Papers

We decompose the spectral edge – the dominant direction of the Gram matrix of parameter updates – into its gradient and weight-decay components during grokking in two sequence tasks (Dyck-1 and SCAN). We find a sharp two-phase lifecycle: before grokking the edge is gradient-driven and functionally active; at grokking, gradient and weight decay align, and the edge becomes a compression axis that is perturbation-flat yet ablation-critical (>4000x more impactful than random directions). Three universality classes emerge (functional, mixed, compression), predicted by the gap flow equation. Nonlinear probes show information is re-encoded, not lost (MLP $R^2=0.99$ where linear $R^2=0.86$), and removing weight decay post-grok reverses compression while preserving the algorithm.

From butterfly wings to remnants of nuclear detonation, aperiodic order repeatedly emerges in nature, often exhibiting reduced sensitivity to boundaries and symmetry constraints. Inspired by this principle, a paradigm shift is introduced in nanomechanical resonator design from periodic to aperiodic structures, focusing on a special class: quasicrystals (QCs). Although soft clamping enabled by phononic stopbands has become a central strategy for achieving high-$Q_m$ nanomechanical resonators, its practical realization has been largely confined to periodic phononic crystals, where band structure engineering is well established. The potential of aperiodic architectures, however, has remained largely unexplored, owing to their intrinsic complexity and the lack of systematic approaches to identifying and exploiting stopband behavior. Here we demonstrate that soft clamping can be realized in quasicrystal architectures and that high-$Q_m$ nanomechanical resonators can be systematically predicted through a data-driven design framework. As a representative demonstration, the 12-fold QC-based resonator obtained numerically a quality factor $Q_m \sim 10^7$ and an effective mass of sub-nanograms at MHz frequencies, corresponding to an exceptional force sensitivity of $26.4$~aN/$\sqrt{\text{Hz}}$ compared to previous 2D phononic crystals. These results establish QCs as a robust platform for next-generation nanomechanical resonators and open a new design regime beyond periodic order.

Large Chunk Test-Time Training (LaCT) has shown strong performance on long-context 3D reconstruction, but its fully plastic inference-time updates remain vulnerable to catastrophic forgetting and overfitting. As a result, LaCT is typically instantiated with a single large chunk spanning the full input sequence, falling short of the broader goal of handling arbitrarily long sequences in a single pass. We propose Elastic Test-Time Training inspired by elastic weight consolidation, that stabilizes LaCT fast-weight updates with a Fisher-weighted elastic prior around a maintained anchor state. The anchor evolves as an exponential moving average of past fast weights to balance stability and plasticity. Based on this updated architecture, we introduce Fast Spatial Memory (FSM), an efficient and scalable model for 4D reconstruction that learns spatiotemporal representations from long observation sequences and renders novel view-time combinations. We pre-trained FSM on large-scale curated 3D/4D data to capture the dynamics and semantics of complex spatial environments. Extensive experiments show that FSM supports fast adaptation over long sequences and delivers high-quality 3D/4D reconstruction with smaller chunks and mitigating the camera-interpolation shortcut. Overall, we hope to advance LaCT beyond the bounded single-chunk setting toward robust multi-chunk adaptation, a necessary step for generalization to genuinely longer sequences, while substantially alleviating the activation-memory bottleneck.

Generating motion-controlled videos–where user-specified actions drive physically plausible scene dynamics under freely chosen viewpoints–demands two capabilities: (1) disentangled motion control, allowing users to separately control the object motion and adjust camera viewpoint; and (2) motion causality, ensuring that user-driven actions trigger coherent reactions from other objects rather than merely displacing pixels. Existing methods fall short on both fronts: they entangle camera and object motion into a single tracking signal and treat motion as kinematic displacement without modeling causal relationships between object motion. We introduce MoRight, a unified framework that addresses both limitations through disentangled motion modeling. Object motion is specified in a canonical static-view and transferred to an arbitrary target camera viewpoint via temporal cross-view attention, enabling disentangled camera and object control. We further decompose motion into active (user-driven) and passive (consequence) components, training the model to learn motion causality from data. At inference, users can either supply active motion and MoRight predicts consequences (forward reasoning), or specify desired passive outcomes and MoRight recovers plausible driving actions (inverse reasoning), all while freely adjusting the camera viewpoint. Experiments on three benchmarks demonstrate state-of-the-art performance in generation quality, motion controllability, and interaction awareness.

The rapid growth of generative artificial intelligence (AI) has introduced unprecedented computational demands, driving significant increases in the energy footprint of data centers. However, existing power consumption data is largely proprietary and reported at varying resolutions, creating challenges for estimating whole-facility energy use and planning infrastructure. In this work, we present a methodology that bridges this gap by linking high-resolution workload power measurements to whole-facility energy demand. Using NLR's high-performance computing data center equipped with NVIDIA H100 GPUs, we measure power consumption of AI workloads at 0.1-second resolution for AI training, fine-tuning and inference jobs. Workloads are characterized using MLCommons benchmarks for model training and fine-tuning, and vLLM benchmarks for inference, enabling reproducible and standardized workload profiling. The dataset of power consumption profiles is made publicly available. These power profiles are then scaled to the whole-facility-level using a bottom-up, event-driven, data center energy model. The resulting whole-facility energy profiles capture realistic temporal fluctuations driven by AI workloads and user-behavior, and can be used to inform infrastructure planning for grid connection, on-site energy generation, and distributed microgrids.

Pluralistic alignment has emerged as a critical frontier in the development of Large Language Models (LLMs), with reward models (RMs) serving as a central mechanism for capturing diverse human values. While benchmarks for general response quality are prevalent, evaluating how well reward models account for individual user preferences remains an open challenge. To bridge this gap, we introduce Personalized RewardBench, a novel benchmark designed to rigorously assess reward models' capacity to model personalized preferences. We construct chosen and rejected response pairs based on strict adherence to (or violation of) user-specific rubrics, ensuring that preference distinctions are uniquely tailored to the individual. In particular, human evaluations confirm that the primary discriminative factor between pairs is strictly personal preference, with both responses maintaining high general quality (e.g., correctness, relevance and helpfulness). Extensive testing reveals that existing state-of-the-art reward models struggle significantly with personalization, peaking at an accuracy of just 75.94%. Crucially, because an effective reward model benchmark should predict a reward model's performance on downstream tasks, we conduct experiments demonstrating that our benchmark exhibits a significantly higher correlation with downstream performance in both Best-of-N (BoN) sampling and Proximal Policy Optimization (PPO) compared to existing baselines. These findings establish Personalized RewardBench as a robust and accurate proxy for evaluating reward models' performance in downstream applications.

Most repository-level code translation and validation techniques have been evaluated on a single source-target programming language (PL) pair, owing to the complex engineering effort required to adapt new PL pairs. Programming agents can enable PL-agnosticism in repository-level code translation and validation: they can synthesize code across many PLs and autonomously use existing tools specific to each PL's analysis. However, state-of-the-art has yet to offer a fully autonomous agentic approach for repository-level code translation and validation of large-scale programs. This paper proposes ReCodeAgent, an autonomous multi-agent approach for language-agnostic repository-level code translation and validation. Users only need to provide the project in the source PL and specify the target PL for ReCodeAgent to automatically translate and validate the entire repository. ReCodeAgent is the first technique to achieve high translation success rates across many PLs. We compare the effectiveness of ReCodeAgent with four alternative neuro-symbolic and agentic approaches to translate 118 real-world projects, with 1,975 LoC and 43 translation units for each project, on average. The projects cover 6 PLs and 4 PL pairs. Our results demonstrate that ReCodeAgent consistently outperforms prior techniques on translation correctness, improving test pass rate by 60.8% on ground-truth tests, with an average cost of $15.3. We also perform process-centric analysis of ReCodeAgent trajectories to confirm its procedural efficiency. Finally, we investigate how the design choices (a multi-agent vs. single-agent architecture) influence ReCodeAgent performance: on average, the test pass rate drops by 40.4%, and trajectories become 28% longer and persistently inefficient.

How does the choice of training data influence an AI model? This broad question is of central importance to interpretability, privacy, and basic science. At its technical core is the data deletion problem: after a reasonable amount of precomputation, quickly predict how the model would behave in a given situation if a given subset of training data had been excluded from the learning algorithm. We present a data deletion scheme capable of predicting model outputs with vanishing error $\varepsilon$ and failure probability $\delta$ in the deep learning setting. Our precomputation and prediction algorithms are only $\tilde{O}(\log(1/\delta)/\varepsilon^2)$ factors slower than regular training and inference, respectively. The storage requirements are those of $\tilde{O}(\log(1/\delta)/\varepsilon^2)$ models. Our proof is based on an assumption that we call stability. In contrast to the assumptions made by prior work, stability appears to be fully compatible with learning powerful AI models. In support of this, we show that stability is satisfied in a minimal set of experiments with microgpt. Our code is available at https://github.com/SamSpo1/microgpt-sketch. At a technical level, our work is based on a new method for locally sketching an arithmetic circuit by computing higher-order derivatives in random complex directions. Forward-mode automatic differentiation allows cheap computation of these derivatives.

In this paper, we derive rates of convergence in the high-dimensional central limit theorem for Polyak-Ruppert averaged iterates generated by the asynchronous Q-learning algorithm with a polynomial stepsize $k^{-\omega},\, \omega \in (1/2, 1]$. Assuming that the sequence of state-action-next-state triples $(s_k, a_k, s_{k+1})_{k \geq 0}$ forms a uniformly geometrically ergodic Markov chain, we establish a rate of order up to $n^{-1/6} \log^{4} (nS A)$ over the class of hyper-rectangles, where $n$ is the number of samples used by the algorithm and $S$ and $A$ denote the numbers of states and actions, respectively. To obtain this result, we prove a high-dimensional central limit theorem for sums of martingale differences, which may be of independent interest. Finally, we present bounds for high-order moments for the algorithm's last iterate.

The growing complexity of neural networks hinders the deployment of distributed machine learning on resource-constrained devices. Split learning (SL) offers a promising solution by partitioning the large model and offloading the primary training workload from edge devices to an edge server. However, the increasing number of participating devices and model complexity leads to significant communication overhead from the transmission of smashed data (e.g., activations and gradients), which constitutes a critical bottleneck for SL. To tackle this challenge, we propose SL-FAC, a communication-efficient SL framework comprising two key components: adaptive frequency decomposition (AFD) and frequency-based quantization compression (FQC). AFD first transforms the smashed data into the frequency domain and decomposes it into spectral components with distinct information. FQC then applies customized quantization bit widths to each component based on its spectral energy distribution. This collaborative approach enables SL-FAC to achieve significant communication reduction while strategically preserving the information most crucial for model convergence. Extensive experiments confirm the superior performance of SL-FAC for improving the training efficiency.

Existing dynamic data pruning methods often fail under noisy-label settings, as they typically rely on per-sample loss as the ranking criterion. This could mistakenly lead to preserving noisy samples due to their high loss values, resulting in significant performance drop. To address this, we propose AlignPrune, a noise-robust module designed to enhance the reliability of dynamic pruning under label noise. Specifically, AlignPrune introduces the Dynamic Alignment Score (DAS), which is a loss-trajectory-based criterion that enables more accurate identification of noisy samples, thereby improving pruning effectiveness. As a simple yet effective plug-and-play module, AlignPrune can be seamlessly integrated into state-of-the-art dynamic pruning frameworks, consistently outperforming them without modifying either the model architecture or the training pipeline. Extensive experiments on five widely-used benchmarks across various noise types and pruning ratios demonstrate the effectiveness of AlignPrune, boosting accuracy by up to 6.3% over state-of-the-art baselines. Our results offer a generalizable solution for pruning under noisy data, encouraging further exploration of learning in real-world scenarios. Code is available at: https://github.com/leonqin430/AlignPrune.

Real-time supervisory control of advanced reactors requires accurate forecasting of plant-wide thermal-hydraulic states, including locations where physical sensors are unavailable. Meeting this need calls for surrogate models that combine predictive fidelity, millisecond-scale inference, and robustness to partial observability. In this work, we present a physics-informed message-passing Graph Neural Network coupled with a Neural Ordinary Differential Equation (GNN-ODE) to addresses all three requirements simultaneously. We represent the whole system as a directed sensor graph whose edges encode hydraulic connectivity through flow/heat transfer-aware message passing, and we advance the latent dynamics in continuous time via a controlled Neural ODE. A topology-guided missing-node initializer reconstructs uninstrumented states at rollout start; prediction then proceeds fully autoregressively. The GNN-ODE surrogate achieves satisfactory results for the system dynamics prediction. On held-out simulation transients, the surrogate achieves an average MAE of 0.91 K at 60 s and 2.18 K at 300 s for uninstrumented nodes, with $R^2$ up to 0.995 for missing-node state reconstruction. Inference runs at approximately 105 times faster than simulated time on a single GPU, enabling 64-member ensemble rollouts for uncertainty quantification. To assess sim-to-real transfer, we adapt the pretrained surrogate to experimental facility data using layerwise discriminative fine-tuning with only 30 training sequences. The learned flow-dependent heat-transfer scaling recovers a Reynolds-number exponent consistent with established correlations, indicating constitutive learning beyond trajectory fitting. The model tracks a steep power change transient and produces accurate trajectories at uninstrumented locations.

Autonomous vehicles deployed in remote environments typically rely on embedded processors, compact batteries, and lightweight sensors. These hardware limitations conflict with the need to derive robust representations of the environment, which often requires executing computationally intensive deep neural networks for perception. To address this challenge, we present CADENCE, an adaptive system that dynamically scales the computational complexity of a slimmable monocular depth estimation network in response to navigation needs and environmental context. By closing the loop between perception fidelity and actuation requirements, CADENCE ensures high-precision computing is only used when mission-critical. We conduct evaluations on our released open-source testbed that integrates Microsoft AirSim with an NVIDIA Jetson Orin Nano. As compared to a state-of-the-art static approach, CADENCE decreases sensor acquisitions, power consumption, and inference latency by 9.67%, 16.1%, and 74.8%, respectively. The results demonstrate an overall reduction in energy expenditure by 75.0%, along with an increase in navigation accuracy by 7.43%.

Automated face recognition has made rapid strides over the past decade due to the unprecedented rise of deep neural network (DNN) models that can be trained for domain-specific tasks. At the same time, large foundation models that are pretrained on broad vision or vision-language tasks have shown impressive generalization across diverse domains, including biometrics. This raises an important question: Do different DNN models—both domain-specific and foundation models—encode facial identity in similar ways, despite being trained on different datasets, loss functions, and architectures? In this regard, we directly analyze the geometric structure of embedding spaces imputed by different DNN models. Treating embeddings of face images as point clouds, we study whether simple affine transformations can align face representations of one model with another. Our findings reveal substantial cross-model compatibility: low-capacity linear mappings substantially improve cross-model face recognition over unaligned baselines for both identification and verification, including across foundation models never trained for face recognition. Alignment patterns generalize across datasets and vary systematically across model families, indicating representational convergence in facial identity encoding. These findings reframe independently trained templates as transferable rather than revocable, with implications for interoperability, ensemble design, and biometric template security.

Online reinforcement learning (RL) serves as an effective method for enhancing the capabilities of Android agents. However, guiding agents to learn through online interaction is prohibitively expensive due to the high latency of emulators and the sample inefficiency of existing RL algorithms. We identify a fundamental limitation in current approaches: the Single State Single Action paradigm, which updates the policy with one-to-one state-action pairs from online one-way rollouts without fully exploring each costly emulator state. In this paper, we propose Android Coach, a novel framework that shifts the training paradigm to Single State Multiple Actions, allowing the agent to sample and utilize multiple actions for a single online state. We enable this without additional emulator overhead by learning a critic that estimates action values. To ensure the critic serves as a reliable coach, we integrate a process reward model and introduce a group-wise advantage estimator based on the averaged critic outputs. Extensive experiments demonstrate the effectiveness and efficiency of Android Coach: it achieves 7.5% and 8.3% success rate improvements on AndroidLab and AndroidWorld over UI-TARS-1.5-7B, and attains 1.4x higher training efficiency than Single State Single Action methods PPO and GRPO at matched success rates.

GROMACS is a de-facto standard for classical Molecular Dynamics (MD). The rise of AI-driven interatomic potentials that pursue near-quantum accuracy at MD throughput now poses a significant challenge: embedding neural-network inference into multi-GPU simulations retaining high-performance. In this work, we integrate the MLIP framework DeePMD-kit into GROMACS, enabling domain-decomposed, GPU-accelerated inference across multi-node systems. We extend the GROMACS NNPot interface with a DeePMD backend, and we introduce a domain decomposition layer decoupled from the main simulation. The inference is executed concurrently on all processes, with two MPI collectives used each step to broadcast coordinates and to aggregate and redistribute forces. We train an in-house DPA-1 model (1.6 M parameters) on a dataset of solvated protein fragments. We validate the implementation on a small protein system, then we benchmark the GROMACS-DeePMD integration with a 15,668 atom protein on NVIDIA A100 and AMD MI250x GPUs up to 32 devices. Strong-scaling efficiency reaches 66% at 16 devices and 40% at 32; weak-scaling efficiency is 80% to 16 devices and reaches 48% (MI250x) and 40% (A100) at 32 devices. Profiling with the ROCm System profiler shows that >90% of the wall time is spent in DeePMD inference, while MPI collectives contribute <10%, primarily since they act as a global synchronization point. The principal bottlenecks are the irreducible ghost-atom cost set by the cutoff radius, confirmed by a simple throughput model, and load imbalance across ranks. These results demonstrate that production MD with near ab initio fidelity is feasible at scale in GROMACS.

Large language models (LLMs) have demonstrated strong capabilities in medical question answering; however, purely parametric models often suffer from knowledge gaps and limited factual grounding. Retrieval-augmented generation (RAG) addresses this limitation by integrating external knowledge retrieval into the reasoning process. Despite increasing interest in RAG-based medical systems, the impact of individual retrieval components on performance remains insufficiently understood. This study presents a systematic evaluation of retrieval-augmented medical question answering using the MedQA USMLE benchmark and a structured textbook-based knowledge corpus. We analyze the interaction between language models, embedding models, retrieval strategies, query reformulation, and cross-encoder reranking within a unified experimental framework comprising forty configurations. Results show that retrieval augmentation significantly improves zero-shot medical question answering performance. The best-performing configuration was dense retrieval with query reformulation and reranking achieved 60.49% accuracy. Domain-specialized language models were also found to better utilize retrieved medical evidence than general-purpose models. The analysis further reveals a clear tradeoff between retrieval effectiveness and computational cost, with simpler dense retrieval configurations providing strong performance while maintaining higher throughput. All experiments were conducted on a single consumer-grade GPU, demonstrating that systematic evaluation of retrieval-augmented medical QA systems can be performed under modest computational resources.

Gaussian process ($GP$) regression is a widely used non-parametric modeling tool, but its cubic complexity in the training size limits its use on massive data sets. A practical remedy is to predict using only the nearest neighbours of each test point, as in Nearest Neighbour Gaussian Process ($NNGP$) regression for geospatial problems and the related scalable $GPnn$ method for more general machine-learning applications. Despite their strong empirical performance, the large-$n$ theory of $NNGP/GPnn$ remains incomplete. We develop a theoretical framework for $NNGP$ and $GPnn$ regression. Under mild regularity assumptions, we derive almost sure pointwise limits for three key predictive criteria: mean squared error ($MSE$), calibration coefficient ($CAL$), and negative log-likelihood ($NLL$). We then study the $L_2$-risk, prove universal consistency, and show that the risk attains Stone's minimax rate $n^{-2\alpha/(2p+d)}$, where $\alpha$ and $p$ capture regularity of the regression problem. We also prove uniform convergence of $MSE$ over compact hyper-parameter sets and show that its derivatives with respect to lengthscale, kernel scale, and noise variance vanish asymptotically, with explicit rates. This explains the observed robustness of $GPnn$ to hyper-parameter tuning. These results provide a rigorous statistical foundation for $NNGP/GPnn$ as a highly scalable and principled alternative to full $GP$ models.

Evaluating robustness under temporal distribution shift remains an open challenge. Existing metrics quantify the average decline in performance, but fail to capture how models adapt to evolving data. As a result, temporal degradation is often misinterpreted: when accuracy declines, it is unclear whether the model is failing to adapt or whether the data itself has become inherently more challenging to learn. In this work, we propose three complementary metrics to distinguish adaptation from intrinsic difficulty in the data. Together, these metrics provide a dynamic and interpretable view of model behavior under temporal distribution shift. Results show that our metrics uncover adaptation patterns hidden by existing analysis, offering a richer understanding of temporal robustness in evolving environments.

In this growing age of data and technology, large black-box models are becoming the norm due to their ability to handle vast amounts of data and learn incredibly complex data patterns. The deficiency of these methods, however, is their inability to explain the prediction process, making them untrustworthy and their use precarious in high-stakes situations. SHapley Additive exPlanations (SHAP) analysis is an explainable AI method growing in popularity for its ability to explain model predictions in terms of the original features. For each sample and feature in the data set, an associated SHAP value quantifies the contribution of that feature to the prediction of that sample. Analysis of these SHAP values provides valuable insight into the model's decision-making process, which can be leveraged to create data-driven solutions. The interpretation of these SHAP values, however, is model-dependent, so there does not exist a universal analysis procedure. To aid in these efforts, we present a detailed investigation of SHAP analysis across various machine learning models and data sets. In uncovering the details and nuance behind SHAP analysis, we hope to empower analysts in this less-explored territory. We also present a novel generalization of the waterfall plot to the multi-classification problem.

Deep neural networks can predict human judgments, but this does not imply that they rely on human-like information or reveal the cues underlying those judgments. Prior work has addressed this issue using attribution heatmaps, but their explanatory value in itself depends on robustness. Here we tested the robustness of such explanations by evaluating whether models that predict human authenticity ratings also produce consistent explanations within and across architectures. We fit lightweight regression heads to multiple frozen pretrained vision models and generated attribution maps using Grad-CAM, LIME, and multiscale pixel masking. Several architectures predicted ratings well, reaching about 80% of the noise ceiling. VGG models achieved this by tracking image quality rather than authenticity-specific variance, limiting the relevance of their attributions. Among the remaining models, attribution maps were generally stable across random seeds within an architecture, especially for EfficientNetB3 and Barlow Twins, and consistency was higher for images judged as more authentic. Crucially, agreement in attribution across architectures was weak even when predictive performance was similar. To address this, we combined models in ensembles, which improved prediction of human authenticity judgments and enabled image-level attribution via pixel masking. We conclude that while deep networks can predict human authenticity judgments well, they do not produce identifiable explanations for those judgments. More broadly, our findings suggest that post hoc explanations from successful models of behavior should be treated as weak evidence for cognitive mechanism.

Despite deep learning models running well-defined mathematical functions, we lack a formal mathematical framework for describing model architectures. Ad-hoc notation, diagrams, and pseudocode poorly handle nonlinear broadcasting and the relationship between individual components and composed models. This paper introduces a categorical framework for deep learning models that formalizes broadcasting through the novel axis-stride and array-broadcasted categories. This allows the mathematical function underlying architectures to be precisely expressed and manipulated in a compositional manner. These mathematical definitions are translated into human manageable diagrams and machine manageable data structures. We provide a mirrored implementation in Python (pyncd) and TypeScript (tsncd) to show the universal aspect of our framework, along with features including algebraic construction, graph conversion, PyTorch compilation and diagram rendering. This lays the foundation for a systematic, formal approach to deep learning model design and analysis.

We introduce a code-based challenge for automated, open-ended mathematical discovery based on the $k$-server conjecture, a central open problem in competitive analysis. The task is to discover a potential function satisfying a large graph-structured system of simple linear inequalities. The resulting evaluation procedure is sound but incomplete: any violated inequality definitively refutes a candidate, whereas satisfying all inequalities does not by itself constitute a proof of the corresponding conjecture's special case. Nevertheless, a candidate that passes all constraints would be strong evidence toward a valid proof and, to the best of our knowledge, no currently known potential achieves this under our formulation in the open $k=4$ circle case. As such, a successful candidate would already be an interesting contribution to the $k$-server conjecture, and could become a substantial theoretical result when paired with a full proof. Experiments on the resolved $k=3$ regime show that current agentic methods can solve nontrivial instances, and in the open $k=4$ regime they reduce the number of violations relative to existing potentials without fully resolving the task. Taken together, these results suggest that the task is challenging but plausibly within reach of current methods. Beyond its relevance to the $k$-server community, where the developed tooling enables researchers to test new hypotheses and potentially improve on the current record, the task also serves as a useful benchmark for developing code-based discovery agents. In particular, our $k=3$ results show that it mitigates important limitations of existing open-ended code-based benchmarks, including early saturation and the weak separation between naive random baselines and more sophisticated methods.

While Learned Data Compression (LDC) has achieved superior compression ratios, balancing precise probability modeling with system efficiency remains challenging. Crucially, uniform single-stream architectures struggle to simultaneously capture micro-syntactic and macro-semantic features, necessitating deep serial stacking that exacerbates latency. Compounding this, heterogeneous systems are constrained by device speed mismatches, where throughput is capped by Amdahl's Law due to serial processing. To this end, we propose a Dual-Stream Multi-Scale Decoupler that disentangles local and global contexts to replace deep serial processing with shallow parallel streams, and incorporate a Hierarchical Gated Refiner for adaptive feature refinement and precise probability modeling. Furthermore, we design a Concurrent Stream-Parallel Pipeline, which overcomes systemic bottlenecks to achieve full-pipeline parallelism. Extensive experiments demonstrate that our method achieves state-of-the-art performance in both compression ratio and throughput, while maintaining the lowest latency and memory usage. The code is available at https://github.com/huidong-ma/FADE.

As large language models (LLMs) are increasingly trained on sensitive user data, understanding the fundamental cost of privacy in language learning becomes essential. We initiate the study of differentially private (DP) language identification and generation in the agnostic statistical setting, establishing algorithms and matching lower bounds that precisely quantify the cost of privacy. For both tasks, approximate $(\varepsilon, \delta)$-DP with constant $\varepsilon > 0$ recovers the non-private error rates: $\exp(-r(n))$ for identification (for any $r(n) = o(n)$) and $\exp(-\Omega(n))$ for generation. Under pure $\varepsilon$-DP, the exponents degrade by a multiplicative factor of $\min\{1, \varepsilon\}$, which we show is tight up to constants. Notably, for generation under pure DP with mild assumptions, the upper bound $\exp(-\min\{1,\varepsilon\} \cdot \Omega(n))$ matches the lower bound up to some constants, establishing an optimal rate. Our results show that the cost of privacy in language learning is surprisingly mild: absent entirely under approximate DP, and exactly a $\min\{1,\varepsilon\}$ factor in the exponent under pure DP.

We provide a computational complexity lens to understand the power of machine learning models, particularly their ability to model complex systems. Machine learning models are often trained on data drawn from sampleable or more complex distributions, a far wider range of distributions than just computable ones. By focusing on computable distributions, machine learning models can better manage complexity via probability. We abstract away from specific learning mechanisms, modeling machine learning as producing P/poly-computable distributions with polynomially-bounded max-entropy. We illustrate how learning computable distributions models complexity by showing that if a machine learning model produces a distribution $\mu$ that models the distribution generated by a cryptographic pseudorandom generator at least as well as the uniform distribution does, then $\mu$ must be close to uniform. We view training as compression that removes the randomness from the data, keeping only its structure, and inference as adding randomness back to generate new outputs. A computationally bounded model cannot distinguish pseudorandomness from true randomness, so training removes both, treating anything it cannot efficiently predict as noise. The same holds at inference, where the randomness added back can be, and in practice is, pseudorandom. In the appendix, we make the connection to compression precise and, using a characterization of pseudoentropy by Vadhan and Zheng, show that the best loss achievable by an efficient model is the next-bit pseudoentropy of the data.

As large language models (LLMs) evolve from static chatbots into autonomous agents, the primary vulnerability surface shifts from final outputs to intermediate execution traces. While safety guardrails are well-benchmarked for natural language responses, their efficacy remains largely unexplored within multi-step tool-use trajectories. To address this gap, we introduce **TraceSafe-Bench**, the first comprehensive benchmark specifically designed to assess mid-trajectory safety. It encompasses 12 risk categories, ranging from security threats (e.g., prompt injection, privacy leaks) to operational failures (e.g., hallucinations, interface inconsistencies), featuring over 1,000 unique execution instances. Our evaluation of 13 LLM-as-a-guard models and 7 specialized guardrails yields three critical findings: 1) *Structural Bottleneck*: Guardrail efficacy is driven more by structural data competence (e.g., JSON parsing) than semantic safety alignment. Performance correlates strongly with structured-to-text benchmarks ($\rho=0.79$, $p<0.01$) while we detect no association with standard jailbreak robustness. 2) *Scale Is Not the Deciding Factor*: Trajectory detection accuracy does not scale monotonically with model size, and general-purpose LLMs consistently outperform specialized safety guardrails. 3) *Bounded Temporal Stability*: Detection accuracy improves with trajectory length within native context bounds as models process dynamic execution behavior, but degrades in extreme long-context regimes due to over-refusal.

High-dimensional data are often assumed to lie on lower-dimensional manifolds. We study how to construct diffusion processes on this data manifold using only point cloud samples and without access to charts, projections, or other geometric primitives. Here, we introduce Implicit Manifold-valued Diffusions (IMDs), a data-driven mathematical formalism for defining stochastic differential equations in the original high-dimensional space that describe drifting Brownian particles evolving intrinsically on the underlying manifold. Our construction hinges on approximating the corresponding infinitesimal generator of the diffusion process using a proximity graph over the data and using the carré-du-champ of the generator, which encodes the local tangent spaces of the manifold and lifts the intrinsic process into ambient coordinates. We show that as the number of samples grows, our discrete diffusion process converges in law on the space of probability paths to its smooth manifold counterpart. We further present an Euler-Maruyama scheme for the numerical integration of IMDs. We validate our framework using numerical experiments on synthetic manifolds and the MNIST data manifold, showing that IMDs remain confined over the manifold and enable its guided exploration. Our work provides the mathematical foundation and practical implementations of diffusion processes on data manifolds, opening new avenues for manifold-aware sampling, exploration, and generative modeling.

Emotion annotation is inherently subjective and cognitively demanding, producing signals that reflect diverse perceptions across annotators rather than a single ground truth. In continuous affect prediction, this variability is typically collapsed into point estimates such as the mean or median, discarding valuable information about annotator disagreement and uncertainty. In this work, we propose a distribution-aware framework that models annotation consensus using the Beta distribution. Instead of predicting a single affect value, models estimate the mean and standard deviation of the annotation distribution, which are transformed into valid Beta parameters through moment matching. This formulation enables the recovery of higher-order distributional descriptors, including skewness, kurtosis, and quantiles, in closed form. As a result, the model captures not only the central tendency of emotional perception but also variability, asymmetry, and uncertainty in annotator responses. We evaluate the proposed approach on the SEWA and RECOLA datasets using multimodal features. Experimental results show that Beta-based modelling produces predictive distributions that closely match the empirical annotator distributions while achieving competitive performance with conventional regression approaches. These findings highlight the importance of modelling annotation uncertainty in affective computing and demonstrate the potential of distribution-aware learning for subjective signal analysis.

Mixture proportion estimation (MPE) aims to estimate class priors from unlabeled data. This task is a critical component in weakly supervised learning, such as PU learning, learning with label noise, and domain adaptation. Existing MPE methods rely on the irreducibility assumption or its variant for identifiability. In this paper, we propose novel assumptions based on conditional independence (CI) given the class label, which ensure identifiability even when irreducibility does not hold. We develop method of moments estimators under these assumptions and analyze their asymptotic properties. Furthermore, we present weakly-supervised kernel tests to validate the CI assumptions, which are of independent interest in applications such as causal discovery and fairness evaluation. Empirically, we demonstrate the improved performance of our estimators compared with existing methods and that our tests successfully control both type I and type II errors.\label{key}