Papers

Neural operators serve as fast, data-driven surrogates for scientific modeling but typically rely on a monolithic, single-pass inference procedure that struggles to resolve high-frequency details, a limitation known as spectral bias. We introduce the Iterative Refinement Neural Operator (IRNO), which augments pre-trained operators with a learned refinement module iteratively applied via fixed-point iteration. IRNO decomposes the prediction into a coarse initialization followed by successive residual corrections, paralleling classical numerical solvers. Under local assumptions, we establish contraction of the induced operator, ensuring convergence to a unique fixed point. To explicitly target high-frequency errors, we propose a progressive spectral loss that adaptively increases penalty on high-frequency components over refinement steps during training. Across physical systems, IRNO consistently lowers error, with up to 56.05% improvement on turbulent flow. On Active Matter, spectral analysis reveals that, relative to base operator, the normalized error ratios decrease to 27.72-36.10% in low-, 5.07-6.68% in mid-, and 1.48-2.04% in high-frequencies, remaining stable beyond the trained iteration count. Code is available at https://github.com/xiaotianliu-dartmouth/Iterative_Refinement_Neural_Operator

Aurora visibility at a given location requires two physically distinct conditions to hold at once: aurora occurring overhead, governed by solar wind-magnetosphere coupling, and observing conditions that permit detection, governed by cloud cover and moonlight. Approaches that conflate the two weaken the space-weather signal and limit cross-site generalizability. We develop Aurora Hunter, a two-stage cascade that separates occurrence prediction from observing-condition assessment. Stage 1 uses 43 physics-driven features to predict P(aurora identifiable in all-sky images) via gradient-boosted trees trained on joint Tromso+Kiruna data (about 16,600 hours, 2015-2023). Stage 2 models P(unobscured image class | identifiable) with logistic regression on 15 observing-condition features, trained on hours with identifiable aurora. The cascade P(visible) = P(identifiable) x P(unobscured | identifiable) achieves retrospective ROC-AUC of 0.958 (Tromso test, 2019-2020) and 0.933 (independent Kiruna, 2024), improving on the occurrence stage used alone by +0.095 and +0.097. Transfer to Skibotn, the one station withheld entirely (2022-2025), is limited. SHAP analysis identifies magnetic local time, the Kp x nightside interaction and the three-hour mean Kp as the dominant features (44% of attribution), consistent with auroral oval physics. Hemisphere-wide occurrence maps combine the Stage 1 amplitude, on a clear-sky basis, with the Feldstein oval parameterization. By providing location-specific visibility probabilities with measured reliability rather than coarse geomagnetic indices, the framework links space weather research to practical observation planning. An operational proof of concept is available at https://aurora-hunter.onrender.com

Mechanistic interpretability typically discovers circuits and then argues what they do from examples and ablations. We introduce Verifiable Transformers, a framework for turning task-localized circuits into bounded, solver-checkable claims: projected functional equivalence, task-relevant invariance, edge necessity, and robustness to continuous final-residual perturbations. At small scale, we directly verify all four properties for quote-closing and bracket-type circuits, including program-mediated circuits whose attention selection is entirely symbolic. At GPT-2 scale, we remove LayerNorm from a sparsemax/LeakyReLU model after training with a +0.0087 OpenWebText loss increase, replace retained attention heads with synthesized restricted-DSL programs, and calibrate only program-local readouts while freezing and hashing every other parameter. The resulting three-edge quote circuit (embedding $\to$ MLP 0 $\to$ program head $\to$ logits) verifies all four properties over a hash-pinned 1,280-prompt domain in linear real arithmetic: 1,280/1,280 equivalence and invariance, 640 edge-necessity witnesses per edge, and robustness at $\epsilon = 0.01$ with minimum certified radius 0.01515. For two alphabet variants of the same opener-detection/copy-type task, untouched gates expose a localization frontier: bracket-type extraction is exact only with all 144 heads, while the constructed quote artifact verifies with three edges. Naive discovery-based verification failed for measured reasons; at scale, we find we must build the object we can verify. The verified object is a calibrated artifact, not the unmodified model, and all claims are bounded to declared domains.

Large language models (LLMs) offer a scalable mechanism to elicit domain-informed prior information for high-dimensional variable selection. However, existing methods such as LLM-Lasso are sensitive to weight quality, with performance degrading substantially when LLM-generated weights are inaccurate. To address this challenge, we first introduce a framework for quantifying the quality of LLM-generated weights, enabling rigorous evaluation of LLM-informed methods across varying weight regimes. We then propose the LLM Sparsity Prior (LSP), which integrates LLM-generated weights into the prior inclusion probabilities of Spike-and-Slab and Spike-and-Slab Lasso models via two interpretable hyperparameters governing global sparsity and weight concentration. Hierarchical hyperpriors on these parameters allow the model to dynamically discount uninformative or misleading weights, improving robustness without sacrificing gains when weights are accurate. Finally, we develop principled prompt engineering strategies and validate the method on a private medical dataset studying Acute Kidney Injury. LSP improves prediction accuracy and identifies clinically relevant features missed by the baselines, with robustness to prompt variation and particular effectiveness in low-data regimes.

The popular Cheon-Kim-Kim-Song (CKKS) scheme enables efficient private inference in neural networks by evaluating them on encrypted data. Since CKKS only supports addition, multiplication, and array rotation operations, turning neural networks into CKKS circuits requires approximating all activation functions (e.g. ReLU) with polynomials over fixed input ranges. Traditionally, these ranges are estimated via heuristic sampling techniques. In this paper, we empirically demonstrate that sampled ranges leave the CKKS network exposed to overflow events, i.e. legitimate inputs that exceed the ranges of the circuit and cause corrupted outputs, and propose an improved design technique that completely avoids them. First, we propose a practical algorithm to induce overflow events in a CKKS circuit, i.e. we automatically craft legitimate inputs that exceed the design ranges within the circuit and cause corrupted outputs. As an example, our algorithm can quickly corrupt around 47% of benign inputs in a CIFAR10 neural network. Second, we propose an end-to-end formal verification approach that constructs executable CKKS circuits. These circuits are guaranteed to be overflow-free and come with certified error bounds on the output difference w.r.t. the original neural network. Our experiments demonstrate that our approach supports a larger set of neural network architectures than prior certified techniques, while also delivering executable CKKS artifacts.

Model-based reinforcement learning improves sample efficiency by learning a world model. However, existing latent world models such as DreamerV3 do not explicitly enforce local smoothness in their learned transition dynamics, leaving a useful inductive bias for transition dynamics learning unexploited. We propose GPLD, a gradient-penalized latent dynamics regularizer for DreamerV3 that applies a row-wise Jacobian penalty to the posterior latent distribution to encourage locally smooth transition learning. We show that this penalty can be interpreted as the continuous-latent analog of finite-difference smoothing of transition laws in discrete embedded-state MDPs, and estimate it efficiently using Hutchinson-style stochastic probes. Empirically, across DeepMind Control proprioceptive tasks, GPLD improves aggregate sample efficiency, with particularly strong gains on higher-complexity locomotion environments. On more challenging quadruped tasks, GPLD reaches high-return behavior earlier and exhibits more consistent late-stage learning over longer horizons. Explicit local smoothness regularization is a simple and effective way to improve latent world models for smooth continuous control environments. Code for GPLD is available at github.com/romils9/gpld-mbrl .

Neural collapse (NC) describes the structured geometry that emerges in the features and weights of trained classifiers. Recent theory suggests NC can be suboptimal in deep architectures, attributing this to an explicit low-rank bias from L2 regularization. We study the deep unconstrained feature model (UFM)-equivalent to a deep linear network with orthogonal inputs-trained without regularization, to isolate how gradient descent and depth alone shape NC. We show that depth induces an implicit low-rank bias: low-rank matrices propagate norm more efficiently through successive multiplications, promoting low-rank alternatives to NC. These alternatives, we argue, correspond to softmax codes: max-margin solutions previously found in width-bottlenecked networks. Analyzing training dynamics under spectral initialization, we identify an early-time repulsion among singular values that drives low-rank emergence, and characterize how depth shrinks NC's basin of attraction. Finally, we show that some effects act in the opposite direction: for randomly initialized networks, increasing width biases training toward higher-rank solutions. Our results provide the first asymptotic and dynamic characterization of implicit bias in deep UFMs trained with unregularized multiclass cross-entropy.

Survival analysis aims to model how covariates and time jointly shape the time-to-event distribution under right censoring. Classical methods such as the Cox model and generalised additive models (GAMs) require interactions and time-varying effects to be manually specified, which is increasingly impractical on rich clinical datasets. We introduce KAPLAN-HR, a B-spline Kolmogorov-Arnold Network (KAN) for nonparametric estimation of the conditional hazard as a joint function of covariates and time. A single-layer KAPLAN-HR model recovers a GAM, while deeper architectures capture interactions and time-varying effects through composition. We establish a convergence rate for the nonparametric KAN hazard estimator that depends only on the smoothness of the underlying KAN representation and not on the covariate dimension, thereby mitigating the curse of dimensionality for KAN-representable targets. In evaluations over six clinical benchmark datasets, KAPLAN-HR matches or exceeds the predictive performance of established statistical and deep learning survival methods.

Efficient attention algorithms are critical to mitigate the quadratic cost of attention in long-context workloads. Prior work utilises block-scaled quantisation techniques on Blackwell GPUs to move attention computation to 4-bit precision to accelerate inference. However, these techniques result in significant quality degradation in long-context settings. We show that the output impact of quantisation error is highly non-uniform and increases with the importance of each query-key interaction, concentrating functionally relevant error in a small number of attention blocks that contain the most important tokens. We propose ThriftAttention, a low-bit attention variant that delivers near-FP16 long-context quality at FP4 inference efficiency. This approach proceeds in two stages. First, a heuristic rapidly selects a small number of important query-key block pairs for FP16 precision. Second, the selected blocks are computed in FP16 and the remaining blocks in FP4, with both paths merged via online softmax into a single output. We demonstrate across long-context benchmarks and model families that by computing only 5% of query-key blocks in FP16, ThriftAttention recovers on average 89.1% of the FP4-to-FP16 performance gap. We show ThriftAttention's advantage grows with sequence length, mitigating the systematic FP4 quality degradation observed at longer contexts. The code is available at https://github.com/joesharratt1229/ThriftAttention.

Feature attribution scores each part of an input by how much it explains a model's output. We argue that in generative language models these scores carry no fixed meaning. A classifier has a single output to explain, but a generative model produces its output token by token, and each generated token is both an output and an input, so explaining the output becomes several distinct questions. We support this claim by proposing the Attribution Contract, a framework that names the question a set of attribution scores answers. A contract specifies the model score being explained, what is held fixed, the target output, the generation process, and which features can receive attribution, choices that matter in practice. For example, eligible features decide what a score explains: under a local next-token contract, the model's own generated tokens receive $29\%$ of the attribution mass, comparable to the $25\%$ on the input question, and reading this as a prompt-level explanation is a mistake we name the self-attribution fallacy. The generative process also shapes feature attribution: on a mixture-of-experts model, Integrated Gradients completeness is unreliable, since recomputing expert routing along the attribution path leaves a large residual that holding the routing fixed substantially lowers, while a masked-diffusion model, which fixes its choices at generation, has a residual that instead decreases with more steps. These results show that Attribution Contracts should be carefully specified when proposing and evaluating attribution methods for generative language models, since the meaning and evaluation of an attribution depend on its explanatory setting.

Mixture-of-Experts Large Language Models (MoE-LLMs) achieve strong performance but incur substantial memory overhead due to massive expert parameters. Mixed-precision quantization mitigates this cost by allocating expert-wise bit-widths based on their importance, approaching the accuracy-memory Pareto frontier and enabling extreme low-bit quantization. However, existing methods rely on layer-wise importance estimation and overlook router shifts induced by quantization, resulting in suboptimal allocation and routing. In this work, we propose Global Expert-level Mixed-precision Quantization (GEMQ) to overcome these limitations via (1) a global linear-programming formulation that captures model-wide expert importance based on quantization error analysis, and (2) efficient router fine-tuning to adapt routing to quantized experts. These components are integrated into a progressive quantization framework that iteratively refines importance estimation and allocation. Experiments demonstrate that GEMQ significantly reduces memory and accelerates inference with minimal accuracy degradation. Source code is available at https://github.com/jndeng/GEMQ .

In a landscape of high-performance distributed ML systems, JAX has emerged as a framework of choice. However, JAX's modular design philosophy leaves it without a standardized checkpointing solution. In this paper, we introduce Orbax, a modular, JAX-native checkpointing library that abstracts the complexities of distributed accelerator systems while also providing flexibility for user-friendly checkpoint manipulations throughout the ML model lifecycle. We demonstrate performance exceeding comparable PyTorch competitors by up to 3.5$\times$ for saving and 2$\times$ for loading. The library is available at https://github.com/google/orbax.

Vision foundation models are widely used as frozen backbones across many downstream tasks, making them a single point of failure under adversarial attack. We study multi-level Floyd-Steinberg error-diffusion dithering as a lightweight, model-agnostic input transformation that disrupts adversarial perturbations while preserving semantic content. Unlike prior work, which was limited to binary dithering, grayscale CIFAR-10, and a single small model trained from scratch, we evaluate across six tasks (classification, segmentation, depth estimation, retrieval, captioning, visual question answering), two model families (DINOv2, PaliGemma), and three attacks of increasing strength (PGD, MI-FGSM, SIA), as well as an adaptive attacker using a straight-through estimator. Our results show that Floyd-Steinberg dithering at intermediate quantization levels, especially when combined with post-processing blur, exceeds or matches all tested baselines, including diffusion-based denoising, with substantially less degradation on clean inputs.

Body shape and circumferences are clinically informative biomarkers for risk stratification, including measures such as waist to hip ratio, limb and trunk girths, yet conventional tools such as manual tape measures and optical scanners often require undressing and sustained poses. These demands slow workflows, compromise dignity, and exclude many older adults and people with limited mobility. To make measurement fast and contactless, we leverage millimeter-wave (mmWave) radar, which preserves privacy and operates through typical clothing, enabling quick full-body acquisition. In this work, we present a new optimization-based framework to recover 3D human shape and extract a comprehensive set of anthropometric measurements from volumetric mmWave data. Our method introduces a weighted registration pipeline that fits a parametric body model (SMPL) directly to the noisy mmWave point cloud. The core of our contribution is a vertex-weighting strategy that modulates a Chamfer energy function for reliable surface alignment and noise elimination. We further stabilize the fit by incorporating a foot-ground plane constraint and pose priors, optimizing directly for the SMPL parameters. Together, these components enable a fast, privacy preserving workflow that delivers high fidelity body shape and measurements through clothing without cameras or disrobing and with minimal cooperation, supporting frequent risk oriented assessments in clinics and care facilities for patients of all ages and mobility levels.

Standard neural network training relies on learning-rate schedules tied to a fixed horizon, leading to strong path dependence and costly re-tuning as data availability changes. Schedule-Free (SF) methods address this by removing explicit schedules, yet SF-AdamW, the current state-of-the-art anytime optimizer, consistently underperforms well-tuned AdamW baselines. We propose SF-NorMuon, a schedule-free spectral optimizer that closes this gap: with a single hyperparameter configuration, SF-NorMuon matches or exceeds tuned AdamW on 125M and 772M parameter language models across $1$–$8\times$ Chinchilla horizons, and trails horizon-aware, cosine-scheduled NorMuon by only $\sim 0.03$ nats, substantially narrowing the gap between anytime and scheduled spectral optimizers. On the theoretical side, we prove a stationarity guarantee for schedule-free spectral dynamics and identify weight decay at the fast iterate as essential for long-horizon stability. SF-NorMuon enables practitioners to obtain high-quality checkpoints at any point during training without committing to a horizon in advance. Fully matching horizon-aware spectral optimizers without a schedule remains an open problem, which we highlight as a direction for future work. By closing the performance gap with tuned AdamW baselines, SF-NorMuon makes horizon-free optimization more practical, taking a step towards truly open-ended, continual learning.

RequestRouter is a lightweight request-boundary controller for reducing the latency and energy cost of single-GPU large language model inference. Rather than serving all requests with one static configuration, the system uses cheap request-level features to select one fixed inference mode per request, including FP16, quantized inference, speculative decoding, prefix caching, continuous batching, and hybrid modes such as GPTQ plus prefix caching and INT8 plus continuous batching. We evaluate RequestRouter using an 8B instruction-tuned language model served through vLLM on NVIDIA A100 GPUs. Across the full-scale A100 evaluation–26,500 fixed-mode evaluations followed by 3,500 online-controller evaluations, for 30,000 measured inference executions in total–the controller achieves a 2.10x mean latency speedup over FP16 and a 0.48x energy ratio on deployment-style workloads. A smaller matched evaluation with repeated measurements provides a controlled statistical check of this result: RequestRouter retains a 1.93x latency speedup (95% CI: 1.88–1.98x) and a 0.523 energy ratio (95% CI: 0.506–0.540), showing that the gains persist under a more tightly controlled protocol. On a separate expanded automatic benchmark evaluation, the routed policy retains 99.6% of FP16 macro accuracy. A 100,000-call CPU microbenchmark measures only 0.00475 ms mean routing overhead (0.00532 ms p99). Thus, simple request-aware routing can recover substantial serving efficiency without retraining or modifying the underlying LLM.

Frontier language models sometimes recognize that they are under evaluation and adjust their behavior which can undermine validity of benchmark results. Yet the field studies it without a shared foundation, conflating flaws of the evaluation with capabilities of the model, and detection with behavioral response. We ground evaluation awareness in social psychology, decomposing it into an environment component and a model component that separates recognition from propensity. We operationalize the environment component through eight categorized trigger factors, such as placeholder entities and grading-style output formats, and study recognition and behavior through chain-of-thought monitoring. Across nine frontier models and four benchmarks, recognition rates depend on the specific pairing of model and benchmark. Recognition rarely associates with behavioral change, and when it does, the direction depends on the type of evaluation perceived. Models are also more sensitive to safety than capability evaluations, placing safety benchmark validity at greater risk. To study which factors each model is sensitive to and how they interact, we propose EvalAwareBench, a factor-controlled benchmark of 100 paired safety-capability tasks where each of the eight factors can be independently toggled, varying evaluative signals while holding the underlying request fixed. Through EvalAwareBench, we find that no single factor uniformly affects all models, but stacking factors progressively raises evaluation awareness across all of them. Our framework and EvalAwareBench provide the tools to measure, attribute, and mitigate evaluation awareness, building the foundation for future solutions.

Model collapse, the progressive degradation of LLMs trained on their own outputs, has been characterized statistically but lacks a linguistic explanation for which structures degrade, in what order, and why. We show that iterated learning theory from cultural evolution fills this gap. We derive five falsifiable predictions, distinguish those uniquely discriminative for the theory from confirmatory ones, and test them by self-training LLaMA-2-7B and Mistral-7B over 10 generations in English, German, and Turkish. The critical discriminative finding: compositionality follows a non-monotonic trajectory (initially rising, then falling) under unfiltered self-training. This signature persists with maximally regular seed data (ruling out noise removal) and is sustained only by task-grounded filtering, not random filtering, providing the first LLM-scale evidence for the compression-communication tradeoff. All predictions are confirmed with large effect sizes (Hedges' $g > 1.6$; $\mathrm{BF}_{10} > 100$), and LLM regularization gradients closely match human behavioral data ($R^2 = 0.94$). These results reframe model collapse as a cultural transmission phenomenon and yield concrete principles for self-training pipeline design.

Video representation learning has seen tremendous progress in recent years. This has been driven by many factors, including the scale of training and the success of self-supervised models trained on next-frame prediction. While these factors have pushed the boundaries of what video models can do, they also introduce their own set of limitations. First, scaling video models can reach prohibitive costs, with recent models needing hundreds of years of video to be trained. Second, learning to predict the next frame (or embedding of the frame) naturally focuses on spatial information and as a result, video models still struggle with temporal understanding. In this paper we propose a novel approach that uses motion as a modality to alleviate both of these core issues. Given the motion in a video in the form of point tracks, we mask some of the tracks and use a masked autoencoder to reconstruct the missing tracks. This allows us to learn a representation in a self-supervised manner which we call TIME (Temporally-Informed Motion Embedding), and is focused on capturing temporal information. Also, as motion is inherently appearance-invariant, TIME needs far fewer examples to generalize well. As a result, without bells and whistles, on temporal tasks TIME performs on par with state-of-the-art models, using up to 4 orders of magnitude less training data. For general tasks, TIME can be used in combination with existing representations, and we observe that it leads to a significant improvement for V-JEPA 2, RVM and VideoMAE on standard benchmarks such as SSV2, EgoExo4D and Diving48. These results point to a new promising video paradigm for both more temporally-aware as well as more scalable models.

Latent steering exploits internal representations of Large Language Models (LLMs) to guide generation, yet interventions on dense states can entangle distinct semantic features. In this paper, we investigate attention query activations as a high-fidelity site for precise control, hypothesizing that manipulating the attention mechanism itself offers sharper steerability than general state interventions. We introduce Prototype-Based Sparse Steering, a framework that applies Sparse Autoencoders (SAEs) specifically to query activations, to decompose them into interpretable features, then apply gradient-based optimization during inference to align the sparse representation with class prototypes of target behaviors. To validate this architectural insight, we first analyze the mechanism in Textualized Gridworld, a controlled environment for verifiable planning constraints. We demonstrate that optimizing sparse query features enables effective navigation of rigid planning requirements (i.e., safe vs. short paths), confirming the method's ability to satisfy objective rules. We then demonstrate the framework's versatility by training SAEs on a high-dimensional educational domain, where the framework steers the cognitive complexity of feedback (i.e., Bloom's Taxonomy). Our experiments establish that sparse query representations provide the necessary disentanglement for unified, interpretable control over both logical planning and stylistic nuance.

How do learners acquire knowledge of what is unacceptable without negative evidence? Construction Grammar proposes statistical preemption: exposure to a conventional form (e.g., "donated the books to the library") preempts structurally possible but unattested alternatives ("*donated the library the books"). We present a computational study that, for the first time, directly dissociates statistical preemption from the competing entrenchment hypothesis in large language models within a single converging design. Across four experiments spanning 120 English verb-construction pairings (dative, causative, locative), we show that (1) LLM surprisal patterns correlate strongly with human acceptability judgments ($r = 0.79$), validated against three independent behavioral datasets; (2) these patterns are driven by competing-form frequency rather than overall verb frequency, confirmed by non-circular partial correlations; (3) preemption sensitivity scales as a power law with model size; and (4) a controlled fine-tuning intervention causally demonstrates that manipulating competing-form frequencies shifts preemption behavior in the predicted direction, with reverse-direction controls ruling out frequency-sensitivity confounds. These results provide converging evidence that neural language models acquire negative linguistic knowledge through distributional competition, the core mechanism posited by Construction Grammar.

Two-phase heat transfer underpins boiling, condensation, immersion cooling, flow boiling, energy conversion, and electronics thermal management, but its coupled interfacial physics make data reuse and model comparison difficult. This narrative review synthesizes open datasets, machine-learning methods, and reusable software for two-phase heat-transfer research, with emphasis on boiling, multimodal sensing, and thermal-management datasets. We organize the review around a spatial-plus-temporal dimensionality taxonomy, denoted S+TD, that classifies data objects by the dimensionality of the measured, simulated, or derived fields, including 0+0D point values, 0+1D time series, 1+1D profiles, 2+0D images, 2+1D videos, 3+0D/3+1D fields, and mixed multimodal records. The taxonomy is used to connect dataset types to AI tasks such as tabular regression, acoustic sequence learning, image segmentation, video analysis, inverse heat-flux reconstruction, surrogate modeling, and multimodal fusion. The review also develops a roadmap for physics-aware open data, including metadata definitions, evidence and reuse-maturity labels, benchmark splits, decoders, baseline models, and community databanks. NED3 resources are discussed as one implementation case within a broader open-data ecosystem rather than as a complete solution. The main conclusion is that progress in two-phase AI now depends as much on findable, decodable, benchmarkable, and physically interpretable data infrastructure as on model architecture.

Foundational Models pretrained on huge amount of data learn representations that evolve across depth, forming a hierarchy of embeddings with distinct semantic content and geometric structure. Contrary to the widespread practice of using only the final layer or shallow mixtures, we show that task-relevant information is distributed non-monotonically across layers and cannot be recovered by naïve aggregation. Through a geometric and empirical study across multiple modalities, we show that effective transfer depends on identifying which layers encode task-discriminative structure and how their embeddings are geometrically organized. We introduce Layer-wise Optimal Embedding Selection (LOES), a constructive spectral method that identifies task-discriminative subspaces by minimizing residual error under orthogonality and isotropy constraints. To align fine-tuning with this selection principle, we further propose Geometric Regularization Loss (GeoReg), which enforces a simplicial structure on class manifolds and stabilizes representation geometry during fine-tuning. Across a wide range of architectures, depths, modalities, and data regimes, LOES consistently outperforms standard baselines, with gains that grow as model depth increases. Beyond accuracy, our method reveals how semantic factors are distributed across layers, thereby enabling cross-lingual and cross-modal interpretability analyses. Together, our results provide strong evidence that layerwise embedding geometry is not incidental but central to how deep models represent and transfer knowledge.

Machine learning methods rely on data. However, gathering suitable data can be challenging due to availability constraints, cost, or the need for domain expertise. Expanding datasets with additional sources is a common response to limited data, yet this practice does not always improve downstream performance and can sometimes lead to a loss of performance, known as negative transfer. We propose RADAR, a simple, geometrically grounded metric for estimating cross-domain transferability in foundation models. RADAR analyzes the layer-wise evolution of representations by measuring angular alignments and relative changes in distance along layer-to-layer displacement trajectories, and by comparing empirical distributions of within-domain and cross-domain dynamics. We hypothesize that domain transferability is related to the divergence between these trajectory distributions. We evaluate the metric across multiple modalities, including cross-lingual sentiment classification with text embedding models and cross-domain image classification with foundation vision models. Across several settings, RADAR provides competitive predictive performance relative to existing transferability metrics on several vision and text benchmarks, with particularly strong results when domain transitions are smooth or cleanly separated. Our ablations further suggest that the effectiveness of transferability estimation depends on the geometry of the model's internal representation space, with different modalities favoring different topological formulations.

World models represent a paradigm shift in generative AI, pursuing predictive understanding and controllable simulation of environments in a structured and generalizable way. We present World Machine, a generative world-modeling architecture for time series. It is a transformer-based architecture with latent states that enables adaptation to different amounts of observed data and contexts. This shows an improvement over traditional transformers, which have a computational and memory cost that scales quadratically with the context. Experiments on a proposed synthetic dataset, Toy1D, validate the approach's feasibility, demonstrate capabilities not found in conventional transformers, and highlight the contributions of each component of the training protocol.

Large language models now write software, draft legal documents, and produce clinical notes, yet fundamental limits, from Turing and Arrow to the No Free Lunch theorems, shape what computation can do. This thesis turns such impossibility results from curiosities into design rules. Its flagship result proves an accuracy ceiling set by architecture alone: past a critical reasoning depth, no amount of training moves it, at any adapter rank, sample size, or loss function. Computable before deployment from layer count and embedding width, this Deterministic Horizon is measured between nineteen and thirty-one across twelve transformer architectures, and fine-tuning on optimal-length traces recovers under four percentage points. The mechanism is a capacity invariant of the residual stream, and an information-theoretic conversion yields super-exponential accuracy decay past the horizon. An unconditional circuit-complexity lower bound for modular exponentiation against constant-depth prime-modulus circuits complements this result. The same argument recasts across subfields: preference learning under any misspecified model jumps discontinuously in sample complexity; multi-stage retrieval pipelines require at least as many independent metrics as stages; standard truthful auctions fail for agents with prompt-dependent valuations; and zero-knowledge verification of neural inference pays a measured overhead of one hundred ten to one hundred ninety times per non-linear activation. Together these form a catalogue of sixteen specifications, each pairing a computable boundary, a quantified violation cost, and a constructive design rule: two compositions are proved, one pairing is an honest obstruction, and four remain open. The impossibility-specification methodology is offered for the generative research programme that trustworthy AI may need. Every fundamental limit of AI is also a design rule.

Deploying language-model agents in production often requires substantial compute and human effort to tune prompts, parsers, validators, and other components of the agent pipeline. Self-evolution offers a promising alternative, but most existing frameworks assume access to frontier models that can reliably diagnose failures, propose revisions, and judge their own updates. We study whether frozen small language models (SLMs) can serve as effective self-evolving agents under resource constraints. We propose PACE (Prompt And Control Logic Evolution), a two-timescale framework that coordinates low-risk prompt refinement with higher-risk control-logic updates. PACE evolves prompts under fixed control logic until prompt-level gains saturate, then considers constrained control-logic updates that are accepted through held-out validation. Across three frozen SLM backbones ranging from 4B to 14B parameters and four controlled benchmarks, PACE achieves the best performance on all 12 backbone–benchmark combinations, improving over vanilla SLM agents by up to +9.2% relative improvement and over the stronger single-mode evolution baseline by up to +5.4% relative improvement. A tau-bench case study further shows that PACE improves multi-turn tool-use success over vanilla and prompt-only evolution. These results suggest that reliable SLM agent self-evolution is possible without updating model weights or relying on frontier-model teachers, and that the key benefit is not any single final solver pattern but autonomous, validated discovery of task-appropriate inference strategies.

One prominent method of evaluating machine learning model trustworthiness is the notion of calibration. In the binary outcome setting, a probabilistic predictor is calibrated if outcomes are realized according to a model's distributional prediction, conditioned on this prediction. Straightforward extensions of binary calibration definitions to probabilistic multiclass classifiers suffer from an exponential complexity blowup as the space of predictions grows exponentially in the number of classes $n$. As a remedy, Noarov and Roth (2023) propose multiclass calibration with predictions that are properties of the outcome distribution, reducing complexity from growing in the number of classes $n$ to the dimension $d$ of the property, called its elicitation complexity. Previous work on approximate property calibration is generally limited to continuous scalar properties, despite many relevant properties of interest being discrete, like the mode or rankings. We characterize the approximate property calibration of discrete properties which are strongly orderable by using Lipschitz continuous properties as an intermediary. This work is the first to our knowledge to provide approximate calibration results for discrete properties. Along the way, we characterize the Lipschitz elicitation complexity of strongly orderable discrete properties by constructing algorithms for designing these Lipschitz properties, which we prove can be post-processed to obtain the original discrete property.

We explore the application of LLM-driven algorithm optimization to several common tasks in quantitative finance. MadEvolve, a general-purpose algorithm optimization framework inspired by DeepMind's Alpha-Evolve, was recently developed to optimize algorithms in computational cosmology. Here we demonstrate the utility of MadEvolve to optimize algorithmic trading strategies and alpha generation at the example of Bitcoin trading. On our simulation and backtesting setup, we achieve significant improvements on all tasks we considered, such as evolving feature sets for signal generation, optimizing separate components of the trading strategy, and jointly evolving the feature pipeline together with the execution strategy. Additionally, we compare our method to other agentic search approaches, specifically Claude Code, and carefully evaluate p-hacking probabilities on our simulation setup. Our findings strongly support the utility of AI-driven agentic and evolutionary algorithms for algorithmic trading and quantitative finance.

Active sensing is traditionally defined as the expenditure of energy, typically in the form of movement, for obtaining information. Here, we propose that the combination of reliance on adaptive sensors, the linkage between movement and sensing, and task-level control inevitably gives rise to the emergence of active sensing movements. In this way, active sensing is not driven by sensory goals, such as minimizing uncertainty about the state, but rather is necessary for task-level control. This hypothesis, that active sensing subserves control, is supported by both empirical data from organisms and mathematical theory. Interestingly, active sensing behaviors often occur in discrete epochs, interspersed with goal-oriented behavior. This suggests that animals switch between two behavioral modes with distinct control policies, an `explore' mode in which animals produce dynamic movements to shape sensory feedback, and an `exploit' mode in which animals produce slower compensatory movements that are directly related to achieving task goals. This strategy for feedback control that relies on adaptive sensors, active sensing, and mode switching is not commonly used in engineered systems despite being ubiquitous in biology. Engineered systems comprising state-of-the-art sensors, actuators, and mechanical designs can outperform animals with respect to “cost functions” such as maximum force generation, precision, and speed. Nevertheless, animals routinely achieve robust, graceful behaviors that are currently unmatched by engineered systems, suggesting that current control systems are insufficient. These insights, expressed in the language of control theory, may be critical for improving robotic sensing and control.