Papers

Pre-training large language models is dominated by the memory cost of storing full-rank weights, gradients, and optimizer states. Low-rank pre-training has emerged to address this, and the space of methods has grown rapidly. A central question remains open: do low-rank methods produce models that generalize comparably to full-rank training, or does the rank constraint fundamentally alter the solutions reached? Existing comparisons rely almost entirely on validation perplexity from single-seed runs, often carried forward from prior literature. Yet perplexity is a poor proxy for solution quality; two methods can match on perplexity while converging to different loss landscape regions and internal representations. We close this gap by characterizing the solutions found by five low-rank pre-training methods, GaLore and Fira (memory-efficient optimizers), CoLA and SLTrain (architecture reparameterizations), and ReLoRA (adapter-style updates with periodic resets), against full-rank training at three model scales (60M, 130M, 350M). We evaluate each along 16 metrics across four dimensions: 1-D loss landscape along random/top-K PCA directions, 1-D interpolation between checkpoints, spectral structure of the weights and learned updates, and activation similarity to full-rank training. We show that low-rank methods are not equivalent to full-rank training, nor to one another, even when validation perplexity is close. Full-rank training settles into a sharper basin than low-rank methods along random directions, while the reverse holds for the top-1 PCA direction. Each method converges to a geometrically distinct basin. Low-rank activations diverge from full-rank in later layers as training progresses, with GaLore tracking full-rank most closely. Further, validation perplexity does not translate to downstream performance at every scale. Adding geometric and spectral metrics improves the prediction.

Most anomaly detection systems output scores rather than calibrated decisions, leaving practitioners to choose thresholds heuristically and without clear statistical interpretation. Conformal anomaly detection addresses this limitation by converting anomaly scores into calibrated p-values that are valid under the statistical assumption of data exchangeability, with a growing literature extending this idea beyond that setting. We present nonconform, a Python package for applying conformal anomaly detection within existing machine-learning workflows, and use it as the basis for an implementation-grounded introduction to the field. The package integrates with scikit-learn, PyOD, and custom anomaly detectors, and provides a unified interface for calibration, p-value generation, and false discovery rate control. It supports several conformalization strategies, ranging from simple split-conformal calibration to more data-efficient and shift-aware extensions. Through a progression from foundational concepts to advanced conformalization strategies, complemented by code examples, the paper connects the statistical ideas behind conformal anomaly detection to their practical use in nonconform. Empirical results demonstrate that the implemented methods enable statistically principled anomaly detection. Together, the package and exposition aim to make core conformal anomaly detection workflows more accessible and reproducible in experimental and production-oriented settings.

Complex reinforcement learning environments frequently employ multi-task and mixed-reward formulations. In these settings, heterogeneous reward distributions and correlated reward dimensions often destabilize the construction of scalar advantages. To address these challenges, we propose Reward-Decorrelated Policy Optimization (RDPO), a reward-processing method designed to explicitly target both failure modes. RDPO first utilizes Magnitude-Aware Quantile normalization to stabilize prompt-level advantage allocation across binary, fractional, and continuous rewards. It then applies Mahalanobis whitening within each active reward subspace to mitigate correlation redundancy prior to aggregation. When applied during the post-training of LongCat-Flash, RDPO enhances instruction following, writing quality, and robustness to hard prompts while remaining broadly competitive on reasoning and coding evaluations.

In this paper, we establish last-iterate convergence rates for off-policy actor–critic methods in reinforcement learning. In particular, under a single-loop, single-timescale implementation and a broad class of policy updates, including approximate policy iteration and natural policy gradient methods, we prove the first $\tilde{\mathcal{O}}(\epsilon^{-2})$ sample complexity guarantee for finding an $\epsilon$-optimal policy under minimal assumptions, namely, the existence of a policy that induces an irreducible Markov chain. This stands in stark contrast to the existing literature, where an $\tilde{\mathcal{O}}(\epsilon^{-2})$ sample complexity is achieved only through nested-loop updates and/or under strong, algorithm-dependent assumptions on the policies, such as uniform mixing and uniform exploration. Technically, to address the challenges posed by the coupled update equations arising from the single-loop implementation, as well as the potentially unbounded iterates induced by off-policy learning, our analysis is based on a coupled Lyapunov drift framework. Specifically, we establish a geometric convergence rate for the actor and an $\tilde{\mathcal{O}}(1/T)$ convergence rate for the critic, and combine the two Lyapunov drift inequalities through a cross-domination property. We believe this analytical framework is of independent interest and may be applicable to other coupled iterative algorithms with unbounded

A quantum compiler is a critical piece in the quantum computing pipeline since it allows an abstract quantum circuit to be run on a physical quantum computer. One extremely important subproblem in quantum compilation is the generation of a logical to physical qubit mapping. Typically in quantum compilers this step is either implemented as a random or a heuristic based assignment that aims to minimize additional (SWAP) gate overhead in the quantum circuit. In this paper, we present an alternative approach to solving the qubit mapping problem. Specifically, we formulate the qubit mapping problem with a combinatorial optimization (CO) objective. We then present a method to find a solution to the CO problem by training a reinforcement learning (RL) policy. We also propose a local search based post-processing algorithm to further reduce the overhead. Our results show a dramatic improvement over conventional techniques in reducing the number of SWAPs. On different real world datasets like MQTBench and Queko circuits, our trained policy achieves a 65-85% reduction in SWAP overhead when compared to existing quantum compilers.

Diagnosing esophageal motility disorders pose significant challenges due to the complexity of high-resolution impedance manometry (HRIM) data and variability in clinical interpretation. This work explores the feasibility of a multimodal Machine Learning (ML)-based classification approach that combines HRIM recordings with patient-specific information and incorporates a graph-based modeling of esophageal physiology. We analyze HRIM recordings with corresponding patient information from 104 patients with esophageal motility disorders. Patient data includes demographic, clinical, and symptom information extracted from structured questionnaires and free-text notes using keyword detection and large language model-based processing. HRIM data is represented as spatio-temporal graphs, where nodes correspond to pressure values along the esophagus and edges encode spatial adjacency and impedance dynamics. A graph neural network (GNN) is applied to learn physiologically meaningful representations, which are fused with patient embeddings for multi-category, multi-class classification of swallow events. The impact of patient features and graph-based modeling is evaluated by ablation studies and comparison to vision-based classifier baselines. The proposed multimodal approach indicates improvements over models that rely solely on HRIM-derived features across all classification categories. Additionally, the graph-based modeling provides gains compared to vision-based baselines. Our experiments systematically assess the complementary contribution of multiple modalities, as well as demonstrate the feasibility of our proposed graph-based approach. Our initial findings demonstrate that integrating patient-level data with graph-based representations of HRIM signals appears to be a promising direction for more accurate classification of esophageal motility disorders.

Understanding how deep neural networks learn useful internal representations from data remains a central open problem in the theory of deep learning. We introduce Neural Low-Degree Filtering (Neural LoFi), a stylized limit of gradient-based training in which hierarchical feature learning becomes an explicit iterative spectral procedure. In this limit, the dynamics at each layer decouple: given the current representation, the next layer selects directions with maximal accessible low-degree correlation to the label. This yields a tractable surrogate mechanism for deep learning, together with a natural kernel-space interpretation. Neural LoFi provides a mathematically explicit framework for studying multi-layer feature learning beyond the lazy regime. It predicts how representations are selected layer by layer, explains how emergence of concepts arises with given sample complexity, and gives a concrete mechanism by which depth progressively constructs new features from old ones through low-degree compositionality. We complement the theory with mechanistic experiments on fully connected and convolutional architectures, showing that Neural LoFi improves over lazy random-feature baselines, recovers meaningful structured filters, and predicts representations aligned with early gradient-descent feature discovery with real datasets.

Graph neural networks (GNNs) have emerged as a fundamental tool for learning from graph-structured data, achieving strong performance across a wide range of applications. However, understanding their generalization capabilities remains challenging due to the complex structural dependencies inherent in such data. Existing generalization analyses largely follow the classical machine learning paradigm, focusing primarily on model complexity while overlooking the fundamental role of graph structure. Therefore, in this work, we systematically investigate this role by asking: does the graph structure actually influence generalization, and if so, by how much? To answer the first question and validate our intuition, we theoretically prove that incorporating more edges into the prediction process transforms the input representations to be overly accommodating to the output model, thereby inducing overfitting. To address the second question, we formulate a structural complexity measure based on the number of effective edges and derive a Rademacher complexity-based generalization bound. In doing so, we demonstrate that GNN generalization depends explicitly on structural complexity, alongside traditional parameter-dependent factors. Motivated by these theoretical findings, we propose a structural entropy regularization method. This approach controls structural complexity by regulating effective edges to balance underfitting and overfitting, ultimately improving the generalization performance of GNNs.

Causal discovery, the problem of inferring the direction of causality, is generally ill-posed. We use the language of structural causal models (SCM) to show that assuming that the causal relations are acyclic and invariant across multiple environments (e.g., the way minimum wage affects employment rate is stable across different geographical regions), only two auxiliary environments are sufficient to infer the causal graph for arbitrary nonlinear mechanisms. Moreover, we demonstrate that this implies identifiability of the SCM functional mechanisms: as a corollary, we show that two auxiliary environments are sufficient to guarantee correct counterfactual inference. We empirically support our theoretical results on synthetic data.

Preprocessing screening is often the most expensive part of a near-infrared spectroscopy calibration workflow. It works because smoothing, derivatives, detrending and related filters change the spectral directions seen by partial least squares (PLS) or Ridge regression, but a full external search repeatedly refits nearly the same linear model. This paper studies the case where that search can be collapsed into one calibration step. For a strict linear preprocessing operator A acting on row spectra as XA^T, the transformed PLS cross-covariance satisfies (XA^T)^T Y = A X^T Y, and Ridge regression depends on the operator-induced kernel X A^T A X^T. These identities let a finite operator bank be screened inside the model while retaining original-wavelength coefficients, and the same identity extends to cheaply evaluated linear operator chains. Sample-adaptive or fitted corrections such as SNV, MSC, EMSC and ASLS are not strict linear; we prove the boundary and keep them as fold-local branches. The cohort has 61 regression and 17 classification rows, with a strict paired regression denominator of N=32 for the eight paper variants. There, AOM-PLS reaches median RMSEP ratios of 0.991/0.990 (simple) and 0.985/1.002 (best) against PLS-default/PLS-HPO, and AOM-Ridge reaches 0.974/0.984 (simple) and 0.918/0.966 (best) against Ridge-default/Ridge-HPO. The operator-adaptive classifier AOM-PLS-DA improves balanced accuracy by a median 0.159 on N=13 datasets (12/13 wins). The practical result is the runtime gap: PLS-HPO takes a median 710.81 s per run, whereas AOM-PLS takes 1.18-1.63 s – 436 to 602 times less PLS fitting time. Linear operator-adaptive calibration thus gives prediction quality comparable to exhaustive preprocessing screening, with orders-of-magnitude less fitting time for PLS.

Constraint-based game content generators that learn local constraints from existing content, such as Wave Function Collapse (WFC), can generate visually satisfying game levels but face challenges in optimizing global properties, such as playability. On the other hand, reinforcement-learning-trained generators can optimize global properties—because such properties can easily be included in reward functions—but the results can be visually dissatisfying. In this paper, we explore ways to combine these methods. Specifically, we constrain the action space of a PCGRL generator with constraints learned by WFC, effectively allowing the PCGRL generator to achieve global properties while being forced to adhere to local constraints. To better analyze how this hybrid content generation method operates, we vary the number and type of inputs, and we test whether to randomly collapse the starting state and exclude rare patterns. While the method is sensitive to hyperparameter tuning, the best of our trained generators produce visually satisfying and playable puzzle-platform game levels—such as Lode Runner levels—with desired global properties.

Myocardial infarction (MI) is a leading cause of death, and its adverse outcomes are urgent to predict. Yet ECG-based prognostic models underperform because deep learning requires large, labelled datasets, which are scarce in medicine. Foundation models can learn from unlabelled ECGs via selfsupervision, but medically relevant training strategies remain underexplored. We propose a pretrained artificial intelligence model that combines patient-specific temporal information using contrastive learning with supervised multitask heads, then fine-tunes on post-MI outcome prediction. The proposed model outperformed a model trained from scratch (0.794 vs 0.608 AUC) showing that clinically structured ECG modelling improves classification in limited data regimes.

Land Surface Temperature (LST) is a key variable for various applications, such as urban climate and ecology studies. Yet, existing satellite-derived LST products provide either high spatial or high temporal resolution, resulting in a fundamental trade-off between the two. To address this trade-off, we combine observations from a geostationary and a polar orbiting satellite and provide LST fields at high spatial and high temporal resolution (1 km at 15-min intervals). We demonstrate their application for intraday forecasting of LSTs. To estimate LST fields at high spatiotemporal resolution, a U-Net model is trained to map LST fields from SEVIRI/MSG (3 km and 15 min resolution) to LST fields from Terra/Aqua MODIS (1 km, 4 overpasses per day) that are collocated in space and time. The presented model has been trained on LSTs across large European cities with a population exceeding 1 million inhabitants, and achieves an RMSE = $1.92${\deg}C and near-zero bias MBE = $0.01${\deg}C on the hold-out test set. As a second step, we present an LST nowcasting model based on ConvLSTM architecture, trained across downscaled LST fields with forecast lead times of 15 to 75 minutes. The nowcasting model outperforms a persistence and a Climatological Rolling Median benchmarks, with RMSEs of $0.57$ to $1.15${\deg}C for the considered lead times and biases ranging from $-0.1$ to $0.14${\deg}C. An additional validation conducted against independent MODIS overpasses confirms robust performance. Our LST forecast model at high spatiotemporal resolution is directly applicable to operational satellite-based LST monitoring.

This work studies lung tumor growth prediction from sparse and irregular longitudinal computed tomography (CT) observations with measurement variability. A Bayesian physics-informed neural network is developed by combining Gompertz growth dynamics with low-dimensional Bayesian inference in the log-volume domain. The framework employs a two-stage inference strategy combining maximum a posteriori (MAP) estimation and Hamiltonian Monte Carlo (HMC) sampling to estimate posterior predictive distributions and uncertainty intervals. The method was evaluated on longitudinal data from the National Lung Screening Trial (30 patients). Results show that the model captures heterogeneous tumor growth patterns while maintaining reasonable prediction accuracy under limited observations. Compared with deterministic modeling approaches, the proposed approach additionally provides calibrated uncertainty estimates. The inferred posterior parameter correlations were consistent with expected biological growth behavior. The proposed framework achieved a cohort-level log-space RMSE of approximately 0.20 together with well-calibrated 95% credible interval coverage across 30 patients. These findings suggest that Bayesian physics-informed modeling may be useful for uncertainty-aware tumor growth assessment when only limited longitudinal follow-up scans are available.

Contrastive reinforcement learning (CRL) learns goal-conditioned Q-values through a contrastive objective over state-action and goal representations, removing the need for hand-crafted reward functions. Despite impressive success in achieving viable self-supervised learning in RL, all existing CRL algorithms rely on off-policy optimisation and are mostly constrained to continuous action spaces, with little research invested in discrete environments. This leaves CRL disconnected from widely used and effective, modern on-policy training pipelines adopted across both single-agent and multi-agent RL in continuous and discrete environments. To establish a first connection, we introduce Contrastive Proximal Policy Optimisation (CPPO). CPPO is an on-policy contrastive RL algorithm that derives policy advantages directly from contrastive Q-values and optimises them via the standard PPO objective, without requiring a reward function or a replay buffer. We evaluate CPPO across continuous and discrete, single-agent and cooperative multi-agent tasks. Whilst the existence of an on-policy approach is inherently useful, we observe that CPPO not only significantly outperforms the previous CRL baselines in 14 out of 18 tasks, but also matches or exceeds PPO's performance, which uses hand-crafted dense rewards, in 12 out of the 18 tasks tested.

Neural Posterior Estimation (NPE) enables rapid parameter inference for complex simulators with intractable likelihoods. NPE trains an inference network to estimate a probability density over parameters given data, typically assumed to be continuous. However, many scientific models involve parameter spaces that are mixed, that is, they contain both discrete and continuous dimensions. We address this limitation by extending NPE to mixed parameter spaces through an inference network that jointly handles discrete and continuous parameters. The inference network factorizes the joint posterior into discrete and continuous components, combining an autoregressive classifier for the discrete parameters with a generative model for the continuous parameters, trained jointly under a single simulation-based objective. In addition, we propose a diagnostic tool to assess the calibration of the mixed posterior approximation. Across tractable toy examples and real-world scientific simulators, our joint inference approach yields accurate and calibrated posteriors. The inference framework is available in the sbi Python package.

Intensive care units (ICU) generate long, dense and evolving streams of clinical information, where physicians must repeatedly reassess patient states under time pressure, underscoring a clear need for reliable AI decision support. Existing ICU benchmarks typically treat historical clinician actions as ground truth. However, these actions are made under incomplete information and limited temporal context of the underlying patient state, and may therefore be suboptimal, making it difficult to assess the true reasoning capabilities of AI systems. We introduce RealICU, a hindsight-annotated benchmark for evaluating large language models (LLMs) under realistic ICU conditions, where labels are created after senior physicians review the full patient trajectory. We formulate four physician-motivated tasks: assess Patient Status, Acute Problems, Recommended Actions, and Red Flag actions that risk unsafe outcomes. We partition each trajectory with 30-min windows and release two datasets: RealICU-Gold with 930-window annotations from 94 MIMIC-IV patients, and RealICU-Scale with 11,862 windows extended by Oracle, a physician-validated LLM hindsight labeler. Existing LLMs including memory-augmented ones performed poorly on RealICU, exposing two failure modes: a recall-safety tradeoff for clinical recommendations, and an anchoring bias to early interpretations of the patient. We further introduce ICU-Evo to study structured-memory agents that improves long-horizon reasoning but does not fully eliminate safety failures. Together, RealICU provides a clinically grounded testbed for measuring and improving AI sequential decision-support in high-stakes care. Project page: https://chengzhi-leo.github.io/RealICU-Bench/

Dynamic graphs are ubiquitous in real-world systems, and building generalizable dynamic Graph Foundation Models has become a frontier in graph learning. However, dynamic graphs from different domains pose fundamental challenges to unified modeling, as their semantic and temporal patterns are inherently inconsistent, making the multi-domain pre-training difficult. Consequently, the widely used "pretrain-then-finetune" paradigm often suffers from severe negative knowledge transfer. To the best of our knowledge, there exists no multi-domain dynamic GFM. In this work, we propose DyGFM, a Dynamic Graph Foundation Model over multiple domains based on decoupled and divergence-conditioned prompting. To disentangle transferable semantics from the domain-specific dynamics, we introduce a dual-branch pre-training strategy with semantic-temporal decoupling. To alleviate negative transfer during domain adaptation, we further develop a cross-domain routing mechanism with divergence-aware expert selection. To enable efficient downstream fine-tuning, we design a divergence-conditioned prompt generator that injects lightweight, learnable graph prompts tailored to semantic and temporal traits. Extensive experiments on continuous dynamic graph benchmarks demonstrate that DyGFM consistently outperforms 12 state-of-the-art baselines on both node classification and link prediction tasks, achieving superior effectiveness and efficiency.

Inference-time alignment techniques offer a lightweight alternative or complement to costly reinforcement learning, while enabling continual adaptation as alignment objectives and reward targets evolve. Existing theoretical analyses justify these methods as approximations to sampling from distributions optimally tilted toward a given reward model. We extend these techniques by introducing reference-model temperature adjustment, which leads to further generalization of inference-time alignment to ensembles of generative reward models combined as a sharpened logarithmic opinion pool (SLOP). To mitigate reward hacking, we propose an algorithm for calibrating SLOP weight parameters and experimentally demonstrate that it improves robustness while preserving alignment performance.

High-Level Synthesis (HLS) compiles algorithmic C/C++ descriptions into hardware, with Quality of Results (QoR)—latency and resource utilization—critically governed by pragma configurations and code structure. Existing natural-language-to-HLS (NL-to-HLS) training approaches prioritize functional correctness while largely ignoring QoR. We observe that reinforcement learning (RL) for HLS does not require absolute synthesis results—only relative comparisons between candidates. Based on this insight, we propose HLS-Seek, a QoR-aware NL-to-HLS framework that avoids full synthesis-in-the-loop RL via a comparative proxy reward model achieving 99.53% Pareto-dominance accuracy. To prevent reward hacking, we introduce uncertainty-aware Monte Carlo (MC) dropout switching that selectively invokes real Vitis HLS synthesis for low-confidence candidates and online updates the proxy, creating a self-improving reward system. HLS-Seek achieves 84.7% syntax correctness pass@1 and 81.4% functional correctness pass@5 on HLS-Eval~\cite{abikaram2025hlseval} with only 7B parameters, surpassing GPT-5.1 on functional pass@5, while achieving 8.5$\times$ faster training than real-reward RL. On QoR evaluation, HLS-Seek achieves the lowest latency on 19/30 kernels and Pareto-dominates HLS-specific baselines on 9 kernels.

We address shortcomings of principal component analysis (PCA) for visualizing high-dimensional data lying on a nonlinear low-dimensional manifold via two-dimensional scatterplots, focusing on a fossil teeth dataset from the early mammalian insectivore Kuehneotherium. While the PCA scatterplot reported by Jolliffe and Cadima (Philosophical Transactions of the Royal Society A, 2016) shows clustering in the region where PC2 < 0, our analysis based on t-SNE and persistent homology (PH) reveals a ring-like structure with no evident clustering and intrinsic dimensionality equal to one. We further propose a generative probabilistic-geometric model in which the data are sampled uniformly from a unit circle. Under this model, pairwise cosine distances follow an arcsine distribution, in qualitative agreement with the observed U-shaped distribution, thereby independently supporting the analysis based on t-SNE and persistent homology.

Vector Quantized Variational Autoencoder (VQ-VAE) has become a fundamental framework for learning discrete representations in image modeling. However, VQ-VAE models must tokenize entire images using a finite set of codebook vectors, and this capacity limitation restricts their ability to capture rich and diverse representations. In this paper, we propose ArcCosine Additive Margin VQ-VAE (ArcVQ-VAE), a novel vector quantization framework that introduces a spherical angular-margin prior (SAMP) for the codebook of a conventional VQ-VAE. The proposed SAMP consists of Ball-Bounded Norm Regularization, which constrains all codebook vectors within a time-dependent Euclidean ball, and ArcCosine Additive Margin Loss, which encourages greater angular separability among latent vectors. This formulation promotes more discriminative and uniformly dispersed latent representations within the constrained space, thereby improving effective latent-space coverage and leading to improved codebook utilization. Experimental results on standard image reconstruction and generation tasks show that ArcVQ-VAE achieves competitive performance against baseline models in terms of reconstruction accuracy, representation diversity, and sample quality. The code is available at: https://github.com/goals4292/ArcVQ-VAE

A key technical difficulty in differential privacy is selecting a privacy budget that satisfies privacy requirements while maximizing utility. A natural and well-studied workaround is to use personalized privacy budgets, which may differ across agents. In this paper, we show that personalized budgets come with major limitations and that for mean estimation, the dominant factor is not full personalization, but rather choosing the right effective privacy budget. This can be achieved through a simple thresholding operator that we describe. Compared with this thresholding baseline, the gains obtained by fully personalized mechanisms are limited. In particular, we precisely quantify the constant-factor improvement in settings with mixed private and public datasets and in private datasets with two levels of privacy requirements. We also establish upper bounds and identify regimes of maximal gain for arbitrary privacy requirements.

LLM-based generation of SystemVerilog Assertions (SVA) is often reported as nearing saturation, with the strongest specialized model reaching ${\sim}76\%$ accuracy on NL2SVA-Human. We show that this aggregate hides a temporal gap: models that appear strong overall still collapse to a few implication templates on bounded-delay and liveness specifications. The core issue is that the dominant recipe, supervised fine-tuning on NL/SVA pairs, optimizes token-level mimicry rather than the property equivalence that defines SVA correctness. We introduce Reward-Weighted On-Policy Distillation (RWOPD), an on-policy distillation method that samples student rollouts, scores them with an open SymbiYosys+Z3 Property-Equivalence Checker (PEC), and applies a verifier-reward-weighted forward-KL gradient from a frozen 14B teacher on verifier-passable rollouts. This keeps the supervision dense at every response token while grounding both selection and loss weight in property-equivalent behavior. RWOPD distills CodeV-SVA-14B into a Qwen2.5-Coder-7B-Instruct student that sets a new state of the art on NL2SVA-Human and NL2SVA-Machine across pass@1, pass@5, and pass@10, surpassing both specialized prior SOTA models and 671B general-purpose baselines.

Large Language Models (LLMs) have become increasingly prevalent in cloud-based platforms, propelled by the introduction of AI-based consumer and enterprise services. LLM inference requests in particular account for up to 90% of total LLM lifecycle energy use, dwarfing training energy costs. The rising volume of LLM inference requests is increasing environmental footprints, particularly carbon emissions and water consumption. To improve sustainability for LLM inference serving in cloud datacenter environments, we propose a novel multi-agent game-theoretic reinforcement learning framework called MARLIN to co-optimize time-to-first token (TTFT), carbon emissions, water usage, and energy costs associated with LLM inference. MARLIN demonstrates a reduction of at least 18% in TTFT, 33% in carbon emissions, 43% in water usage, and 11% in energy costs compared to state-of-the-art LLM inference management frameworks.

Flow Matching is a powerful framework for learning transport maps between probability distributions. Yet its standard single-parameter formulation is not designed to capture multi-parameter variations where the resulting transport should be path-independent. Path independence is crucial because it ensures that transformations depend only on the initial and target distributions, not on the specific path. In this work, we introduce Path-independent Flow Matching (PiFM), a method for learning vector fields whose induced flows yield path-independent transport between distributions. We show that PiFM generalizes Flow Matching to higher-dimensional parameter domains while enforcing structural conditions that ensure consistency of composed transformations. In addition, we show that, under suitable assumptions, PiFM approximates the Wasserstein barycenter, linking the framework to a notion of distributional interpolation. To enable practical training, we propose a tractable, simulation-free objective that regresses onto multi-parameter conditional probability paths. We showcase empirically that PiFM outperforms other approaches on both synthetic and real world data in interpolating path-independent trajectories and generating desired out of distribution samples.

Transformers predict over a representation of a sequence. The same data can be written as bytes, characters, or subword tokens, and these representations may be lossless. Yet, under a fixed context window, they need not expose the same information to the model. This raises a basic question: how does the choice of representation change what a finite-context predictor can achieve? We study this question on Markov sources and uncover two complementary phenomena. First, we observe that moving to smaller representation units can hurt prediction even when the context window is enlarged to cover the relevant source history. To explain this, we introduce fragmentation: a lossless recoding that replaces each source symbol by several smaller units. We prove that fragmentation can strictly increase the optimal finite-context log-loss, showing that the gap is not merely an optimization or capacity issue, but can be intrinsic to the representation. This gives a theoretical account of the finite-context gap observed in byte- and character-level models such as ByT5 and CANINE relative to subword-tokenized models. Second, we study the opposite direction: greedy tokenization – BPE, WordPiece, and related methods – which groups source symbols into larger units. We show that tokenization can make a short token window behave like a longer source-context window, and we give a loss guarantee describing when this is achievable. The guarantee depends on how reliably token windows span the needed source history, together with the compression rate of the tokenizer. This also yields a simple diagnostic for real tokenizers: measuring how much source context a fixed token window reliably contains. Together, the two directions establish a finite-context information-theoretic framework for reasoning about representation choices in Transformers.

Calibration is commonly evaluated by comparing model confidence with its empirical correctness, implicitly treating reliability as a function of the confidence score alone. However, this view can hide substantial structure: models may be systematically overconfident on some kinds of inputs and underconfident on others, causing global reliability diagnostics to obscure localised calibration failures. To address this, we formulate the problem of discovering hidden miscalibration regimes without assuming access to predefined data slices. We define the corresponding miscalibration field and propose a diagnostic framework for estimating it. Our approach learns a calibration-aware representation of the input space and estimates signed local miscalibration by kernel smoothing in the learned geometry. Across four real-world LLM benchmarks and twelve LLMs, we find that input-dependent calibration heterogeneity is prevalent. We further show that the discovered fields are actionable: they support local confidence correction and reduce calibration error in systematically miscalibrated regions where confidence-based methods such as isotonic regression and temperature scaling are less effective.

Linear attention and state-space models offer constant-memory alternatives to softmax attention, but often struggle with in-context associative recall. The Delta Rule mitigates this by writing each token via one step of online gradient descent. However, its step size relies on a single scalar gate that ignores the feature-wise curvature of the inner objective. We propose Online Scaled DeltaNet (OSDN), which augments the scalar gate with a diagonal preconditioner updated online via hypergradient feedback. Crucially, this right-preconditioning is algebraically equivalent to a per-feature scaling of the write-side key. This equivalence allows OSDN to strictly preserve the hardware-friendly chunkwise parallel pipeline of DeltaNet without incurring high-dimensional state overhead. Theoretically, by exploiting the exact-quadratic structure of the inner regression loss, we establish super-geometric convergence against a right-Newton comparator and prove an algorithm-aligned token-local residual contraction bound. To handle non-stationary contexts, we further introduce Adaptive Preconditioner Forgetting (APF) to dynamically refresh stale calibration. Empirically, OSDN demonstrates strong performance across scales. At the 340M-parameter scale, OSDN improves JRT-style in-context recall by 32% over DeltaNet. Scaling to 1.3B parameters, it achieves a 39% reduction in the recall residual ratio while maintaining parity on general downstream tasks (e.g., perplexity and LongBench) – demonstrating that our online-preconditioning mechanism effectively transfers and amplifies at the billion-parameter scale.

Perception and planning in physical systems often require inverting a continuous forward map $p \mapsto y = f(p)$, e.g., a renderer or a simulator, that can be queried but not inverted in closed form. Learned inverses are limited by their approximation error, whereas iterative solvers in observation space can stall at spurious stationary points. We instead learn a representation $r(y)$, insensitive to perturbations of $y$ normal to the data manifold, in which $p \mapsto r(f(p))$ is a well-conditioned bijection. In these coordinates, the least-squares inverse problem has no spurious interior stationary points, so that Gauss-Newton iteration with the exact forward model converges to the accuracy the data permit; the learned model need only be qualitatively correct. We introduce Twincher, an exactly invertible architecture of localized area-preserving rotations, trained with Jacobian-level objectives and a domain-growing curriculum. On synthetic maps with tunable nonconvexity, the worst-case residual reaches machine precision within five iterations once the query budget suffices, whereas an MLP-based baseline improves only gradually. In 6D localization of industrial objects in real depth images of the MVTec ITODD benchmark, given instance masks, twinchers trained only on CAD renderings make the iteration converge from a fixed initial guess within a few steps, without feature matching, pose regression, or pose hypotheses; in observation space, the same iteration fails for up to half of the views. All validation poses pass the strictest BOP thresholds and fit the sensor data about as closely as the ground-truth poses; 80% of all 239 instances (67% with visible masks only) are solved in agreement with the sensor data, and the number of training iterations grew only 1.5-fold from two to four pose parameters. Code is open source.