP code as $T \to 0$, and mathematically identical to the transfer-matrix computation for a 1D Ising-like spin chain. We contribute (i) a fused Triton kernel making BCJR tractable on a single consumer GPU ($6.57\times$ speedup, fp32 parity); (ii) a quantitative drift-budget theory of when BCJR-QAT can escape the QTIP-PTQ Voronoi basin, verified across four experiments; and (iii) a positive empirical result on Llama-3.2-1B at 2 bpw under end-to-end forward-KL distillation: with the right schedule (skip the high-$T$ phase to avoid an overshoot we diagnose), single-layer BCJR-QAT beats QTIP-PTQ by $\mathbf{-0.084}$ PPL on WikiText-2, and multi-layer compounding is super-additive.

Learning a physical model from video data that can comprehend physical laws and predict the future trajectories of objects is a formidable challenge in artificial intelligence. Prior approaches either leverage various Partial Differential Equations (PDEs) as soft constraints in the form of PINN losses, or integrate physics simulators into neural networks; however, they often rely on strong priors or high-quality geometry reconstruction. In this paper, we propose CausalGS, a framework that learns the causal dynamics of complex dynamic 3D scenes solely from multi-view videos, while dispensing with the reliance on explicit priors. At its core is an inverse physics inference module that decouples the complex dynamics problem from the video into the joint inference of two factors: the initial velocity field representing the scene's kinematics, and the intrinsic material properties governing its dynamics. This inferred physical information is then utilized within a differentiable physics simulator to guide the learning process in a physics-regularized manner. Extensive experiments demonstrate that CausalGS surpasses the state-of-the-art on the highly challenging task of long-term future frame extrapolation, while also exhibiting advanced performance in novel view interpolation. Crucially, our work shows that, without any human annotation, the model is able to learn the complex interactions between multiple physical properties and understand the causal relationships driving the scene's dynamic evolution, solely from visual observations.

Modern LLM APIs often reveal only top-$K$ logit scores and censor the remaining vocabulary. We study the per-position distribution-recovery limits of this access model. For censoring threshold $τ$, the compatible teacher distributions form an identified set whose total-variation diameter is exactly $U_K=(V-K)\exp(τ)/(Z_A+(V-K)\exp(τ))$, where $Z_A$ is the observed partition function. For KL recovery, we give a computable binary-endpoint lower bound and an asymptotically matching small-ambiguity upper bound, with an extension to reference-aware attackers. Experiments on a Qwen3 math-reasoning teacher reveal a layered extraction hierarchy: on-task top-$K$ distillation recovers 12% of private capability, full-logit distillation recovers 56% despite 99% KL closure, and generation-based extraction recovers 96%. Top-$K$ censoring therefore limits per-position distribution recovery but does not by itself prevent capability extraction, separating fidelity from transfer in prompt-only logit distillation.

This paper aims at analyzing the regularization effect that data augmentation induces on supervised regression methods in the proportional regime, where the number of covariates grows proportionally to the number of samples. We provide a tight characterization of the test error, measured in mean squared error, in terms only of the population quantities of the true data, as well as first and second order statistics of the augmentation scheme. Our results are valid under misspecified feature maps, and for any network architecture where only the last readout layer is trained, and the rest of the network is either frozen or randomly initialized. We specify our results in the case of Gaussian data, and show that our asymptotic characterization is tight in this setting.

On-policy self-distillation (self-OPD) densifies reinforcement learning with verifiable rewards (RLVR) by letting a policy teach itself under privileged context. We find that when this guidance spans the full response, all-token KL spends gradients on mostly redundant positions and amplifies privileged-information leakage, causing entropy rise, shortened reasoning, and out-of-distribution degradation in long-horizon math training. We propose Token-Routed Alignment for Critical rEasoning (TRACE), which distills only on annotator-marked critical spans: forward KL on key spans of correct rollouts, optional reverse KL on localized error spans, and GRPO on all remaining tokens, with the KL channel annealed away after a short warm-up. Our analysis explains TRACE through two effects: forward KL provides non-vanishing lift to teacher-supported tokens that the student under-allocates, while span masking and decay keep cumulative privileged-gradient exposure finite. On four held-out math benchmarks plus GPQA-Diamond, TRACE improves over GRPO by 2.76 percentage points on average and preserves the Qwen3-8B base OOD score on GPQA-Diamond, where GRPO and all-token self-OPD baselines degrade. Gains persist under online self-annotation (+1.90 percentage points, about 69% of the strong-API gain), reducing the concern that TRACE merely imports external annotator capability. Across scales, the best routed action is base-dependent: on Qwen3-8B it is forward KL on key spans, while on Qwen3-1.7B it shifts to reverse KL on error spans.

Weight decay remains one of the most widely used regularization mechanisms for training convolutional neural networks, yet it is still commonly applied as a fixed coefficient shared by all layers throughout training. This uniform treatment ignores that different layers may follow different structural dynamics and therefore may require different regularization strengths. In this work, we propose OUIDecay, an adaptive layer-wise and time-dependent weight decay scheduler for CNNs driven by the Overfitting-Underfitting Indicator (OUI), an activation-based metric previously shown to provide early information about regularization quality. OUIDecay uses a lightweight batch-based formulation of OUI to monitor the structural behavior of each layer online and periodically rescales its weight decay relative to the other layers in the network. Unlike gradient-based adaptive decay methods, our approach relies on functional information extracted from activation patterns and does not require validation data. Experiments on EfficientNet-B0 with Stanford Cars, ResNet50 with Food101, DenseNet121 with CIFAR100, and MobileNetV2 with CIFAR10 show that OUIDecay achieves the best mean best-validation-loss in 7 out of 8 evaluated settings. These results indicate that activation-driven weight decay adaptation is a practical and effective alternative to fixed decay and gradient-based adaptive decay, while keeping the method lightweight and suitable for online use.

Generative (diffusion) priors demonstrate remarkable performance in addressing inverse problems in imaging. Yet, for scientific and medical imaging, it is crucial that reconstruction techniques remain stable and reliable under imperfect settings. Typical definitions of stability encompass the notion of ''convergent regularization'', robustness to out-of-distribution data, and to inaccuracies in the forward operator or noise model. We evaluate these properties numerically. Furthermore, we benchmark generative approaches against modern optimization-based methods inspired by the widely used variational techniques. Our results give insights for which settings and applications generative priors can deliver state-of-the-art reconstructions, and on those in which they fall short or may even be problematic.

In this paper, we study the problem of sampling from a distribution under the constraint of differential privacy (DP). Prior works measure the utility of DP sampling with density ratio-based measures such as KL divergence. However, such formulations suffer from two key limitations: 1) they fail to capture the geometric structure of the support, and 2) they are not applicable when the supports of the distributions differ. To deal with these issues, we develop a novel framework for DP sampling with Wasserstein distance as the utility measure. In this formulation, we propose Wasserstein Projection Mechanism (WPM), a minimax optimal mechanism based on Wasserstein projection. Furthermore, we develop efficient algorithms for computing the proposed mechanisms approximately and provide convergence guarantees.

Mode connectivity has been widely studied, yet the role of the optimizer remains underexplored. We revisit it through optimizer-induced implicit regularization, asking how connectivity behaves when restricted to solutions constrained by a given optimizer. For two-layer ReLU networks, we show that solutions from a single optimizer -- AdamW, Muon, or others in the Lion-$\mathcal{K}$ family -- form a connected set at sufficiently large width, a result not implied by prior work. We then characterize how optimizer-induced regions interact: at large width two different regions can be disjoint or overlap depending on regularization, while in our small-width example AdamW and Muon converge to disconnected zero-loss components separated by a provable loss barrier. Empirically, in GPT-2 pretraining, we observe same-optimizer paths preserve each model's spectrum while cross-optimizer paths traverse a smooth transition. Our results reveal optimizer-dependent structure beyond classical mode connectivity literature.

Reinforcement Learning with Verifiable Rewards (RLVR) has become the standard paradigm for LLM mathematical reasoning, where Group Relative Policy Optimization (GRPO) serves as the mainstream algorithm. We point out two understudied inefficiencies existing in GRPO. First, the fixed KL penalty coefficient overly restricts policy exploration at stages where the model requires significant deviation from the reference policy. Second, uniform sampling of training questions ignores that moderately difficult problems provide the most informative gradient signals for optimization. We propose Exploration-Prioritized Policy Optimization (EXPO) with two lightweight plug-in modules. The Accuracy-Conditioned KL Scaling (AKL) dynamically adjusts KL regularization strength through a smooth nonlinear function of batch average accuracy, relaxing the penalty when the model underperforms and strengthening it when the model achieves good results. The Gaussian Curriculum Sampling (GCS) assigns sampling weights to questions following a Gaussian distribution centered at moderate accuracy around 0.5, focusing training on the model's learning frontier. We conduct extensive experiments on DeepSeek-R1-Distill-Qwen-1.5B and Qwen3-8B-Base over six mathematical reasoning benchmarks. The results show EXPO steadily surpasses vanilla GRPO. It obtains an absolute gain of 13.34 on AIME 2025 pass@32, rising from 63.33 percent to 76.67 percent, and achieves an average pass@32 improvement of 2.66 on the 8B model. The much larger performance gains on pass@32 compared with pass@1 demonstrate that EXPO effectively enlarges the model's exploration boundary under a fixed inference cost budget.

Diffusion large language models (dLLMs) offer a promising paradigm for parallel text generation, but in practice they face an accuracy-parallelism trade-off, where increasing tokens per forward (TPF) often degrades generation quality. Existing acceleration methods often gain speed at the cost of accuracy. To address this limitation, we propose TAD, a Temporal-Aware trajectory self-Distillation framework. During data construction, we condition a teacher model on both the prompt and the ground-truth response to generate decoding trajectories, recording the intermediate masked states throughout the process. Based on how many decoding steps remain before each masked token is revealed, we partition masked positions into near and distant subsets. For near tokens, we train the student with a hard cross-entropy loss using the teacher trajectory tokens as labels, encouraging confident predictions for tokens that are about to be decoded. For distant tokens, we apply a soft KL divergence loss between the teacher and student token distributions, providing softer supervision and preserving future planning knowledge. This temporal-aware partition naturally gives rise to two deployment configurations: a Quality model that prioritizes accuracy and a Speed model that favors more aggressive acceleration. Experiments show that TAD consistently improves the accuracy-parallelism trade-off. On LLaDA, it raises average accuracy from 46.2% to 51.6% with the Quality model and average AUP from 46.2 to 257.1 with the Speed model. Our code is available at: https://github.com/BHmingyang/TAD

Emojis are widely used in online financial communication, but it is unclear whether they provide transferable sentiment signals across languages, platforms, and asset communities. This study examines the extent to which emoji usage, semantics, and sentiment polarity remain stable across financial communities, and how these layers influence zero-shot sentiment transfer. Using large corpora of Twitter and StockTwits posts in four languages, we measure cross-community divergence and evaluate sentiment models trained under emoji-only, text-only, and text+emoji inputs. We find that emoji frequencies differ across communities, especially across languages, but their semantics and sentiment polarity are largely stable. Cross-asset transferability shows minimal degradation, while cross-language transfer remains the most challenging. Including emojis consistently reduces transfer gaps relative to text-only models. These results indicate that financial communication exhibits a partially shared ``emoji code,'' and that emojis provide compact, language-independent sentiment cues that improve model generalization across markets and platforms.

We consider a mutual information (MI) regularized version of optimal density control of a discrete-time linear system. MI optimal control has been proposed as an extension of maximum entropy optimal control to trade off between control performance and benefits provided by stochastic inputs. MI regularization induces stochasticity in the policy, which poses challenges for applications of MI optimal control in safety-critical scenarios. To remedy this situation, we impose Gaussian density constraints at specified times to directly control state uncertainty. For this MI optimal density control problem, we propose an alternating optimization algorithm and derive the closed form of each step in the algorithm. In addition, we reveal that the alternating optimization of the MI optimal density control problem coincides with that of the so-called generalized Schrödinger bridge problem associated with the discrete-time linear system.

Reinforcement Learning has significantly advanced the reasoning capabilities of Multimodal Large Language Models (MLLMs), yet the resulting policies remain brittle against real-world visual degradations such as blur, compression artifacts, and low-resolution scans. Prior robustness techniques from vision and deep RL rely on static data augmentation or value-based regularization, neither of which transfers cleanly to critic-free RL fine-tuning of autoregressive MLLMs. Reinforcing reasoning against such corruptions is non-trivial: naively injecting degraded views during rollout induces reward poisoning, where perceptual occlusions trigger hallucinated trajectories and destabilize optimization. We propose ROMA, an RL fine-tuning framework that modifies the optimization dynamics to reinforce reasoning against visual degradation while preserving clean-input performance. A dual-forward-pass strategy uses teacher forcing to evaluate corrupted views against clean-image trajectories, avoiding new rollouts on degraded inputs. For distributional consistency, we apply a token-level surrogate KL penalty against the worst-case augmentation; to prevent policy collapse under regularization, an auxiliary policy gradient loss anchored to clean-image advantages preserves a reliable reward signal; and to avoid systematically incorrect invariance, correctness-conditioned regularization restricts enforcement to successful trajectories. On Qwen3-VL 4B/8B across seven multimodal reasoning benchmarks, our method improves robustness by +2.4% on seen and +2.3% on unseen corruptions over GRPO while matching clean accuracy.

If a person can solve a task, can measuring their brain make it easier to train a model to solve that task too? Recent NeuroAI work suggests that supplementing task training with neural recordings can modestly improve model performance and robustness. However, it is unclear when there should be a benefit from using neural data and how much benefit to expect. We formulate this question mathematically, and begin to address it theoretically using a simple, analytically tractable linear gaussian model of task targets and neural recordings. For a multimodal estimator trained on both brain data and task labels, we derive scaling laws for how performance scales with the numbers of brain and task samples. From these laws we derive relative value and exchange rates between brain samples and task samples, quantifying how much extra task samples neural data is worth as a function of task-brain alignment, neural and task noise, latent dimension, and brain data sample size. We also analyze test distribution shift, to identify conditions where brain-regularized learning can produce substantial robustness gains through learned invariances. Finally, under a fixed collection budget, we characterize the regimes in which brain data is worth collecting. Our results provide a foundation for understanding how valuable brain data could be for improving machine learning.

Joint-Embedding Predictive Architectures (JEPAs) provide a simpleframework for learning world models by predicting future latent representations.However, JEPA training is subject to a bias-variance tradeoff.Without sufficient structural constraints, excessive representationalvariance causes the model to collapse to trivial solutions.The recent LeWorldModel (LeWM) shows that this issue can be alleviated bysimply constraining latent embeddings with an isotropic Gaussian prior.However, latent representations inherently lie on low-dimensional manifoldswithin a high-dimensional ambient space, and enforcing an isotropic Gaussianprior directly in this ambient space introduces an overly strong bias.In this work, we propose ame, which seeks a favorable operatingpoint on the bias-variance frontier by applying Gaussian constraints inmultiple random subspaces rather than in the originalembedding space.This design relaxes the global constraint while preserving itsanti-collapse effect, leading to a better balance between trainingstability and representation flexibility.Extensive experiments across fourcontinuous-control environments demonstrate that consistentlyoutperforms LeWM with very clear margins.Our method is simple yet effective, and serves as a strong baseline for future JEPA-based world model research.fdefinedeeemodeThe code is available at https://github.com/intcomp/Sub-JEPA.

\emph{Kullback-Leibler} (KL) regularization is ubiquitous in reinforcement learning algorithms in the form of \emph{reverse} or \emph{forward} KL. Recent studies have demonstrated $ε^{-1}$-type fast rates for decision making under reverse KL regularization, in contrast to the standard $ε^{-2}$-type sample complexity. However, for forward-KL-regularized objectives, existing statistical analyses are either not applicable or result in $\tilde{O}(ε^{-2})$ slow rates. We take the first step towards addressing this problem via a streamlined analysis of forward-KL-regularized offline CBs. We give the first $\tilde{O}(ε^{-1})$ upper bounds in tabular and general function approximation settings, both under notions of \emph{single-policy concentrability}. In particular, our convex-analytical pipeline unifies these settings by exploiting the pessimism principle in a novel way and completely bypasses the proof routines in previous works based on the mean value theorem, which might be of independent interest. Moreover, we provide rate-optimal lower bounds, manifesting the tightness of our upper bounds in terms of statistical rates. Our lower bounds also demonstrate that the forward-KL-regularized sample complexity recovers the unregularized slow rate in the low-regularization regime, similarly to the reverse-KL regularization.

Learned representations are often invariant to rotational transformations, leaving individual dimensions non-identifiable and interchangeable. We study how Matryoshka Representation Learning (MRL) induces a task-aligned privileged basis distinct from variance-based or regularizer-induced orderings. In the linear setting, we prove that full-prefix MRL recovers the ordered principal directions, and can be computed efficiently using shared statistics. Empirically, we demonstrate that MRL yields consistent per-dimension structure aligned with task signal, where coordinate magnitude reflects informativeness.

Bias audits of large language models now operate within governance frameworks such as the EU AI Act, making benchmark reliability a security concern in its own right. Many current benchmarks, however, collapse bias into a single scalar from one prompt format and one surface label. This design misses two failure modes that can be exploited without changing model weights. Across prompts, meaning-preserving format changes shift bias endorsement by more than $0.7$ on a fixed statement pool. Within a response, the discrete Selection and free-text Elaboration can take opposing stances, so an apparently clean aggregate may hide substantial internal inconsistency (a ``cancellation trap''). Selection-only and elaboration-only rankings are therefore nearly uncorrelated across eight LLMs (Spearman $ρ= 0.238$, $p = 0.570$): LLaMA3-70B ranks in the middle under selection-only scoring but highest under elaboration-only scoring on the same responses. We introduce \textsc{BiAxisAudit}, a protocol that reports each bias score together with a reliability estimate on two orthogonal axes. The across-prompt axis evaluates each statement under a factorial grid of task format, perspective, role, and sentiment, treating bias as a distribution rather than a point estimate. The within-response axis uses Split Coding to recover Selection and Elaboration as separate signals, measured by the Inconsistency Rate and Divergence Net Imbalance. Across eight LLMs with $80{,}200$ coded responses each, task format alone explains as much variance as model choice; $63.6\%$ of pooled bias signals (up to $85.2\%$ per model) appear in only one coding layer, and prompt-dimension interactions exceed main effects. The instrument also separates real bias reductions from apparent reductions caused by cross-layer redistribution: some prompt configurations reduce both BER and IR, whereas others suppress only selection-layer bias.

Sardinian, a Romance language with roughly one million speakers, has minimal presence in modern NLP. Commercial services do not support it, and current language models do not produce it reliably. We present LLiMba, a 3B parameter Sardinian-ready model adapted from Qwen2.5-3B-Instruct through continued pretraining (CPT) and supervised fine-tuning (SFT) on a single 24 GB consumer GPU. The corpus contains 11.5 million tokens of Sardinian spanning LSC, Logudorese, and Campidanese, augmented with 2.4 million tokens of related Romance text as replay against register blurring. After CPT the model reaches a perplexity of 6.76 on held out Sardinian and outperforms the base across all six FLORES-200 directions. We compare five SFT configurations under matched conditions: full fine-tuning, LoRA r64, rsLoRA r128, rsLoRA r256, and DoRA r256. rsLoRA r256 wins on every direction into Sardinian, reaching 28.5 BLEU from English against 17.3 after CPT and 21.0 with full fine-tuning. The rank ablation places r128 between LoRA r64 and rsLoRA r256 on BLEU but reveals failure modes invisible to the metric, including leakage across scripts no other variant produces. LoRA r64 retains less factual content from SFT than configurations at higher rank and produces more confident fabrications, though all methods fabricate on content absent from training. DoRA r256 yields the smallest gap between training and evaluation but the worst factual accuracy. The findings indicate that adapter capacity matters more than the choice among LoRA variants for adapting a Romance pretrained base to a low resource Romance target, that stronger regularization is not uniformly beneficial, and that translation metrics smoothly order configurations whose qualitative behavior differs categorically. Perplexity comparisons across scripts must account for byte fallback tokenization, which deflates the metric for scripts other than Latin.

The new wave of adversarial attacks that utilize gradient-related vulnerabilities in neural network-based classifiers makes Network Intrusion Detection Systems more open to such threats. Although state-of-the-art adversarial training methods have shown promising results in producing more robust classifiers, their interpretability and defense ability are limited due to their lack of understanding of how adversarial attacks propagate in different layers of network classifiers. In this paper, we present an insightful approach, called LARAR (Layer-wise Adversarial Robustness using Adaptive Regularization), that incorporates additional layer-wise vulnerability analysis and adaptive weighting in conventional adversarial training methods. Additionally, we utilize 'Auxiliary Classifiers' in our approach. LARAR provides interpretable layer-wise vulnerability scores, achieves a clean accuracy of 95.01%, and provides better robustness against adversarial attacks (FGSM, PGD, and transfer attacks) on the UNSW-NB15 dataset. Through the identification of vulnerable layers, the proposed framework reduces computational complexity and enables the early detection of adversarial samples, thus enhancing the effectiveness and interpretability of adversarial defense mechanisms in NIDS.

Flow Matching has recently emerged as a popular class of generative models for simulating a target distribution $μ_1$ from samples drawn from a source distribution $μ_0$. This framework relies on a fixed coupling between $μ_0$ and $μ_1$, and on a deterministic or stochastic bridge to define an interpolating process between the two distributions. The time marginals of this process can then be approximately sampled by estimating the transition rates, or more generally the generator, of its Markovian projection. This framework has recently been extended to the case of discrete source and target distributions, under the name Discrete Flow Matching (DFM). However, theoretical guarantees for such models remain scarce. In this paper, we study two DFM models on $\mathbb{Z}_m^d = \{0,\ldots,m-1\}^d$, sampled through time discretization, and derive non-asymptotic associated bounds for both of them. In contrast to previous work, we establish non-asymptotic bounds in Kullback--Leibler divergence for the early-stopped version of the target distribution. We also derive explicit convergence guarantees in total variation distance with respect to the true target distribution. Importantly, these bounds rely only on an approximation error assumption, relaxing standard score assumptions used in earlier works, while also yielding improved dependence on the vocabulary size $m$ and the dimension $d$.

Pretraining large language models (LLMs) with next-token prediction has led to remarkable advances, yet the context-dependent nature of token embeddings in such models results in high intra-class variance and inter-class similarity, thus hindering the efficiency of representation learning. While similarity-based regularization has demonstrated benefit in supervised fine-tuning and classification tasks, its application and efficacy in large-scale LLM pretraining remains underexplored. In this work, we propose the SimReg, an embedding similarity regularization loss that explicitly encourages token representations with the same ground-truth label within each sequence to be more similar, while enforcing separation from different-label tokens via a contrastive loss. Our analysis reveals that this mechanism introduces gains by enlarging multi-classification margins, thereby enabling more efficient classification. Extensive experiments across dense and Mixture-of-Experts (MoE) architectures demonstrate that SimReg consistently accelerates training convergence by over 30% and improves average zero-shot downstream performance by over 1% across standard benchmarks. Further ablation studies and analyses offer practical insights into hyperparameter tuning and loss effectiveness.

Transparent decision-making is essential for traffic signal control (TSC) systems to earn public trust. However, traditional reinforcement learning-based TSC methods function as black boxes with limited interpretability. Although large language models (LLMs) can provide natural language reasoning, reinforcement finetuning for TSC remains unstable because feedback is sparse and delayed, while most actions produce only marginal changes in congestion metrics. We introduce OracleTSC, which stabilizes LLM-based TSC through two mechanisms: (1) a reward hurdle mechanism that filters weak learning signals by subtracting a calibrated threshold from environmental rewards, and (2) uncertainty regularization that maximizes the probability of the selected response to encourage consistent decisions across sampled outputs. Experiments on the LibSignal benchmark show that OracleTSC enables a compact LLaMA3-8B model to substantially improve traffic efficiency, achieving a 75% reduction in travel time and a 67% decrease in queue length compared with the pretrained baseline while preserving interpretability through natural language explanations. OracleTSC also demonstrates strong cross-intersection generalization: a policy trained on one intersection transfers to a structurally different intersection with 17% lower travel time and 39% lower queue length without additional finetuning. These results suggest that uncertainty-aware reward shaping can improve the stability and effectiveness of reinforcement fine-tuning for TSC.

We introduce the Sinkhorn treatment effect, an entropic optimal transport measure of divergence between counterfactual distributions. Unlike classical quantities such as the average treatment effect, this measure captures differences across entire distributions. We analyze this divergence as a statistical functional and show it can be written as a smooth transformation of counterfactual mean embeddings with an appropriate kernel. This characterization allows us to establish first-order pathwise differentiability in general, and second-order pathwise differentiability under the null hypothesis of equal counterfactual distributions. Leveraging this smoothness, we construct debiased estimators and use them to obtain asymptotically valid tests for distributional treatment effects with a fixed entropic regularization parameter. Because the power of the test depends on this unknown parameter, we further propose an aggregated test that combines evidence across a grid of regularization choices. Experiments on simulated and image data demonstrate the practical advantages of our estimator and testing procedure.

The manifold hypothesis (MH) is often used to explain how machine learning can overcome the curse of dimensionality. However, the MH is only applicable in regimes where the training data provides a sufficiently dense sample of the underlying low-dimensional data manifold, or where such a low-dimensional manifold is conceivably present. We describe the regimes where the MH is not applicable as sparse. In this paper, we demonstrate that models succeed in the sparse regime by exploiting a highly structured local geometry, a property we formalize as normal alignment. We prove that normal-aligned classifiers -- whose input-output Jacobians are rank-one and align perfectly with the training data -- minimize the training objective under norm constraints and achieve maximal local robustness under a non-zero Jacobian constraint. For continuous piecewise-affine deep networks, normal alignment manifests geometrically as centroid alignment within the network's induced power diagram partition and results from the feature-learning regime. Motivated by these theoretical insights, we introduce GrokAlign, a regularization strategy that actively induces normal alignment. We demonstrate that GrokAlign significantly accelerates the training dynamics of deep networks relevant to the grokking phenomenon. Furthermore, we apply the principle of normal alignment to Recursive Feature Machines (RFMs) to introduce Recursive Feature Alignment Machines (RFAMs). We show that RFAMs exhibit greater adversarial robustness compared to RFMs when trained on tabular data.

Sparse training reduces the memory and computational costs of deep neural networks. However, sparse optimization methods, e.g., those adding an $\ell_1$ penalty, often control sparsity only indirectly through a regularization parameter $λ$, whose mapping to the final sparsity rate is non-trivial. In our experiments, we found this parameter sensitivity to be particularly pronounced for Bregman-based optimizers. Specifically, the two variants LinBreg and AdaBreg reach the same sparsity at $λ$ values that differ by up to two orders of magnitude, requiring expensive trial-and-error sweeps to achieve a user-specified sparsity. To address this, we propose an adaptive regularization scheme that updates $λ$ based on the difference between the model's current sparsity and the target sparsity. We analyze the resulting algorithm and evaluate it on automatic speaker verification with ECAPA-TDNN and ResNet34 on VoxCeleb and CNCeleb. The proposed method reliably achieves sparsity targets ranging between 75% and 99%. It also converges faster than the oracle-tuned non-adaptive baseline during early training and matches or surpasses its final performance in equal error rate. We further show that the adaptive scheme inherits key properties from its non-adaptive counterpart, including improved out-of-distribution robustness over the dense baselines.

On-Policy Distillation (OPD) has emerged as a dominant post-training paradigm for large language models, especially for reasoning domains. However, OPD remains unstable in practice due to the high gradient variance of its single-sample Monte Carlo estimator, and recipes for stable training are still immature. We propose vOPD (On-Policy Distillation with a control variate baseline), which casts OPD as policy-gradient RL and stabilizes it by introducing a control variate baseline-canonically a value function -- from the RL literature. We show that the OPD value function admits a closed form as the per-token negative reverse KL divergence between the student and the teacher, available directly from the already-computed forward pass with no additional critic or inference. Existing stabilization methods either compute the full token-level reverse KL over the entire vocabulary, adding significant overhead, or restrict it to a top-k support, biasing the objective. vOPD instead preserves the lightweight single-sample estimator, subtracting the value function as a detached baseline to keep the gradient unbiased while reducing variance. Furthermore, we show that a top-k approximation of the baseline further lowers cost without compromising performance. Across mathematical and scientific reasoning benchmarks, vOPD consistently outperforms vanilla OPD and matches the most expensive full-vocabulary baseline, offering an efficient stabilization of On-Policy Distillation through principled RL variance reduction.

We study the problem of \emph{architecture selection} for deep learning models trained to solve partial differential equations (PDEs), asking when transformer-based architectures with learned attention outperform Fourier-domain neural operators. We introduce the \textbf{Multi-Scale Attention Transformer} (\msat{}), a deep learning architecture that encodes spatiotemporal solution histories as token sequences and trains end-to-end via a composite supervised objective with optional physics-informed regularization terms. We conduct a comprehensive empirical evaluation against nine baselines -- including physics-informed neural networks (PINNs), neural operators (FNO, DeepONet, GNOT), and state-space models (Mamba-NO) -- across five benchmark problems from the PINNacle suite, using identical train/test splits and reference data for all methods. \msat{} achieves state-of-the-art generalization on complex geometry problems ($L^2_\mathrm{rel} = 0.0101$ on Heat2D-CG, a $3.7\times$ improvement over FNO) at $34\,\mathrm{s}$ total inference vs.\ $120{,}812\,\mathrm{s}$ for Mamba-NO. Ablation studies over the physics regularization component reveal a precise inductive bias tradeoff: physics priors reduce test error on diffusion-dominated problems but degrade generalization on chaotic and recirculating-flow regimes, directly characterizing the prior misspecification boundary. Approximation error bounds as a function of domain boundary complexity $κ$ provide a theoretical basis for these empirical findings and a principled rule for architecture selection.

Despite extensive progress in point cloud robustness, existing methods primarily rely on augmentation strategies or defense mechanisms while overlooking the geometric nature of adversarial fragility. We hypothesize that adversarial vulnerability in 3D networks arises from a manifold misalignment between the latent geometry learned by the model and the intrinsic geometry of the underlying surface. Small, geometry-preserving perturbations along the input manifold often induce disproportionate distortions in feature space, potentially leading to misclassifications. We formalize this phenomenon by developing a geometric interpretation of 3D robustness that links classical adversarial theory to the intrinsic structure of point clouds. Motivated by this analysis, we introduce Manifold-Aligned Point Recognition (MAPR), a framework that regularizes the latent geometry by aligning predictions across intrinsic perturbations. MAPR augments each point cloud with intrinsic features capturing local curvature and diffusion structure, and applies a consistency loss that preserves invariance to intrinsic, geometry-preserving perturbations. Without relying on adversarial training or additional data, MAPR consistently improves robustness under multiple adversarial attacks across several datasets, achieving average robustness gains of +20.02 and +8.83 percentage points over vanilla models on ModelNet40 and ScanObjectNN, respectively.

Reinforcement learning with verifiable rewards improves LLM reasoning but often induces overthinking, where models generate unnecessarily long reasoning traces. Existing methods mainly rely on length penalties or early-exit strategies; however, the former may degrade accuracy and induce underthinking, whereas the latter assumes that substantial portions of reasoning traces can be safely truncated. To obtain a compression signal without these limitations, we revisit the training dynamics of existing compression methods. We observe that the length--accuracy correlation is initially negative but continually increases during compression, indicating that shorter responses are initially more likely to be correct but gradually lose this property as the policy moves toward underthinking. Based on this observation, we formalize overthinking: a negative correlation indicates an overthinking regime, while a positive one indicates underthinking. When overthinking, the shortest correct responses are shorter than the group-average response length in expectation, making them natural compression targets already present in on-policy rollouts. We therefore propose \emph{Implicit Compression Regularization} (ICR), an on-policy regularization method whose compression signal comes from a virtual shorter distribution induced by the shortest correct responses in rollout groups, guiding the policy toward concise yet correct trajectories. Training dynamics show that ICR maintains a better length--accuracy correlation during compression, indicating that short responses remain better aligned with correctness instead of drifting toward underthinking. Experiments on three reasoning backbones and multiple mathematical and knowledge-intensive benchmarks show that ICR consistently shortens responses while preserving or improving accuracy, achieving a stronger accuracy--length Pareto frontier.

Language model (LM) alignment improves model outputs to reflect human preferences while preserving the capabilities of the base model. The most common alignment approaches are (i) reinforcement learning, which maximizes the expected reward under a KL-divergence constraint, and (ii) best-of-$N$ alignment, which selects the highest-reward output among $N$ independent samples. Despite their widespread use, the fundamental limits of reward improvement under a KL budget remain poorly understood. We characterize the information-theoretic limits of KL-regularized alignment by deriving the maximum achievable expected reward gain for a fixed KL-divergence budget. Our first result provides a closed-form expression for the optimal reward improvement, governed by a Jeffreys divergence term rather than the $\sqrt{\texttt{KL}}$ used in prior analyses. We further reformulate this expression as a covariance under the base model, yielding a practical estimator that predicts achievable alignment gains from base model samples alone. We extend our analysis to the proxy reward setting, showing that the gap between ideal and proxy alignment (reward hacking) grows with the magnitude of reward error and when the KL penalty factor decreases. We then prove that reward ensembling mitigates reward hacking, providing a theoretical justification for this technique used in practice. Empirically, we compute the KL-reward Pareto frontier for two tasks for LMs, safety and summarization, and show that best-of-$N$ closely approaches the theoretical limit, while PPO and GRPO remain substantially suboptimal. Our theoretical results shed light on several empirically observed phenomena in the alignment literature and suggest that algorithmic improvements are needed to achieve optimal alignment without high inference costs.

Data selection methods address a critical challenge in LLM post-training: effectively leveraging scarce, high-fidelity target data alongside abundant but imperfectly aligned general training data. In this work, we move beyond the data-selection framing and introduce Dr. Post-Training (Data-Regularized Post-Training), a novel framework that reconceptualizes general training data as a data-induced regularizer that prevents overfitting to the scarce target objective, rather than serving as a pool for selection. Specifically, our framework proposes that at each training step, construct a feasible set of model update directions using the general training data, and project the model update direction specified by the scarce target data onto that feasible set. Standard training and existing data selection methods arise as special cases with different choices of the data-induced regularizer, and these methods correspond to different points on a bias--variance spectrum with different regularization strength. Building on this view, we propose a family of methods offering a richer design space and more flexible bias--variance tradeoffs. For practical LLM-scale use, we introduce careful system optimizations that realize these methods with minimal overhead. Extensive experiments across SFT, RLHF, and RLVR show that our methods consistently outperform state-of-the-art data selection baselines, and system benchmarks confirm their efficiency.

Reinforcement Learning from Human Feedback (RLHF) has become a cornerstone technique for post-training large language models. While most existing approaches rely on the reverse KL-regularization, recent empirical studies have begun exploring alternative divergences (e.g., forward KL, chi-squared) as regularizers in RLHF. However, a unified theoretical understanding of general $f$-divergence regularization remains under-explored. To fill this gap, this work develops a comprehensive theoretical framework for online RLHF with a general $f$-divergence regularized objective. Rather than treating each possible divergence function individually, we adopt a holistic perspective across the entire function class and propose two algorithms based on distinct sampling principles. The first extends the classical optimism principle with a carefully designed exploration bonus, while the second introduces a new method that exploits the sensitivity of the optimal policy to reward perturbations under $f$-divergence regularization. Theoretical analysis shows that $O(\log T)$ regret and $O(1/T)$ sub-optimality gap are achievable, establishing provable efficiency of both algorithms and, to the best of our knowledge, the first performance bounds for online RLHF under general $f$-divergence regularization.

Learning rate is a critical component of reinforcement learning (RL). This work uses global and local clocks to distinguish two types of learning rates. The former is of the standard form $α_t$ that depends only on the time step $t$ (i.e., a global clock). The latter is of the form $α_{ν(S_t, t)}$, where $ν(s, t)$ counts the number of visits to state $s$ until time $t$ (i.e., a local clock). In discounted RL, an RL algorithm that is convergent with a local clock is always also convergent with a global clock, and vice versa. We are not aware of any counterexample. The key contribution of this work is to show that this nice correspondence breaks down in average-reward RL. Specifically, we construct a counterexample showing that although differential temporal difference learning is convergent with a local clock, it can diverge with a global clock. This counterexample closes the open problem in Wan et al. [2021], Blaser et al. [2026].

Optimizers play an important role in both pretraining and finetuning stages when training large language models (LLMs). In this paper, we present an observation that full finetuning with the same optimizer as in pretraining achieves a better learning-forgetting tradeoff, i.e., forgetting less while achieving the same or better performance on the new task, than other optimizers and, possibly surprisingly, LoRA, during the supervised finetuning (SFT) stage. We term this phenomenon optimizer-model consistency. To better understand it, through controlled experiments and theoretical analysis, we show that: 1) optimizers can shape the models by having regularization effects on the activations, leading to different landscapes around the pretrained checkpoints; 2) in response to this regularization effect, the weight update in SFT should follow some specific structures to lower forgetting of the knowledge learned in pretraining, which can be obtained by using the same optimizer. Moreover, we specifically compare Muon and AdamW when they are employed throughout the pretraining and SFT stages and find that Muon performs worse when finetuned for reasoning tasks. With a synthetic language modeling experiment, we demonstrate that this can come from Muon's strong tendency towards rote memorization, which may hurt pattern acquisition with a small amount of data, as for SFT.

Reinforcement learning with verifiable rewards (RLVR), due to the deterministic verification, becomes a dominant paradigm for enhancing the reasoning ability of large language models (LLMs). The community witnesses the rapid change from the Proximal Policy Optimization (PPO) to Group Relative Policy Optimization (GRPO), in which GRPO reduces the complicated advantage estimation with simple estimation over grouped positive and negative rollouts. However, we note that negative rollouts may admit no gradation of failure severity, and the combinatorial vastness makes penalizing a few sampled negatives unlikely to cover a meaningful reward signal under sparse binary rewards. In this work, we propose Positive-Only Policy Optimization (POPO), a novel RLVR framework in which learning can occur exclusively via online positive rollouts. Specifically, POPO utilizes bounded importance sampling over the positive rollout set. Thus, no disjoint negative rollouts are used for the gradient guidance. We show that implicit negative gradients can emerge naturally through reinforcing the positive probability via rollouts redistribution. Next, POPO stabilizes the policy optimization through two mechanisms. First, it applies a siamese policy network with a momentum-based adaptation law for stabilized policy evolution. Second, we replace the KL-divergence with a bounded similarity penalty term in the siamese representation space. We conduct extensive experiments using publicly available, well-established text-LLM models, e.g., the Qwen family, across all-level mathematical benchmarks. Our experiment demonstrates that POPO achieves performance comparable to, or even superior to GRPO. Notably, we show that POPO can achieve 36.67% in AIME 2025 with Qwen-Math-7B, outperforming GRPO 30.00%. Our ablation and sweep studies further illustrate the necessity and robustness of POPO components.

Weight decay is widely used as a regularizer in large language models, yet its precise role in shaping Transformer loss landscapes remains theoretically underexplored. This paper provides the first rigorous functional-analytic characterization of the standard Transformer objective--cross-entropy loss with $L^2$ regularization--by proving it satisfies Villani's criteria for coercive energy functions. Specifically, we show that the regularized loss $\mathcal{F}$ is infinitely differentiable, grows at least quadratically, has Gaussian-integrable tails, and satisfies the differential growth condition $-Δ\mathcal{F} + \tfrac{1}{s}\|\nabla\mathcal{F}\|^{2} \to \infty$ as $\|θ\| \to \infty$ for all $s>0$. From this structure, we derive explicit log-Sobolev and Poincaré constants $C_{\mathrm{LS}} \leq λ^{-1} + d/λ^{2}$, linking the regularization strength $λ$ and model dimension $d$ to finite-time convergence guarantees for noisy stochastic gradient descent and PAC-Bayesian generalization bounds that tighten with increasing $λ$. To validate our theory, we introduce a scalable Villani diagnostic $Ψ_s(θ) = -Δ\mathcal{F} + s^{-1}\|\nabla \mathcal{F}\|^2$ and estimate it efficiently using Hutchinson trace probes in models with over 100M parameters. Experiments on GPT-Neo-125M across Penn Treebank and WikiText-103 confirm the predicted quadratic growth of $Ψ_s$, spectral inflation of the Hessian, and exponential convergence behavior consistent with our log-Sobolev analysis. These results demonstrate that weight decay not only improves generalization empirically but also establishes the mathematical conditions required for fast Langevin mixing and theoretically grounded curvature-aware optimization in deep learning.

We present EDDY (Exact-marginal Diversification via Divergence-free dYnamics), a guidance mechanism for diffusion and flow matching models that promotes diversity among samples generated while maintaining quality. EDDY exploits symmetries of the Fokker-Planck equation, using drift perturbations that change particle trajectories while preserving the evolving marginal distribution. We instantiate this principle through kernel-based anti-symmetric pairwise matrix fields, constructed from the repulsive directions. The resulting divergence-free dynamics promote diversity at the joint particle level while preserving each particle's marginal distribution without any additional training. As computing the guidance can be computationally expensive in cases such as text-to-image generation with perceptual embeddings, we propose practical approximations as an effective and efficient solution. Experiments on synthetic distributions and text-to-image generation show that EDDY improves diversity while maintaining strong distributional fidelity compared to common baselines.

Despite extensive research into mitigating distribution shifts, many existing algorithms yield inconsistent performance, often failing to outperform baseline Empirical Risk Minimization (ERM) across diverse scenarios and necessitating newer algorithms which can handle scenarios where existing algorithms currently underperform. Furthermore, high algorithmic complexity frequently limits interpretability and offers only an indirect means of addressing spurious correlations. We propose eXplaining to Learn (eX2L): an interpretable, explanation-based framework that decorrelates confounding features from a classifier's latent representations during training. eX2L achieves this by penalizing the similarity between Grad-CAM activation maps generated by a primary label classifier and those from a concurrently trained confounder classifier. On the rigorous Spawrious Many-to-Many Hard Challenge synthetic data benchmark, eX2L achieves an average accuracy (AA) of 82.24% +/- 3.87% and a worst-group accuracy (WGA) of 66.31% +/- 8.73%, outperforming the current state-of-the-art (SOTA) by 5.49% and 10.90%, respectively. Beyond its competitive performance, eX2L demonstrates that functional domain invariance can be enforced by explicitly decoupling label and nuisance attributes at the group level.