Transfer learning presumes that a representation learned on source tasks carries structure that remains usable on related target tasks. Standard evaluations probe this through target accuracy or distributional discrepancy, yet leave unspecified which structural invariant is meant to transfer. We supply that invariant categorically. A source task category $\mathcal A$, a target task category $\mathcal B$, and a task-change functor $J:\mathcal A\to\mathcal B$ determine, for every invariant-valued source representation $F:\mathcal A\to\mathcal V$, the universal transferred invariant $\operatorname{Lan}J F$. Given a target invariant $G:\mathcal B\to\mathcal V$, we define the transfer discrepancy $\operatorname{Comp}J(F,G)=\sup{b\in\operatorname{Ob}(\mathcal B)} d{\mathcal V}\bigl((\operatorname{Lan}_J F)(b),G(b)\bigr)$, evaluating transfer not by an objectwise comparison of source and target, but by comparing the target invariant against the one forced by the prescribed task transformation. We prove finite cokernel formulas for $(\operatorname{Lan}_J F)(b)$ in chain complexes and persistence modules, indexed by the comma category $J\downarrow b$. For persistence-valued finite-type one-parameter invariants, the discrepancy is computed exactly by bottleneck distances between barcodes. Controlled experiments on neural latent point clouds then test whether the score recovers the correct task functor and flags representation collapses that preserve classification accuracy while destroying transfer-relevant topology.

Perception in dense, unstructured urban traffic remains a major challenge for autonomous driving because of the wide variety of road users, frequent occlusions, irregular motion patterns, and the lack of standardized road layouts. Although recent LiDAR based 3D object detectors have shown strong performance in structured driving scenarios, most are developed and evaluated for limited field of view settings, and their behavior under full surround 360-degree sensing is still not well understood. This paper studies a 360-degree LiDAR perception pipeline for autonomous driving, with particular attention to panoramic sensing, azimuthal sector wise spatial processing, and transformation equivariant feature extraction in complex urban scenes. The paper presents a practical 360-degree perception framework that combines sector wise panoramic processing with rotation equivariant sparse convolutions and evaluates its behavior on a custom Ouster OS0 LiDAR dataset collected across diverse Indian urban traffic conditions. The results show generally stable detection across several object classes, with the strongest performance for cars at 92.02/90.51, buses at 80.53/76.34, and trucks at 78.59/74.16, while lower scores for pedestrians at 67.45/61.02, cyclists at 73.21/69.54, and motorcyclists at 71.20/68.13 reflect the greater difficulty of detecting smaller and more variable road users in dense urban scenes.

This discussion argues that sequential statistical inference can naturally contribute to LLM trustworthiness. In deployment, LLM systems are queried repeatedly, conditioned on evolving contexts, and incorporate user or tool feedback, and may exhibit behavioral shifts after model updates or distribution changes. The discussion is organized around three tasks: representation, modeling LLM interactions as dependent stochastic processes rather than isolated prompt--response pairs; validity, developing uncertainty guarantees that remain meaningful under dependence, repeated use, and adaptation; and monitoring, using sequential alarms and change-point detection to identify shifts in calibration, hallucination rates, refusal behavior, fairness, or other task-relevant properties. This perspective complements recent surveys by viewing trustworthy LLM deployment as a problem of statistical process control.

This paper develops a model-theoretic framework for verifying context-conditioned language-model behavior by replacing benchmark labels with finite semantic certificates. The first problem is finite determinacy: when do examples in a context force the answer to a query without changing model parameters? In finite-field linear task families, we prove an exact row-space criterion, compute the residual hypothesis count, derive full and query-local identification curves, and show that extracting a smallest forcing subcontext is NP-complete even for binary outputs. The second problem is threshold emergence: when does an apparent benchmark jump reflect a semantic transition rather than a discontinuity of the scoring map? We prove an anti-mirage theorem separating thresholded metrics from semantic confidence and give a rate-sensitive crossing bound for latent commitments becoming visible above threshold. The common semantic object is a confidence functional on definable events. We show that it is a Boolean probability measure, equivalently a Keisler measure on the relevant type space, whose measure-one formulas form a proper filter and whose Stone-space representation is invariant under definitional expansion. The resulting calculus provides finite context certificates, pair-separator hitting sets, query teaching dimension, prompt-preservation criteria, and scale-limit witnesses. Exact-arithmetic ancillary scripts reproduce the finite-field and threshold calculations and generate the data used by the figures.

Edge services increasingly use federated learning to personalize on-device models while keeping sensitive data local. In practice, deployments must handle heterogeneity in both client resources and local data distributions. Model-heterogeneous federated learning lowers client cost by allowing each client to train a subnet of a shared supernet, but most subnet-allocation policies are driven by device constraints and do not explicitly account for statistical heterogeneity. This paper proposes Heterogeneity-Aware Subnet Allocation (HASA), a train-only rule that assigns subnet widths based on client heterogeneity scores computed from local training data while enforcing a fixed size-weighted compute budget. This design enables budget-matched comparisons with alternative allocation policies. On an article-title next-word prediction benchmark with seven clients, HASA improves unweighted mean client test accuracy over uniform allocation across 10 matched seeds, increasing mean client test accuracy from 13.82 percent to 14.32 percent, and improves worst-client accuracy on average. In a matched-budget comparison with representative partial-training baselines, HASA achieves the strongest worst-client and tail-client accuracy on this benchmark. A directionality ablation shows that assigning smaller subnets to more heterogeneous clients degrades both mean and tail performance. A cross-domain image-classification study further shows that the effectiveness of heterogeneity-aware allocation depends on how well the heterogeneity score reflects clients' need for additional model width.

With the growth of Vision Transformers in safety-critical domains like autonomous systems and medical imaging, ensuring their reliability against soft errors is paramount. While ViTs offer state-of-the-art accuracy, their massive parameter counts render exhaustive fault injection campaigns infeasible. To bridge this gap, a statistical fault injection framework is presented, leveraging finite-population sampling theory to provide formal reliability guarantees. It is demonstrated that failure rates are bounded within a 1% margin at 99\% confidence using only a few thousand samples, regardless of model scale. This methodology achieves up to a 10,700 times reduction in experimental cost compared to exhaustive approaches, while preserving the ability to localize vulnerabilities across architectural components. Through extensive evaluation of different architectures like ViT-Tiny and ViT-Small, a highly non-uniform reliability landscape is uncovered. It is shown that while only 3% of FP32 bit-flips result in failure, the vast majority of these events lead to catastrophic accuracy collapse. Specific vulnerabilities are localized to normalization layers and critical exponent bits within the IEEE-754 format, providing a mathematical foundation and actionable insights for the design of hardened, edge-deployed ViT architectures.

We present a Graph Neural Network (GNN) framework for the classification of finite groups according to their solvability. Using graph representations associated with finite groups, including Cayley graphs (CG), the proposed model is trained to distinguish solvable and non-solvable groups using structural graph information alone. The framework is evaluated on groups outside the training dataset in order to investigate the extent to which GNNs can learn algebraic properties arising in group theory. More broadly, the present work explores the relationship between algebraic structure and graph-based geometric representations of finite groups. The present study is intended as a proof-of-concept investigation of whether GNNs can learn algebraic properties of finite groups from graph-based representations

NVFP4 is a recently introduced hardware-supported FP4 format that improves the fidelity of 4-bit quantization through fine-grained block scales. However, existing NVFP4 scale initialization methods still primarily rely on AbsMax initialization, which leaves a noticeable gap to the optimal solution. To address this, we propose ScaleSweep, a simple and efficient scale optimization method that sweeps over feasible block scale candidates and selects the candidate that minimizes a target objective. We further provide a theoretical analysis of NVFP4 quantization and derive both lower and upper bounds for the required sweep range under mean square error (MSE) and weighted mean square error (WMSE) between the original tensor and the quantized reconstructed tensor. The proposed bounds substantially reduce the sweep space while preserving the optimal candidate, enabling negligible overhead compared with the baseline quantization operators. Experiments on Llama and Qwen models demonstrate that ScaleSweep consistently improves quantization performance over existing initialization methods and further narrows the gap to full precision. In particular, under aggressive end-to-end quantization of weights, activations, KV cache, and query states, ScaleSweep preserves more than 93% of the full-precision performance.

While sparse autoencoders provide features more interpretable than individual neurons, reliably characterizing them remains challenging. We propose Query Lens, which extends Logit Lens to enable more comprehensive and faithful interpretations of sparse features. By jointly considering encoder-side key features and decoder-side value features, we identify both the inputs that activate a feature and the outputs it promotes. We also account for indirect, module-mediated effects that arise when the feature is processed by downstream modules, going beyond the direct effect captured by Logit Lens. In experiments, we find that Query Lens yields coherent token signatures for features that remain uninterpretable under Logit Lens. Finally, we propose the Subspace Channel Hypothesis, suggesting that downstream modules read features through layer-specific subspaces.

Deep neural networks often contain redundant hidden units. Removing individual weights can reduce parameter count, but unstructured sparsity is not always easy to exploit in standard dense implementations. This paper develops a structured pruning framework in which complete neurons are removed using multi-armed bandit (MAB) algorithms. Each candidate neuron is treated as an arm; pulling an arm temporarily masks that neuron, measures the change in loss on a sampled mini-batch, restores the neuron, and updates an estimate of its safe-removal reward. The framework supports stochastic policies, including Epsilon-Greedy, Softmax, UCB1 and Thompson Sampling, and multiplicative-weight policies, including Hedge-style multiplicative weights and EXP3. We evaluate the method on tabular classification, tabular regression and deep neural-network benchmarks covering image, text and reasoning tasks. Statistical comparisons using the Friedman test followed by the Nemenyi post-hoc test show significant differences between methods. On tabular classification tasks, UCB1 obtains the highest mean rank among pruning policies and improves on the unpruned neural network. On regression tasks, UCB1 obtains the highest mean rank and is statistically competitive with, or superior to, several standard regression models according to R^2. On deep-learning tasks, UCB1 and Thompson Sampling obtain the strongest ranks, and several MAB policies significantly outperform the unpruned model, magnitude-based neuron pruning and greedy activation-variation pruning. The results show that MAB-based neuron pruning is an effective and computationally practical approach for structured model reduction.

Reliable measurement of income and consumption is essential for monitoring poverty and inequality in low- and middle-income countries, yet full household surveys are costly and difficult to implement regularly. This paper examines whether reduced survey instruments can preserve key distributional information. We apply Random Forest Recursive Feature Elimination (RF-RFE) to the 2018/19 Nigeria General Household Survey-Panel to identify the income sources, consumption categories and household characteristics that best classify individuals within the welfare distribution. The analysis focuses on three outcomes: poverty status, location in the quintile distribution and position relative to the Gini-based inequality line. The survey's post-planting and post-harvest periods allow us to assess performance under different seasonal contexts. Results show that RF-RFE achieves strong classification accuracy with few predictors. For consumption, poverty status and inequality-line position are accurately predicted using a small set of expenditure categories, while quintile classification reaches about 80 percent accuracy for seasonal consumption and 60--65 percent for annual consumption predicted from a single seasonal visit. For income, poverty status reaches around 90 percent accuracy with five predictors, and inequality-line position is largely captured by labour earnings. The findings suggest that machine-learning methods can help improve survey design and reduce data requirements while retaining much of the distributional information needed to measure and monitor poverty and inequality.

Visual evidence has long been treated as a reliable form of legal proof, but advances in artificial intelligence (AI) are undermining that assumption. This article asks how well humans and frontier multimodal large language models (MLLMs) can distinguish authentic evidentiary photographs from AI-generated counterparts in the object-centric scenarios typical of civil disputes. We built Synthetic Legal Evidence Detection (SLED-1400), a dataset of 200 authentic evidence images paired with 1,200 synthetic counterparts produced by six contemporary text-to-image generators across ten evidence categories. The same stimuli and response format were used in a controlled web experiment with 136 lay participants and in a standardized evaluation of four MLLMs (GPT-5.1, Gemini-3-Pro, Gemini-3-Flash, Qwen3-VL-235B). Human accuracy was 64.8% overall, and 48.5% and 51.0% on the two strongest generators (Gemini-3-Pro-Image and Flux-2-Max), indistinguishable from chance. MLLMs never misclassified an authentic image (100% specificity), but missed most synthetic outputs from the harder generators, with average MLLM detection at 5.9% on Gemini-3-Pro-Image outputs. Human and MLLM errors were largely uncorrelated, while the four MLLMs were strongly correlated with each other. Neither group is a reliable standalone authenticator. We argue that visual evidence in legal proceedings should be treated as inherently contestable, and that a workable procedural response must combine trained human review, MLLM screening, and provenance infrastructure such as C2PA Content Credentials.

State-of-the-art GRPO-style methods for speech-aware large language model post-training suffer from coarse credit assignment, broadcasting the same terminal-reward advantage to every token in a response. This ignores useful structure within rollout batches, where speech-conditioned completions often share prefixes before diverging at important decisions. We propose Low-rank Exploration with Adaptive Forking (LEAF), a retrospective tree-based RL method that recovers this structure without online branching or additional decoding. LEAF samples complete responses, selects high-surprisal boundaries, groups responses by shared prefixes, and assigns span-level advantages using descendant rewards. We theoretically justify LEAF's span-level credit assignment and boundary-selection design. Empirically, LEAF improves over GRPO across speech question answering and speech translation benchmarks under the same rollout and low-rank adaptation budget. Notably, smaller LEAF-trained models outperform current state-of-the-art, full-parameter baselines.

We present a systematic study of fine-tuning OpenAI's Whisper large-v3 for Swiss German ASR, using 1,367 hours of broadcast speech paired with Standard German subtitles as weak supervision. Through 16 iterative training runs on an NVIDIA DGX Spark (Grace Blackwell, 128 GB unified memory, up to 1 PFLOP FP4), we compare LoRA and full fine-tuning of the 1.55B-parameter model, investigate hallucination root causes, and quantify the effect of data quality, subtitle alignment, and training strategy. Our best model achieves 25.6% measured WER on the All Swiss German Dialects Test Set (ASGDTS) in an honest evaluation on strictly disjoint data. A harmonized error analysis separating genuine errors from valid stylistic variation (tense, word order, Swiss orthography) yields a content WER (cWER) of 13.8%, counting only actual recognition failures. Bias-corrected estimation reduces this to 8.5%, suggesting the true error rate is roughly one third of measured WER. We demonstrate that published state-of-the-art Swiss German ASR results (17.1-17.5% WER) are inflated by benchmark contamination: a vanilla Whisper model self-trained on the ASGDTS test set with zero Swiss German data achieves 13.88% WER, surpassing all published systems. Experiments with Phi-4-multimodal show an even stronger memorization effect (3.9% WER), revealing that the benchmark primarily measures convention matching rather than dialectal comprehension. We release two models, a LoRA adapter (25.32% WER, 13.9% cWER) and a full fine-tuned model (25.60% WER, 13.8% cWER), among the few publicly available, honestly evaluated Whisper models for Swiss German, under Apache 2.0 with full reproducibility, requiring no institutional data agreements.

Advances in machine learning and computational power have unlocked the predictive potential of the human genome, yet biologists now demand that these models also elucidate the underlying biological mechanisms. While interpretable machine learning (IML) techniques have been increasingly applied to bridge this gap, there has been a pervasive reliance on anecdotal validation: the vast majority of research relies on a single IML method and reports only isolated successful instances. Through a benchmarking study on transcription factor binding, we demonstrate the risks of current practices. We show that different IML methods can often (1) yield contradictory explanations for the same predictions, (2) fail to localize known regulatory motifs, and (3) fail to faithfully reflect the model's internal decision process. In light of this, we argue for a validation framework analogous to clinical trials: just as trials require rigorous design and adverse-event reporting, genomic interpretability must move beyond cherry-picked plausibility toward systematic assessment of consistency, faithfulness, and biological validity. To facilitate this, we propose a tiered framework to guide rigorous evaluation and reporting of genomic IML methods.

Very low birth weight infants (VLBWI) are at high risk of mortality and severe neurodevelopmental impairment, including cerebral palsy, yet reliable discharge-time prognostic stratification remains challenging in high-dimensional and data-limited clinical settings. To address this problem, we propose QDSP, an interpretable structured learning framework that integrates Quota-guided Subspace Sampling (QSS) and Differentiable-decision-guided Structure Perception (DSP). The QSS module constructs stability-aware and low-redundancy feature subspaces through bootstrap-based feature consistency estimation, whereas the DSP module employs differentiable soft oblique decision structures to model nonlinear clinical interactions while preserving traceable decision evidence. The proposed framework was evaluated on a real-world VLBWI cohort comprising 51 infants and further validated on three public medical tabular datasets. On the primary cohort, QDSP achieved an accuracy of 0.9200 and an AUC of 0.9714, outperforming representative machine learning and deep tabular learning baselines, including XGBoost, TabNet, and TabPFN. Across external datasets, QDSP maintained competitive discrimination and calibration under varying sample sizes and clinical distributions. In addition, SHAP-based analyses and differentiable decision-path tracing identified clinically relevant predictors, including cystic periventricular leukomalacia (cPVL) and birth weight, consistent with established neonatal pathophysiological evidence. These results suggest that QDSP provides an interpretable and robust framework for discharge-time risk stratification in VLBWI and may support early individualized clinical decision-making in neonatal intensive care settings.

Analyzing attention weights has become a standard approach for interpreting the information flow of Large Language Models (LLMs). However, this approach has significant limitations as it neglects the geometric properties of the value vectors being aggregated. To address this gap, we introduce \emph{Contribution Weights}, a projection-based metric that quantifies a token's influence by accounting for it's attention weight, value magnitude, and directional alignment with the layer output. We demonstrate that contribution weights provide a more faithful measure of token importance, consistently outperforming attention-based metrics in identifying semantically critical tokens across different decoder-only models, tasks, and datasets. Further, our metric enables novel mechanistic analysis of \emph{attention sinks}. While previous work characterized sinks as passive repositories for excess attention, we reveal they serve an active functional role, suppressing information through a convex relationship between sink rate and output norm, stabilizing representations by opposing the semantic drift of low-confidence tokens.

Large language models (LLMs) exhibit strong reasoning capabilities, yet most LLM-based agents are statically deployed and unable to improve through task interactions. Existing experience-driven methods often rely on memory or heuristics without enhancing the model's ability to learn, treating it as a passive executor and leading to early performance plateaus and limited long-term improvement. To address this issue, we propose MetaEvo, a two-stage framework for continual agent evolution that focuses on improving how the model learns from tasks experience, rather than solely on what it stores. MetaEvo first applies preference-based optimization to enhance the model's ability of principle abstraction, then enables the accumulation and reuse of these principles within a modular agent architecture. Experimental results on diverse reasoning benchmarks demonstrate that MetaEvo consistently outperforms strong baselines, maintains reliable improvement across iterations. These findings validate the effectiveness of meta-optimization in enabling agents to learn from experience and continually enhance their reasoning capabilities.

LLM-based LEGO assembly generation requires both semantic grounding and physical feasibility. We identify a data-induced failure mode, PhysHack, in which the assemblies satisfy physical-validity constraints while producing structures that are geometrically misaligned, semantically inconsistent, or poorly calibrated. To address this challenge, we propose a model-based data selection approach that uses only a small fraction of the training data while improving physically grounded LEGO assembly generation. Building on the selected trajectories, we introduce PVPO, a sample-efficient reinforcement learning method that couples physical feasibility with voxel-space geometric rewards. Our results show that physical validity alone is an insufficient proxy for reliable physical reasoning: models can learn to generate valid structures without preserving semantic or geometric fidelity. Experiments across model backbones and test-time scaling settings demonstrate that PVPO improves structural and semantic alignment, physical validity, structural stability, and calibration, while reducing reliance on extensive post-hoc rejection sampling. In particular, results on calibration show that PVPO mitigates PhysHack by making test-time selection more predictive of semantic and structural quality.

We introduce the Laplace-Fourier Neural Operator (LFNO), a unified framework for modeling dynamical systems across transient and steady-state regimes by integrating the spectral advantages of Laplace and Fourier Neural Operators. LFNO employs a dual-branch architecture that explicitly decomposes system dynamics into transient and steady-state components. We evaluate LFNO on nine benchmarks, including three ODE systems (Duffing, Lorenz, and Pendulum) and six PDE systems (Euler-Bernoulli beam, Heat, Reaction-diffusion, Brusselator, Burgers, and Navier-Stokes). LFNO significantly outperforms existing operators on ODE systems, where transient dynamics dominate, and consistently surpasses LNO while achieving performance competitive with FNO on PDE benchmarks. Furthermore, LFNO offers improved stability and physical interpretability through its component-wise decomposition. These results demonstrate that LFNO provides a robust and unified approach for learning complex dynamical systems across multiple temporal scales.

We formulate data propagation through the Transformer, the machine learning architecture powering large language models, as a nonlinear control system on the space of probability measures. For the mean-field Transformer model with self-attention and affine feed-forward layers, we prove that Gaussian distributions remain exactly Gaussian along the induced flow. This invariance reduces the infinite-dimensional measure dynamics to a finite-dimensional bilinear control system governing the evolution of the mean and covariance, reformulates the expressive capacity of Transformers as a reachability problem for prescribed Gaussian moments, and reveals a novel connection with Riccati-type equations from classical filtering and control. For time-varying controls, we prove exact finite-time reachability of any target Gaussian distribution whose covariance matrix has the same rank as the initial one, this rank constraint being an intrinsic invariant of the dynamics. For time-invariant parameters, we derive explicit spectral conditions leading either to asymptotic stability toward positive-definite equilibria or to finite-time blow-up of the covariance. Numerical experiments complement the theory by showing that practical Transformers with Gaussian inputs remain close to moment-matched Gaussian distributions through early and intermediate layers, while Transformers with prescribed attention matrices reproduce the predicted covariance regimes: bounded evolution in stabilizing configurations and blow-up in destabilizing ones.

Ordinal Regression (OR) aims to predict target values with inherent order, underpinning critical applications across diverse domains, from recommender systems to computer vision. Though having evolved from naive regression to discretization-based classification and generation, existing paradigms remain fundamentally constrained by quantization artifacts and the lack of global ordinal topological perception. These methods typically enforce rigid boundary delineations, failing to capture the non-stationary semantic transitions inherent to ordinal data. In this paper, we propose a novel paradigm where OR is formulated as a Continuous Generative Ordinal Regression task. Under the novel paradigm, we introduce DiffOR, a unified framework that leverages diffusion models to recover continuous ordinal values via iterative denoising, thereby enabling the dynamic learning of soft semantic transitions. To explicitly preserve ordinal topology, we devise a Dual-Decoupling Strategy: Spatially, Multi-scale Increment Aggregation decomposes targets into hierarchical continuous increments; Temporally, Dynamic Denoising Perception synchronizes denoising steps with feature frequencies, ensuring robust coarse-to-fine refinement. Theoretically, we show that the proposed method can significantly enhance both representation capability and mechanistic interpretability. Extensive experiments on 12 benchmarks across four domains validate DiffOR's consistent superiority over state-of-the-art methods, establishing a new standard that demonstrates strong potential as a general-purpose solution for universal ordinal regression.

We propose a topological framework for comparing trained Graph Neural Networks (GNNs) by mapping the Stochastic Block Models (SBMs) induced on the graphon-signal space of a Message Passing Neural Network (MPNN) onto the unit $n$-sphere $\sphere^{n-1}\subset\R^n$. The construction rests on three classical pillars: the \emph{compactness} of the cut-distance graphon space $(\Wo,\cutdist)$ \citep{lovasz2006limits,lovasz2012large}, the Frieze--Kannan \emph{weak regularity lemma} together with its graphon-signal extension due to \citet{levie2023graphon}, and the Lipschitz continuity of MPNNs with respect to the cut-distance. We show that, for any prescribed tolerance $\varepsilon>0$, a trained MPNN $Φ$ acting on a sufficiently large graph factors (up to $\varepsilon$) through a step-graphon-signal of bounded complexity, and we construct an explicit measure-preserving map $Ψ_n\colon[0,1]\to\sphere^{n-1}$ that places the SBM regions on disjoint spherical caps. This produces a problem-agnostic, low-dimensional ``fingerprint'' of a trained GNN that is amenable to visual inspection and to nearest-neighbour search across model zoos, enabling \emph{transfer-learning candidate retrieval} without retraining. We discuss the obstruction posed by concentration of measure in high dimension -- a phenomenon directly relevant to LLM-scale embeddings. We close with five concrete future research directions: hyperbolic and Grassmannian alternatives to the spherical model, Gromov--Wasserstein distances on graphon-signals as an isometry-free alternative to the $n$-sphere map, an information-geometric (Fisher) reformulation of the SBM manifold, persistent-homology fingerprints of layer-wise embedding clouds, and a spectral-distance baseline derived from the graphon eigendecomposition.

Pre-training data mixtures are commonly tuned by running small-scale experiments and extrapolating to the target training budget. When high-quality data is scarce and must be repeated, this extrapolation frequently fails, but the source of the failure has not been isolated. We show that a primary culprit is a repetition mismatch: because high-quality datasets are small, their repetition rate changes as the training budget grows, shifting the optimal mixture in ways that small-scale proxy experiments do not anticipate. A subsampling procedure that matches the target repetition rate controls for this effect. In a two-source setting combining limited high-quality data with web crawl, a single repetition-controlled experiment using only 1/16 of the target tokens recovers a mixture within 0.05 of the optimum for a 757M parameter model, compared to an error of 0.75 without repetition control. Achieving comparable accuracy without repetition control requires three to four horizons, consuming 44 to 94% of the target token budget. With three data sources, the larger mixture space requires more than a single experiment to constrain, but the approach remains effective: at the 757M scale, just two repetition-controlled horizons recover the optimal mixture, outperforming baselines that instead require the full two-source experiments to construct. Our results reveal that repetition dynamics, not scale alone, shape whether small-scale mixture experiments generalize. More broadly, they suggest that data repetition deserves treatment as a first-class variable in mixture optimization, rather than an inconvenient side effect of limited data.

Fine-tuning often introduces spurious correlations alongside task knowledge, causing systematic failures on underrepresented groups. Existing mitigations require retraining, group labels, or curated counterfactual data. We show a simple post-hoc intervention reduces shortcut reliance without any of these: truncating the tail of the SVD of $ΔW = W_\mathrm{ft} - W_\mathrm{base}$ reduces the spurious-group gap while preserving task accuracy. Across three instruction-tuned models ($0.5$B--$7$B) and four classification benchmarks, top-$k$ truncation reduces the gap on every cell at $<2$ pp accuracy loss, by up to $5\times$ on CivilComments. We propose this works because the shortcut response sits in the tail of the singular ordering of $ΔW$, a claim about how truncation behaves rather than about the raw singular values, which are broadly distributed and look the same across all four datasets. A controlled boundary case in which fine-tuning has only a shortcut to learn shows the predicted FT-to-base collapse, and bottom-/random-$k$ and matched-rank LoRA controls rule out generic low-rank approximation and rank-constrained training as the explanation. We read this as preliminary evidence that the singular basis of $ΔW$ is a useful coordinate system for studying what fine-tuning has learned.

Vision-language agents increasingly consume screenshots, documents, and user interfaces before writing to memory, sending messages, or invoking external tools. We study a concrete failure mode in this setting: action-boundary propagation, where sensitive or unsafe visible text is copied from an image into downstream tool arguments. We present VisualLeakBench, a diversified 500-image benchmark spanning UI, chat, document, form, and dashboard scenes, and evaluate a stratified 100-image agent subset with four production VLM systems under two workflows: note capture and external handoff. At baseline, target strings are propagated into tool arguments in 78.8% of PII cases and 85.5% of rendered unsafe-text cases. Under a defensive system prompt, rendered unsafe-text propagation remains high at 52.6%, while PII tool propagation falls to 2.0%, largely by suppressing tool use rather than preserving utility. Rates are tool-surface dependent: search-like tools suppress PII propagation, but rendered unsafe text still crosses tool boundaries. We measure visual-to-tool propagation rather than downstream instruction execution. We additionally provide a labeled-target oracle upper-bound diagnostic that localizes most failures at the tool boundary while leaving response-side leakage as residual risk.

Personal AI agents must increasingly operate across APIs, shells, web surfaces, and desktop GUIs, yet many systems remain tuned to a single interface and offer limited support for user teaching and auditability. We present Syll, an open-source, self-hosted multimodal agent harness that unifies MCP/API tools, CLI execution, and visual GUI control in a modular runtime, enabling agents to coordinate computer use across heterogeneous interfaces while streamlining how users and agents exchange information. At the core of Syll is a bidirectional user-agent interaction layer: users teach procedures through direct demonstration, which Syll compiles into reusable skills; agent execution is translated back into multimodal evidence -- logs, keyframes, and approval checkpoints -- for inspection and control. Syll further externalizes memory, skills, routines, and governance as editable local artifacts, supporting straightforward inspection, extension, and downstream development. Our implementation has been validated on production desktop applications including Adobe Photoshop, Adobe Audition, Stardew Valley, macOS Finder and others. We report mechanism-oriented studies that validate multimodal routing, teachable GUI replay, and persistent local artifacts. We hope Syll can serve as a practical open-source foundation for personal automation that users can teach, inspect, and continuously extend.

The widespread use of image classification models in high-risk, real-world situations necessitates making these models robust to slight disturbances or perturbations, such as blurring or sharpening, in the input images. While vision transformers (ViTs) play an integral role in many modern-day multi-modal models like Vision-Language-Models (VLMs) and Vision-Language-Action (VLA) models, they have received a lack of attention in the setting of robustness. In this work, we analyze the effects of adversarial fine-tuning, a popular method for improving model robustness to image perturbations, on a ViT's performance on perturbed and regular images through a mechanistic lens. We adversarially train a ViT on low-frequency and high-frequency image corruptions, and attempt to explain changes in downstream model performance through an examination of the model's attention mechanisms, internal representations, and knowledge evolution. Overall, our results suggest that, while fine-tuning on inputs with common corruptions improves model performance and certainty on new instances of corrupted data, these improvements do not transfer to other classes of corruptions not seen in the training. Additionally, despite observing changes in visual attention and knowledge evolution across layers, we found that adversarial training did not lead to fundamental changes in the sparse representations learned by ViTs.

Offline reinforcement learning requires careful conservatism to mitigate distribution shift, yet most existing methods apply a fixed penalty uniformly across all states regardless of local data coverage. We present UNIQ (Uncertainty-Informed Quantile), an offline RL method that introduces state-adaptive conservatism through conformally calibrated uncertainty estimation. Built on the Implicit Q-Learning (IQL) backbone, UNIQ trains a multi-expectile value ensemble, computes distribution-free uncertainty estimates using split conformal prediction, and maps the resulting signal to a state-dependent expectile that relaxes conservatism in well-covered regions while strengthening it in uncertain regions near the data frontier. On D4RL MuJoCo benchmarks, UNIQ consistently improves over IQL, with the largest gains observed on Walker2d and replay-heavy tasks. At the same time, UNIQ operates at near-IQL memory cost (approximately 250 MB peak VRAM), providing roughly a 10x reduction compared to EDAC. Rather than pursuing overall state-of-the-art performance, we position UNIQ as a practical mechanism contribution that improves the performance-efficiency trade-off in offline reinforcement learning.

Pathology foundation models are pretrained on large streams of WSI-derived patches, while supervision during data construction is often slide-level, sparse, or heterogeneous. This mismatch makes it difficult to understand and control which biological patterns enter the pretraining data. We propose SlideCheck, a lightweight pretraining data guidance tool built on frozen pathology foundation model patch features. Rather than serving as a standalone patch diagnostic model, SlideCheck provides explicit abnormality and malignancy scores for organizing, filtering, and auditing pathology pretraining data. SlideCheck uses a dual-head MLP to separately model broad abnormal morphology and malignant evidence. A regularized feature-space scorer provides a supervised anchor for patch-level evidence estimation, while score-attention agreement combines patch scores with WSI-level MIL attention to mine high-confidence pseudo labels. The same scores are then used to construct broad-positive ViT pretraining subsets, where a patch is selected if either abnormality or malignancy evidence exceeds a threshold. Experiments show that SlideCheck-defined data distributions influence the downstream behavior of self-supervised ViT pretraining, indicating that biological composition is an important controllable factor in pathology foundation model development. Curated subsets can approach full-data performance, suggesting that explicitly scored patch pools may support more efficient and auditable pretraining data construction. These findings position SlideCheck as a data guidance and auditing layer for transforming large, undifferentiated patch pools into controllable and reusable pretraining datasets.