Breast cancer screening programmes increasingly seek to move from one-size-fits-all intervals to risk-adapted and personalized strategies. Deep learning (DL) has enabled image-based risk models with stronger 1- to 5-year prediction than clinical models, but leading systems (e.g., Mirai) typically use convolutional backbones, high-resolution inputs (>1M pixels) and simple multi-view fusion, with limited explicit modelling of contralateral asymmetry. We hypothesised that combining complementary inductive biases (convolutional and transformer-based) with explicit contralateral asymmetry modelling would allow us to match state-of-the-art 3-year risk prediction performance on lower-resolution mammograms, suggesting that using less detail in a more structured way can recover state-of-the-art accuracy. We present MamaDino, a mammography-aware multi-view attentional DINO model. MamaDino fuses frozen self-supervised DINOv3 ViT-S features with a trainable CNN encoder at 512x512 resolution and aggregates bilateral breast information via a BilateralMixer to output a 3-year breast cancer risk score. We train on 53,883 women from OPTIMAM (UK) and evaluate on matched 3-year case-control cohorts: an in-distribution test set from four screening sites and an external out-of-distribution cohort from an unseen site. At breast level, MamaDino achieves discrimination comparable to Mirai on internal and external tests while using ~13x fewer input pixels. Adding the BilateralMixer improves patient-level discrimination to AUC 0.736 (vs Mirai 0.713) in-distribution and 0.677 (vs Mirai 0.676) out-of-distribution, with overlapping confidence intervals. Performance was evaluated across age, ethnicity, scanner manufacturer, tumour type and grade. These findings suggest that complementary inductive biases and explicit contralateral modelling can achieve Mirai-comparable discrimination at lower resolution.
Time series anomaly detection (TSAD) supports the analysis of web service key performance indicators (KPIs) by identifying anomalous intervals. Recent large language model (LLM)-based agentic approaches have achieved promising results through reasoning and tool use. Despite their effectiveness, precise anomaly localization remains challenging, as diverse anomaly patterns demand targeted numerical evidence and iterative refinement to accurately determine anomaly boundaries. In this work, we propose AnomaMind, a coarse-to-fine agentic diagnosis framework for TSAD. AnomaMind first identifies suspicious regions from time-series visualizations, then adaptively acquires targeted numerical evidence through multi-turn tool interactions, and finally refines anomaly decisions through self-reflection. This process is supported by a tool engine that integrates knowledge memory and numerical analysis. To coordinate this diagnostic process, we design a hybrid inference strategy that employs frozen general-purpose LLMs for initial localization, tool orchestration, and evaluation, alongside a specialized LLM detector for evidence-grounded interval prediction. We optimize only the specialized detector through supervised fine-tuning (SFT) on diagnostic trajectories and reinforcement learning (RL) with verifiable detection rewards. Experiments on four benchmarks spanning web traffic, service KPIs, and system-monitoring signals, under both in-domain and multi-domain training settings, demonstrate competitive performance and support the effectiveness of coarse-to-fine adaptive diagnosis through targeted evidence acquisition and self-reflective refinement.The code is available at https://github.com/Xiaoyu-Tao/AnomaMind-TS.
Large language models spend most of their inference cost on attention over long contexts, yet empirical behavior suggests that only a small subset of tokens meaningfully contributes to each query. We formalize this phenomenon by modeling attention as a projection onto the convex hull of key vectors and analyzing its entropic (softmax-like) relaxation. Our main theoretical contribution is a face-stability theorem showing that, under a strict complementarity margin (a support gap (\Delta) certified by KKT multipliers), entropic attention concentrates on a constant-size active face: the total mass assigned to inactive tokens decays exponentially as (\exp(-\Omega(\Delta/\varepsilon))), while the error on the active face scales linearly in the temperature/regularization parameter (\varepsilon). This yields a practical criterion for when sparse long-context decoding is safe and provides a principled knob to trade accuracy for compute. Building on these guarantees, we introduce Vashista Sparse Attention, a drop-in mechanism that maintains a small candidate set per query through a paging-style context selection strategy compatible with modern inference stacks. Across long-context evaluations, we observe stable constant-size effective support, strong wall-clock speedups, and minimal quality degradation in the regimes predicted by the support-gap diagnostics. Finally, we discuss deployment implications for privacy-sensitive and air-gapped settings, where interchangeable attention modules enable predictable latency and cost without external retrieval dependencies.
Mode collapse, the failure to capture one or more modes when targeting a multimodal distribution, is a central challenge in modern variational inference. In this work, we provide a mathematical analysis of annealing based strategies for mitigating mode collapse in a tractable setting: learning a Gaussian mixture, where mode collapse is known to arise. Leveraging a low dimensional summary statistics description, we precisely characterize the interplay between the initial temperature and the annealing rate, and derive a sharp formula for the probability of mode collapse. Our analysis shows that an appropriately chosen annealing scheme can robustly prevent mode collapse. Finally, we present numerical evidence that these theoretical tradeoffs qualitatively extend to neural network based models, RealNVP normalizing flows, providing guidance for designing annealing strategies mitigating mode collapse in practical variational inference pipelines.
The rise of Mean Field Games (MFGs) has fostered a growing family of algorithms designed to solve large-scale multi-agent systems. However, the field currently lacks a standardized evaluation protocol, forcing researchers to rely on bespoke, isolated, and often simplistic environments. This fragmentation makes it difficult to assess the robustness, generalization, and failure modes of emerging methods. To address this gap, we propose a comprehensive benchmark suite for MFGs, focusing on the discrete-time, discrete-space, stationary setting. We introduce a taxonomy of problem classes, ranging from no-interaction and monotone games to potential and dynamics-coupled games, and provide prototypical environments for each. Furthermore, we present MF-Garnets, a method for generating random MFG instances to facilitate rigorous statistical testing. We benchmark a variety of learning algorithms across these environments, including a novel population-based approach (MF-PSO). To our knowledge, this is the first benchmark for MFGs.
Learning Gibbs distributions using only sufficient statistics has long been recognized as a computationally hard problem. On the other hand, computationally efficient algorithms for learning Gibbs distributions rely on access to full sample configurations generated from the model. For many systems of interest that arise in physical contexts, expecting a full sample to be observed is not practical, and hence it is important to look for computationally efficient methods that solve the learning problem with access to only a limited set of statistics. We examine the trade-offs between the power of computation and observation within this scenario, employing the Ising model as a paradigmatic example. We show that the couplings of the model can be learned to a desired accuracy from statistics whose order grows linearly with the $\ell_1$ width of the model and only logarithmically with the inverse of the accuracy. In particular, for a fixed accuracy, the required order is proportional to the width. This allows us to infer the model's structure, using statistics whose order grows logarithmically with the inverse of the weakest coupling, and also to learn its magnetic fields. When the width and the accuracy are fixed, the running time is polynomial in the number of variables. We also discuss a setting where prior information about structure of the model is available and show that the learning problem can be solved efficiently with even more limited observational power.
Training on benchmark test items is a classic way to overestimate generalization to unseen problems. Standard decontamination methods miss semantic duplicates (SemDs): training items equivalent in content to a test item, but distant in string space. We call training on such items soft contamination; when the resulting gains extend to held-out items of the same benchmark (but not to related benchmarks), we call the result shallow generalization. We first embed Olmo 3's entire post-training dataset and search it for SemDs of benchmark items; we then finetune Olmo 3 and Qwen models on natural and synthetic semantic duplicates. We find (1) that soft contamination is widespread: the Olmo corpus contains SemDs of 76% of the CodeForces test set and 100% of MBPP's; (2) that finetuning on SemDs of a benchmark raises accuracy on its held-out items by up to 22pp in controlled experiments and 10-11pp at realistic contamination rates, though the effect varies across benchmarks; (3) that the gains from training on SemDs mostly do not transfer to conceptually similar benchmarks, i.e. we find shallow generalization; (4) that finetuning on items with similar embeddings that are not SemDs produces no such gain, indicating that the mechanism is not mere topical similarity; and (5) that at realistic contamination rates, soft contamination can raise held-out performance about as much as exact duplicates (or more). Benchmark gains here are therefore confounded - in a way that standard decontamination does not address.
In structurally constrained molecular optimization, state-of-the-art methods treat every new input structure as a separate instance, restarting an expensive oracle-driven search from scratch and scaling poorly to settings with many starting structures or expensive oracles. While amortized approaches that learn a transferable policy could in principle remove this bottleneck, existing methods struggle to generalize to diverse structural constraints at inference time. We present AMORTIX, a Graph Transformer model that natively supports such constraints, optimizing molecular structures in a single rollout with zero inference-time oracle calls. A central challenge for amortized training in this domain is that optimization difficulty varies drastically across starting structures. We address this by centering rewards within groups of completions that share the same starting structure. We evaluate on goal-directed scaffold decoration, including single- and multi-target kinase inhibitor design, and on a few-shot prodrug case study. AMORTIX outperforms both amortized baselines and search-based instance optimizers on kinase scaffold decoration, and ablations across 23 objectives show that group-relative centering improves over REINFORCE and PPO and is substantially more consistent across scaffold splits. The prodrug case study further demonstrates transfer of learned modification rules to unseen drug structures.
Linear recurrent networks (LRNNs) and linear state space models (SSMs) promise computational and memory efficiency on sequence modeling tasks, yet their diagonal state transitions limit expressivity. Dense and/or nonlinear architectures (e.g., LSTMs) on the other hand are provably more expressive, but computationally costly. Here, we explore how expressivity in LRNNs can be increased via richer state mixing across time and channels while maintaining competitive efficiency. Specifically, we introduce two structured LRNN architectures: (i) Higher-order Linear Recurrent Units (H-LRU), which generalize recurrences to arbitrary order, mixing multiple past states, and (ii) Block-Diagonal LRUs (BD-LRU), which enable dense intra-block channel mixing. To ensure stable training, we introduce a selective gate normalization scheme that allows for scalable window and block sizes. To maintain efficiency, we implement block-diagonal parallel scans in PyTorch and Triton that keep the throughput competitive with diagonal LRNNs for moderate orders (H-LRU) and block sizes (BD-LRU). Consistent with prior theoretical studies on the limitations of diagonal models, we empirically demonstrate that our architectures benefit from the increased expressivity of structured state mixing, on synthetic sequence modeling and on language modeling tasks that require tracking structure. Our results show that the structure of state mixing is a critical driver of performance in LRNNs, offering a practical pathway to closing the efficiency-expressivity gap in linear sequence models.
In binary classification systems, decision thresholds translate model scores into actions. Choosing suitable thresholds relies on the specific distribution of the underlying model scores but also on the specific business decisions of each client using that model. However, retraining models inevitably shifts score distributions, invalidating existing thresholds. In multi-tenant Score-as-a-Service environments, where decision boundaries reside in client-managed infrastructure, this creates a severe bottleneck: recalibration requires coordinating threshold updates across hundreds of clients, consuming excessive human hours and leading to model stagnation. We introduce MUSE, a model serving framework that enables seamless model updates by decoupling model scores from client decision boundaries. Designed for multi-tenancy, MUSE optimizes infrastructure re-use by sharing models via dynamic intent-based routing, combined with a two-level score transformation that maps model outputs to a stable, reference distribution. Deployed at scale by Feedzai, MUSE processes over a thousand events per second, and over 55 billion events in the last 12 months, across several dozens of tenants, while maintaining high-availability and low-latency guarantees. By reducing model lead time from weeks to minutes, MUSE promotes model resilience against shifting attacks, saving millions of dollars in fraud losses and operational costs.
Large Language Models (LLMs) are advancing rapidly, yet the benchmarks used to measure this progress are becoming increasingly unreliable. Score inflation and selective reporting have eroded the authority of standard benchmarks, leaving the community uncertain about which evaluation results remain trustworthy. We introduce the Benchmark Health Index (BHI), a pure data-driven framework for auditing evaluation sets along three orthogonal and complementary axes: (1) Capability Discrimination, measuring how sharply a benchmark separates model performance beyond noise; (2) Anti-Saturation, estimating remaining headroom before ceiling effects erode resolution and thus the benchmark's expected longevity; and (3) Impact, quantifying influence across academic and industrial ecosystems via adoption breadth and practice-shaping power. By distilling 106 validated benchmarks from the technical reports of 91 representative models in 2025, we systematically characterize the evaluation landscape. BHI is the first framework to quantify benchmark health at a macro level, providing a principled basis for benchmark selection and enabling dynamic lifecycle management for next-generation evaluation protocols.
Instantaneous Quantum Polynomial Circuit Born Machines (IQP-QCBMs) have been proposed as quantum generative models that combine a classically tractable training objective - based on the maximum mean discrepancy (MMD) - with a potential quantum advantage motivated by sampling-complexity arguments. While recent works have explored this model across various application domains, fundamental questions remain: does the model suffer from exponentially vanishing loss gradients, known as the barren plateau problem - a pervasive obstacle in quantum machine learning - and how do regimes of trainability relate to regimes of possible quantum advantage? Here, we address both questions analytically. To study trainability, we derive closed-form expressions for the variances of the partial derivatives of the MMD loss function and establish general upper and lower bounds. We explicitly characterize how trainability depends on the generator set and the spectrum of the chosen kernel, identifying regimes in which low Hamming-weight kernels avoid exponential gradient suppression under structured topologies. Regarding potential quantum advantage, we reformulate the anti-concentration property in terms of the same generator-set quantities that govern trainability, enabling a unified analysis. We show that sparse IQP architectures can produce classically intractable output distributions while simultaneously remaining trainable, at least at lower-weight frequencies. Our analytical results corroborate the numerical observations of prior work and provide principled guidelines for designing scalable IQP-QCBM architectures.
We study multi-task demand learning for retail pricing, where firms seek to estimate heterogeneous linear price-response functions across many decision contexts. Each context has rich covariates but limited price variation, motivating transfer across tasks. However, such transfer is complicated by endogeneity: prices may be arbitrarily correlated with unobserved task-level demand determinants. We propose a meta-learning framework that identifies the conditional mean of task-specific causal demand parameters despite this confounding, provided each task contains at least two distinct, locally exogenous price points. The key is to carefully design the information available to the learner. We condition on the full price history to address cross-task confounding, but withhold the demand outcomes associated with two locally exogenous price points to preserve identifiability. These withheld outcomes provide supervision through a randomized query. We establish that this information design is maximally uniformly valid for the squared-loss objective: any refinement of the conditioning set that reveals information about the withheld outcomes or the query index admits a data-generating process under which the objective no longer identifies the conditional mean causal target. Synthetic experiments demonstrate improved estimation of causal demand parameters under strong confounding, while retail experiments show improved held-out demand prediction relative to standard transfer-learning baselines.
Steering methods influence Large Language Model behavior by identifying semantic directions in hidden representations, but are typically realized through inference-time activation interventions that apply a fixed, global modification to the model's internal states. While effective, such interventions often induce unfavorable attribute-utility trade-offs under strong control, as they ignore the fact that many behaviors are governed by a small and heterogeneous subset of model components. We propose Steer2Edit, a theoretically grounded, training-free framework that transforms steering vectors from inference-time control signals into diagnostic signals for component-level rank-1 weight editing. Instead of uniformly injecting a steering direction during generation, Steer2Edit selectively redistributes behavioral influence across individual attention heads and MLP neurons, yielding interpretable edits that preserve the standard forward pass and remain compatible with optimized parallel inference. Across safety alignment, hallucination mitigation, and reasoning efficiency, Steer2Edit consistently achieves more favorable attribute-utility trade-offs: at matched downstream performance, it improves safety by up to 17.2%, increases truthfulness by 9.8%, and reduces reasoning length by 12.2% on average. Overall, Steer2Edit provides a principled bridge between representation steering and weight editing by translating steering signals into interpretable, training-free parameter updates. Our code is available at https://github.com/Trustworthy-ML-Lab/Steer2Edit
Autoregressive large language models (LLMs) scale well by expressing diverse tasks as sequences of discrete natural-language tokens and training with next-token prediction, which unifies comprehension and generation under self-supervision. Extending this paradigm to multimodal data requires a shared, discrete representation across modalities. However, most vision-language models (VLMs) still rely on a hybrid interface: discrete text tokens paired with continuous Vision Transformer (ViT) features. Because supervision is largely text-driven, these models are often biased toward understanding and cannot fully leverage large-scale self-supervised learning on non-text data. Recent work has explored discrete visual tokenization to enable fully autoregressive multimodal modeling, showing promising progress toward unified understanding and generation. Yet existing discrete vision tokens frequently lose information due to limited code capacity, resulting in noticeably weaker understanding than continuous-feature VLMs. We present Kelix, a fully discrete autoregressive unified model that closes the understanding gap between discrete and continuous visual representations.
Reinforcement Learning (RL) is increasingly applied to large-scale decision-making problems like logistics, scheduling, and recommender systems, but existing algorithms struggle with the curse of dimensionality in such large discrete action spaces. We propose Distance-Guided Reinforcement Learning (DGRL), combining Sampled Dynamic Neighborhoods and Distance-Based Updates to enable efficient RL in problems with up to $10^{36}$ actions. Unlike prior methods, DGRL performs stochastic volumetric exploration and transforms policy optimization into a stable regression task, decoupling gradient variance from action space cardinality. On structured tasks, DGRL provably guarantees local value improvement. DGRL naturally handles action spaces without regular semantic structure and hybrid continuous-discrete action spaces. We demonstrate performance improvements of up to 66% against state-of-the-art benchmarks across regularly and irregularly structured environments, while simultaneously improving convergence speed and computational complexity.
Constraint-based causal discovery requires many conditional independence (CI) tests. Generative CI tests are attractive in nonlinear and high-dimensional settings because they can model complex conditional distributions, but often suffer from high per-test costs and limited reusability across query-dependent conditional resampling tasks in graph search. We propose a generative nonlinear CI testing and graph search framework for PC-style causal discovery. We introduce Langevin-RePaint for conditional imputation. Building on Langevin-RePaint and joint flow matching modeling, we propose the Flow Matching based Conditional Independence Test (FMCIT), which reformulates generative CI testing from query-specific conditional modeling into a reusable conditional resampling framework. We establish theoretical guarantees for FMCIT's type-I error control. FMCIT reuses a single trained generative model to generate randomized copies throughout graph search. We then introduce GPC-FMCIT, which embeds FMCIT into a guided PC pipeline with screening, guidance, refinement, and orientation. By using low-cost screening to shrink the search space, constructing edge-specific candidate conditioning sets from local graph structure, and invoking FMCIT only in a budget-controlled refinement stage, GPC-FMCIT reduces both the deployment cost of individual nonlinear CI queries and the overall query complexity. On synthetic nonlinear structural equation models, GPC-FMCIT improves CPDAG recovery over PC-stable baselines and requires far fewer FMCIT queries than direct FMCIT-based PC search, and on hESC gene-expression data it yields smaller graphs than PC/FisherZ in a descriptive comparison. Overall, our results suggest that, through the joint design of reusable conditional resampling and controlled graph search, generative nonlinear CI testing can become a practical component of PC-style causal discovery.
Large Language Model (LLM) code agents increasingly resolve repository-level issues by iteratively editing code, invoking tools, and validating candidate patches. In these workflows, agents often write tests on the fly, but the value of this behavior remains unclear. This raises a central question: how does agent-written testing relate to task outcomes and resource use during repository-level issue resolution? To better understand the role of agent-written tests, we analyze 5,000 recorded agent runs produced by six strong LLMs on SWE-bench Verified using the lightweight mini-SWE-agent scaffold. Our results show that test writing is common for most models. GPT-5.2 is an exception: it rarely writes new test files while achieving a resolution rate close to the top-ranked model. Test-writing frequency varies widely across models. Within a given model, it differs less between resolved and unresolved tasks. When tests are written, their files contain substantially more statements that print runtime values than assertions that check conditions. We then compare baseline prompts with revised prompts that encourage or discourage writing new test files for four models. Agents start or stop writing test files as requested on 36.8%-93.2% of tasks. Resolved-task totals differ by 0-13 out of 500 per model, with no statistically significant difference, while API calls or token use differ substantially for three models. Similar resolution totals can therefore hide substantial differences in how agents test and how many resources they use.
We study inference-time reward-guided alignment for generative models. Existing methods often rely on either architecture-specific adaptations or computationally costly inference procedures. We introduce Learnable Chernoff Baselines (LCBs) as a method for efficiently and approximately sampling from the exponentially tilted kernels that arise from KL-regularized reward alignment. Using only black-box sampling access to the pretrained model, LCBs implement a form of rejection sampling with adaptively selected acceptance probabilities, which allows fine-grained control over inference-compute scaling. We establish total-variation guarantees to the ideal aligned model, and demonstrate in both continuous and discrete diffusion settings that LCB sampling closely matches ideal rejection sampling while using substantially fewer queries to the pretrained model.
Modern data workflows are inherently adaptive, repeatedly querying the same dataset to refine and validate sequential decisions, but such adaptivity can lead to overfitting and invalid statistical inference. Adaptive Data Analysis (ADA) mechanisms address this challenge; however, there is a fundamental tension between computational efficiency and sample complexity. For $T$ rounds of adaptive analysis, computationally efficient algorithms typically incur suboptimal $O(\sqrt{T})$ sample complexity, whereas for arbitrary distributions, algorithms with $O(\log T)$ sample complexity are computationally intractable under standard cryptographic assumptions. In this work, we shed light on this trade-off by identifying a natural class of data distributions under which both computational efficiency and logarithmic sample complexity in $T$ are achievable. We propose a computationally efficient ADA mechanism that attains $O(\log T)$ sample complexity, optimal in its dependence on $T$, when the data distribution is dense with respect to a known prior. This setting includes, in particular, feature-label data distributions arising in distribution-specific learning. As a consequence, our mechanism also yields a sample-efficient (i.e., $O(\log T)$ samples) statistical query oracle in the distribution-specific setting. Moreover, although our algorithm is not based on differential privacy, it satisfies a relaxed privacy notion known as Predicate Singling Out (PSO) security (Cohen and Nissim 2020) whenever the number of its possible update transcripts is polynomial in the sample size. Our results thus reveal an inherent connection between adaptive data analysis and privacy beyond differential privacy.