Decision models target structured tasks by directly predicting choices or scores, avoiding autoregressive decoding latency. However, transferring a generative LLM's ranking knowledge to a decision model remains challenging: the teacher expresses preferences through a ranked list of document ID sequences, while the student determines the ranking using scores from the first-token position. To bridge this gap, we introduce DecisionRank, a decision model trained through item-space on-policy distillation (OPD). Item-space OPD constructs the student's rollout by sorting its document scores into a ranked sequence. A generative teacher takes the resulting ranking list as inputs and provides conditional feedback to directly supervise those scores. Combined with offline ranking supervision, DecisionRank achieves state-of-the-art performance on both general and reasoning-intensive retrieval benchmarks such as TREC DL, BEIR, and BRIGHT. Across four benchmarks, item-space OPD consistently outperforms alternative training objectives. Even at 1.7B parameters, DecisionRank surpasses its 27B generative teacher on TREC DL19--23. At 8B, it outperforms its generative listwise counterpart while running 24.5faster on TREC DL and 16.7faster on BRIGHT.
Objective: To distinguish the transferable predictive value of physiological trajectories from that of observation patterns in ICU mortality prediction, and to assess whether measurement timing affects performance independently of observation quantity. Materials and Methods: In-hospital mortality was predicted at 48 hours using MIMIC-III, MIMIC-IV, and eICU. Nested representations progressively integrated latest-value snapshots (R1), longitudinal physiological summaries (R2), and measurement frequency and timing features (R3). Models fitted on 7,762 MetaVision patients were evaluated on 21,682 MIMIC-IV patients and 43,878 eICU patients across 204 hospitals. The effect of measurement timing was assessed while holding the number of retained observations constant. Results: Physiological trajectories improved external prediction in eICU (R2-R1 log-loss difference -0.00922, 95% CI -0.01148 to -0.00704). Observation-pattern features improved locally developed eICU models (R3-R2 -0.00257, 95% CI -0.00370 to -0.00146), but not clearly after transfer from MetaVision (-0.00012, 95% CI -0.00206 to +0.00208). R3 was favoured in 10 of 30 source bootstrap refits. At matched observation counts, retaining measurements spread throughout the window reduced log loss relative to random retention by 0.03280 for R3 versus 0.00703 for R2. Removing the final 12 hours increased log loss relative to count-matched random removal by 0.01947 for R3 versus 0.00822 for R2. Under reduced availability, the local R3 increment changed from beneficial to harmful. Discussion: Physiological trajectories transferred more consistently, while observation-pattern features were particularly sensitive to the timing and availability of measurements. Conclusion: Clinical prediction models should be evaluated not only in new patient populations but also under the measurement conditions that supply their inputs.
Understanding how gradient descent learns features remains a central challenge in neural network theory. We study this question in a minimal multi-index setting that already exhibits rich multi-phase gradient dynamics: a well-specified two-neuron ReLU teacher-student model under isotropic Gaussian inputs, trained by population gradient flow from small initialization. We show that the dynamics are organized by two directions: the "easy" bisector direction, which carries the leading signal, and the "hard" splitting direction, which governs specialization. To characterize the loss structure and learning behavior, we analyze the population squared-loss landscape and show that every critical point with both student weights nonzero is either a global minimum or a saddle. We then track the gradient-flow trajectory, showing that student neurons first collapse toward a bisector saddle before escaping and specializing to teacher neurons. Our results provide both a landscape and dynamics account of the multi-phase symmetry-breaking behavior that arises in a simple multi-index model and under standard algorithmic and architectural choices.
In the random allocation sampling scheme each user's data is used in k steps chosen randomly and uniformly from a sequence of t steps. This scheme is known to improve on Poisson sampling in private optimization and in communication-efficient aggregation. The existing privacy loss distribution (PLD) accountant for this scheme runs in time quadratic in the inverse of the accuracy parameter α (Feldman and Shenfeld, 2026). We give an algorithm, multiscale FFT, that represents likelihood ratios on a grid that is uniform within each dyadic interval and computes the exact rounding of every pairwise sum using FFTs within each interval. It keeps the guarantees of the accountant and improves the dependence of its running time on α from O(α^{-2}) to O(α^{-1}(1/α)). In addition, we demonstrate that rounding can be performed via the mean-preserving connect-the-dots projection of (Doroshenko et al., 2026). This reduces the errors that rounding introduces in the mean of the privacy loss and in the χ^2 divergence of the computed pair from the order of α to O(α^2/(t)). In our experiments, our algorithm achieves speedups of several orders of magnitude at high accuracy and gives improved privacy bounds for DP-SGD with PREAMBLE.
In real-world deployments, large language model (LLM) systems are increasingly evolving from single-model invocation toward heterogeneous deployments in which different models are dynamically invoked for inference. When the same context is reused across models, however, each newly selected target model typically has to process it again, repeatedly incurring substantial computational costs. Therefore, we propose ACCT (Asymmetric Cross-LLM Context Transfer), which targets KV cache transfer between models of different scales within the same model family. ACCT transforms the KV cache produced by a source model into sparse KV states that can be directly consumed by a target model, thereby enabling cross-model reuse of context computation and decoupling context computation from downstream inference. ACCT consists of three lightweight components: the Cartographer identifies query-relevant context tokens for each target layer and KV head; the Allocator determines how many context tokens each target layer and KV head retains under a shared total KV budget; and the Compiler transforms selected source KV into sparse target KV. The target model can therefore perform inference using the transferred KV without reprocessing the original context. We evaluate ACCT in two complementary settings comprising five general-capability benchmarks and six highly context-dependent tasks. Experiments show that ACCT matches or exceeds the target model's original inference performance even when reducing the KV cache size by 75%. At a context length of approximately 19K tokens, ACCT's KV conversion achieves a 6.5-21.2 times speedup over target context prefill. These results demonstrate the feasibility of transferring context computation across models under a limited KV budget.
The quadratic scaling of attention is a fundamental bottleneck limiting the long-context capabilities of large language models (LLMs). A variety of approaches have attempted to circumvent this, but they either fail to match attention's expressiveness (e.g., sliding-window or linear attention variants with fixed-size state) or only achieve constant-factor speedups (e.g., sparse attention methods whose indexers themselves scale quadratically). In this paper, we derive an asymptotically subquadratic sparse attention method based on a hierarchical, tree-structured indexer, which we call Zyphra Hierarchical Attention (ZHA). ZHA constructs a tree of recursively compressed key summaries over the KV cache, which future queries traverse from the top down to retrieve a sparse subset of keys. Crucially, by iteratively growing the tree's depth, the total cost including routing and final attention grows sublinearly in sequence length, resulting in overall subquadratic (but superlinear) scaling for attention. We demonstrate that ZHA can be cheaply applied to existing pretrained models with only a short KL distillation training stage over a few billion tokens, and maintains performance on long-context retrieval and challenging reasoning, mathematical, and agentic coding evaluation tasks. Moreover, we show that our method is highly robust to tree geometry including very deep trees of up to 6 levels. Finally, we develop a detailed performance model and show that our approach scales subquadratically, dramatically outperforms competing methods beyond 1 million context, and ultimately has the potential to extend effective context windows to the hundreds of millions or billions of tokens.
Atmospheric fronts are associated with much of the high-impact weather in the mid-latitudes, yet their identification and classification remain largely manual and subjective. Existing deep learning approaches rely primarily on gridded reanalysis or forecast data, limiting their applicability to satellite observations available for time-critical applications. We introduce SATFRONT, a machine learning framework for detecting cold, warm, and occluded fronts over Europe from geostationary satellite observations, with reanalysis data as an optional additional input. The model is trained on frontal analyses from five European national meteorological services and evaluated against UK Met Office analyses for 2024. Using satellite observations alone, SATFRONT achieves a Critical Success Index (CSI) of 0.428 at a 150 km tolerance, within the range of agreement between analyses from different meteorological services (0.277-0.501) and higher than that of an operational rule-based front detector that uses satellite data (0.133). Adding ERA5 reanalysis increases the CSI to 0.591. ERA5 substantially improves the placement of individual fronts, although its effect on their annual spatial distribution varies by front type. These results demonstrate that cold, warm, and occluded fronts can be detected directly from geostationary satellite observations, providing a basis for automated frontal analyses at the high temporal frequency required for time-critical applications.
Distributionally robust optimization (DRO) minimizes worst-case risk under distributional uncertainty, but its decisions can be suboptimal in terms of their excess risk, or regret. We study this gap for multivariate least-squares regression, assuming that the observed source law and unknown target law differ by an adversarial 2-Wasserstein perturbation of size . We first identify the asymptotic complexity of robust decision-making in the 0 limit, characterized by a problem-specific condition number κ. Here, a wide spectrum of simple algorithms, including DRO, achieve the instance-optimal rate of O(κ^2). For fixed > 0, we prove that DRO still achieves the minimax rate if the problem has bounded residual effective rank. On the other hand, we identify significant failure modes where DRO is provably suboptimal by dimension-dependent factors. Fortunately, we prove that two alternative procedures, minimax regret optimization (MRO) and spectral-norm regularization (SNR), attain the minimax rate. We derive an exact convex reformulation of MRO and an accelerated proximal algorithm for SNR, along with high-probability finite-sample guarantees. We support these theoretical results with numerical experiments separating the performance of MRO and SNR from DRO.
Can't agree on where to eat? Zest's AI concierge can combine everyone's tastes, put restaurant options to a vote, and make the reservation.
Physics informed machine learning provides a framework to study and simulate fluid dynamics, potentially at a fraction of the computational cost of traditional numerical methods, with the goal of improving the explainability, accuracy and robustness of purely data driven machine learning methods. In this work, we extend the hard constraint projection (HCP) method developed by Chen et al. [Journal of Computational Physics, 445, 110624, 2021] from a linear partial differential equation (PDE) to a nonlinear PDE: the 1D Burgers' equation. To strictly impose a discrete approximation of the governing equation, the Burgers' equation is first locally discretised (using a finite difference stencil) then decomposed into a constraint matrix and a prediction matrix. Using tools from linear algebra, the constraint matrix is used to construct a projection matrix. A physics informed neural network (PINN) [Raissi et al. Journal of Computational Physics, 378, 686, 2019] is used to predict the velocity, and a final unlearnable layer projects the velocity prediction to the hyperplane of exact solutions to a discretised form of the Burgers' equation. Our initial results indicate that the HCP-PINN initially trains similarly to an equivalent PINN, and as the PINN loss plateaus, the HCP-PINN loss continues to decline. When tested on a previously unseen uniform grid of points covering the domain, the HCP-PINN exhibits higher accuracy, converging to a data error approximately an order of magnitude lower than the comparable PINN. We also observe that the inclusion of the HCP layer results in better observance of the governing PDE. Our findings suggest that the HCP-PINN can be more robust to the selection of physics collocation points, and less dependent on collocation point quantity.
Inverse reinforcement learning (IRL) seeks to infer the reward function underlying observed behavior, but its predictive accuracy and interpretability depend on how that reward is computed from the underlying features. While prior IRL reward functions have been implemented through a variety of techniques including linear weights, neural networks, or other function approximators, none make assumptions about the geometry of the underlying feature space. Inspired by geometric organization observed in neural circuits, we introduce Hyperbolic inverse reinforcement learning (Hyperbolic IRL), which parameterizes reward as a function of Hyperbolic distance between the feature vector and a learned parameter vector, as well as its Euclidean distance variant. We tested our methods against a range of linear and nonlinear IRL baselines under both deterministic and stochastic policies. Our Hyperbolic approach achieved comparable performance to state-of-the-art models on established benchmarks and excelled in the low data limit and across dense environments. Applied to human navigation among multiple exhibits in a museum room, Hyperbolic IRL reveals reward structure that depends not only on individual exhibit-distance features, but also on non-separable conjunctions between them. Together, these results demonstrate that a Hyperbolic reward parameterization can improve both the predictive accuracy and interpretability of IRL, providing a framework for studying how non-separable, conjunctive structure can begin to explain behavior.
Deep reinforcement learning (RL) is a powerful framework for sequential decision-making. Despite many high-profile successes in recent years, deep RL is known to be brittle. An emerging line of research attributes this brittleness to a loss of plasticity in neural networks. However, plasticity loss has so far been understudied in stochastic environments. We study this interplay systematically, training SAC agents on three DeepMind Control tasks augmented with controlled levels of observation and action noise. We find that plasticity-preserving methods increase learning robustness on the two harder tasks, with their advantage growing as observation noise increases, while combining resets with layer normalization underperforms resets alone at every observation noise level tested. Further, stochasticity can accelerate the loss of plasticity in harder environments, whereas in easier settings noise can act as a regularizer that sustains plasticity. Lastly, plasticity loss is most pronounced in the early stages of training. In the most challenging stochastic settings, TD instability rather than gradual plasticity loss becomes the dominant failure mode.
World action models (WAMs) learn latent variables from observations and use these to both predict the evolution of the scene and to generate actions. The prediction has to be of the whole scene, whereas only a portion of the scene causes actions. WAMs fitted to demonstrations do not, in general, guarantee that they will generate actions solely from the latent causes of the demonstrated actions (the action-relevant latent variables), as the model could just as well learn to imitate in this way whilst entangling the action-relevant latent variables with the nuisance variables. We therefore separate out what causes action from what prediction depends on. We prove that under a sparsity constraint the latent variables used to generate actions are an invertible function of the action-relevant latent variables alone and the true structure is identified up to a permutation of the latent variables. This requires neither sufficient temporal dependence nor any assumptions on how the unobserved regimes evolve and holds even when the action-relevant latent variables are statistically dependent on the rest. We therefore propose CWAM (Causal WAM), which incorporates the constraint into a pretrained WAM by adding a sparse routing path to its action head, with the world modeling objective remaining unchanged. Experiments on both synthetic data and real-world robots validate the theory as well as showcase the improved efficiency in closed-loop success.
Starbucks could eventually show you a personalized digital menu when you pull up to the drive-thru. During an earnings call on Thursday, Starbucks CEO Brian…
