Papers

Automating empirical research is a long-standing direction of AI. Recent automatic research (AutoResearch) agents bring this goal within reach, as modern LLMs show the capability to independently implement solutions and learn from the execution outcomes. Behind these gains, post-training (especially RL) plays a central role. In this paper, we identify a fundamental tension when scaling RL for these agents: the two components of every AutoResearch trajectory (agent generation and environment execution) scale in very different manners, since all generation shares compute through batching, while each execution occupies its exclusive sandbox and real machine time. As a result, the environment execution dominates the training cost and becomes the bottleneck as trajectories grow. To resolve this tension, we propose World Model RL (WMRL), which replaces environment execution with a world model to remove this bottleneck. Additionally, the world model can be imperfect, as its rewards are corrupted by bias and noise. Therefore, we further equip WMRL with two mitigations, Online Debiasing and Inverse-Variance Denoising, which offset the bias and suppress the noise respectively. Theoretically, we prove that both mitigations of WMRL strictly improve the convergence guarantee. Empirically, WMRL accelerates training by 3-4x on various tasks at different agent scales, while exceeding the performance of standard RL baselines. Moreover, our post-trained 4B and 9B agents outperform much larger open-weight agents of 48B and 120B on held-out benchmarks. Beyond AutoResearch, WMRL also transfers to post-training embodied VLA policies, which demonstrates the generalizability of our method.

Predictive explanation methods attribute a model output; they do not, by themselves, attribute an intervention effect on the real-world outcome. We introduce the Causal Attribution Score (CAS), a compact score architecture for causal explanation. CAS starts from an identified interventional coalition game, allocates the joint intervention contrast with causal Shapley contributions, and converts those raw outcome-scale effects into Local CAS, Signed Local CAS, and two complementary Global CAS summaries. The innovation is not a new Shapley formula, but a local-to-global causal reporting layer with an explicit intervention target. In the known-truth benchmark, eight repeated primary-interaction simulations (n = 2,200 each, three actions) gave mean Local CAS MAE of 0.107 for coalition-aware CAS, compared with 0.173 for one-at-a-time normalisation and 0.213 for a global normalised absolute ATE vector. The paired advantage over one-at-a-time normalisation increased from -0.003 under additivity to 0.091 under strong interactions. On both empirical DoubleML datasets, 401(k) eligibility/net financial assets (n = 9,915) and Pennsylvania reemployment bonus/unemployment duration (n = 5,099), predictive SHAP/TreeSHAP rankings differed materially from Feature-CAS rankings of treatment-effect modifiers. In Pennsylvania, dep1 (exactly one dependent) moved from predictive global rank 13 to Feature-CAS rank 2 and was the leading local Feature-CAS modifier. These results isolate the added value of separating what predicts the outcome from what explains heterogeneity in an estimated causal effect.

Dance imitation integrates motor planning, sensorimotor integration, and social cognition, offering a sensitive framework to characterize motor behavior in autism. In this work, we explore a computational analysis framework to identify potential biomarkers that allow the design and development of improved medical and human-machine systems. We analyzed 3D motion capture data from autistic and neurotypical adults performing dance imitation under solo and socially-framed duo conditions. Methodologically, using Dynamic Time Warping, we quantified movement consistency and propose the Social Context Sensitivity Index (SCSI) to measure modulation of variability by social framing. These features were then used on a classifier to discriminate subjects into autistic or neurotypical groups. Results show that neurotypical adults exhibited increased movement variability in socially-framed imitation, especially in upper and lower limbs, whereas autistic adults maintained consistent movement across contexts. Classification achieved 79.2% balanced accuracy in distinguishing groups. These findings suggest that social context sensitivity in motor imitation constitutes a robust biomarker of autism-related motor behavior, highlighting the importance of social modulation in motor assessments and informing the development of inclusive human-centric technologies.

Deep learning models are widely used for time series analysis in domains such as healthcare, finance, energy systems, and environmental monitoring. However, these models remain vulnerable to adversarial attacks, where small input perturbations cause severe degradation in predictive performance. Commonly used gradient-based attacks, iterative first-order methods, are computationally burdensome, as they repeatedly backpropagate through the victim model to compute input gradients during a number of iterative refinement steps. We propose a GENerative ADversarial Attack (GENADA) that learns a generative model to produce deceptive perturbations directly in a single forward pass and a procedure to train it. Variants include single-step and iterative generative attack schemes. The validation considers attacks on several neural models and datasets in the time-series domain, a controlled, low-dimensional setting. Empirically, GENADA achieves comparable attack quality to strong baselines while requiring less time to generate perturbations during inference.

Deep graph neural networks(GNNs) suffer from oversmoothing- a progressive collapse of node representation towards a low information subspace as network depth increases because the normalized graph propagation operator is repeatedly applied directly to the hidden representations. In this work we study Householder Graph Neural Network (HouseGNN). Rather than updating the hidden state like standard GCN, HouseGNN uses the aggregated neighbourhood message solely to estimate a reflection direction; the node embedding is then updated by a Householder reflector followed by GroupSort, yielding a piecewise orthogonal layer that preserves Euclidean norm at every node and at every depth. We prove three core properties: (i) every internal layer preserves the node-wise Euclidean norm; (ii) the Householder reflector is scale scale and sign-invariant in the message; and (iii) pairwise distance between nodes can change through mismatch between node-wise orthogonal operators.

We describe a simple rejection-sampling-based algorithm to perform length-squared sampling on an $n \times n$ positive-semidefinite (psd) matrix: that is, to sample a column with probability proportional to its squared $\ell_2$-norm. The algorithm runs in just $O(n)$ expected time, which is significantly sublinear in the input matrix size. The runtime is optimal, even when the input is assumed to be diagonal. Our result has several applications. Length-squared sampling is used by a number of sublinear time algorithms for matrix problems, like low-rank approximation and eigenvalue approximation. Often, it is assumed that the algorithm is given access to the matrix column norms, and thus can perform length-squared sampling efficiently. Our result shows that, at least for psd matrices, we can remove this assumption. We also discuss an application to an asymptotically optimal algorithm for estimating the Frobenius norm of a psd matrix to relative error. Finally, we show that our sampling algorithm yields a very simple sublinear time algorithm for the robust psd low-rank approximation problem introduced by Bakshi et al. (FOCS, 2020), which nearly matches the more complex method developed there.

Organizations decide whom to treat under a budget and want to know what a targeting rule would have earned before deploying it. Off-policy evaluation promises this from logged data, but the deployable rule is a deterministic top-k policy: it removes all averaging over actions, so weak overlap hits the estimate directly. We benchmark six estimators across five datasets and two known-effect sweeps, and validate the mechanisms against a non-simulated paired reference. First, weak overlap is governed by logger-target action alignment, not by logging sharpness alone: what governs support is the logger's probability of the target's actions. Sharpening a logger built from the target's own score barely moves overlap over the tested range; action-level disagreement collapses it. Effective sample size ranks this risk across logging environments, but is weak at ranking candidates within the single log a practitioner holds, and its cut point does not transfer. Second, the optimizer's curse is not fixed by cross-fitting the outcome nuisance. When the rule is fit on the data used to evaluate it, cross-fitting the nuisance alone leaves the reuse bias in place and makes it worse. Honest policy-level splitting avoids the reuse by targeting the learning procedure's value – a change of estimand, not a de-biasing of the full-sample policy. Third, propensity-estimation error is the largest degradation we measure: an out-of-fold estimate hurts IPS more than any other stress we apply, leaves doubly-robust estimation almost unchanged, and can invert the overlap diagnostic itself. Logging is synthesized and propensities floored at 0.02, so every failure occurs with bounded weights; the floor also reduces the two tuned hybrids to their untuned parents, leaving four practically distinct estimators, and all exact-value surfaces are synthetic or semi-synthetic. We release the benchmark; public data only.

Clinical prediction models are often developed as if the outcome of interest were cleanly observed for every patient. This assumption fails when treatment decisions make the clinically relevant outcome permanently unobservable. As a case study of this problem, we consider post-cardiac-arrest neurological prognostication using a cohort of 2,497 patients, including 1,429 patients whose outcomes were rendered indeterminate by treatment decisions. These patients with indeterminate outcomes were reviewed by independent clinical experts, who provided their guesses of counterfactual outcomes about what would have happened to the patients. We refer to these patients as uncertain cases. We also have patients for whom we observe their clinically relevant outcomes; we refer to these patients as certain cases. We propose a framework for evaluating prediction models that explicitly splits the evaluation between certain and uncertain cases. Here, we cannot easily evaluate both types of cases in a uniform manner as the available target labels differ. We then propose a simple prediction model that uses target labels from both certain and uncertain cases in a manner that allows us to trade off between them. Across the proposed neural model and a collection of tabular baselines, models with similar certain-case AUROC can nevertheless differ substantially in both certain-case Brier score and their probability estimates for uncertain cases. Improving alignment with target labels of uncertain cases for our proposed model generally comes at the cost of worse accuracy on certain cases, highlighting an explicit tradeoff that standard evaluation conceals. These results show that when treatment decisions determine whether clinically meaningful outcomes remain observable, conventional evaluation metrics can miss important failure modes in the very patients for whom prognostic support matters most.

Is it possible to find an objective classification criterion that organizes the complexity of gust-induced loads across many flight conditions? And one that remains as interpretable as a labelling based on coarse parameters, such as the flight attitude? Our approach encodes a large number of experimental observations through a machine-learned representation and applies a summarization procedure to select a minimal subset of highly significant exemplars. The exemplars provide a similarity-based objective classification criterion of all the observations, they can be more conveniently inspected by experts and can become subject of more refined experiments. We demonstrate the approach on a database of 3480 pressure-load measurements induced by random gusts on a flying-wing model across six flight attitudes. We find nine fundamental response types that recur across multiple attitudes; analysis of a type's transient response enables physical intuition into the underlying fluid mechanics.

Trained transformer models develop privileged bases: coordinate axes whose statistics differ from the rest of the residual stream. But what kind of direction does such a basis select? We investigate the prediction direction, the unembedding direction of the token a model currently predicts, and find that it functions as a content-defined privileged anchor. Measured with respect to this anchor, residual-stream variation is geometrically and behaviorally stratified by proximity to the prediction. The stratification holds in all eighteen models tested (dense and mixture-of-experts, 7B-120B, base and instruction-tuned). A narrow, scale-invariant prediction interface concentrates readout-relevant structure, while the vast prediction-distal complement expands with model scale. Because the prediction direction sits nearly orthogonal to the principal variance axes, variance-based analyses recover this organization only partly, and the shortfall grows with prompt heterogeneity. Anchoring reveals a steep geometric gradient: prediction-proximal regions are highly structured and cluster related prompts, while the complement is flatter and anti-discriminates among prompt groups. The interface is a narrow slice but functionally decisive. Disrupting the variance directions closest to the prediction causes immediate divergence and frequent task-frame shifts; disrupting the next level down delays divergence and preserves framing. The complement is weakly readout-aligned per direction yet causally and temporally load-bearing, and behavior is driven by direction rather than magnitude. These results establish the prediction direction as a privileged anchor distinct from previously described coordinate axes, and give a geometric account of how high-dimensional computation coexists with linear readout.

Sleep stage classification is important for the diagnosis and management of sleep disorders, yet most automatic staging studies evaluate models against a single reference hypnogram despite known inter-scorer variability. This study investigates whether multi-scored datasets can be used to construct more reliable reference labels from the collective behavior of multiple experts. We use the publicly available DOD-H and DOD-O datasets. EEG (C3-M2) and chin EMG signals were segmented into 30-s epochs, and 30 features were extracted from each modality, yielding 60 features for EEG+EMG. We propose a learning-based hypnogram (LBH) that models the stage-specific behavior of each scorer using confusion matrices derived from machine-learning models. After column normalization, these matrices estimate the probability of each true sleep stage given each scorer's label; probabilities are aggregated across scorers to assign the final label for each epoch. LBH was evaluated with random forest, support vector machine, and multilayer perceptron classifiers under EEG-only and EEG+EMG settings, and compared with the dataset hypnogram (DH) and best-scorer hypnogram (BSH). LBH consistently improved overall performance. The best results were obtained with random forest and EEG+EMG, reaching 86.07% accuracy, 85.46% precision, and 85.29% F1-score on DOD-H, and 86.04% accuracy, 85.21% precision, and 84.70% F1-score on DOD-O. These findings suggest that personalized scorer modeling can improve reference hypnogram construction without discarding information from individual experts.

An unconditional risk bound on automated decisions can be satisfied without automating anything, since a selector that never acts drives the bound to zero. We show this is structural: any risk certificate is defined over a decision contract, the inputs a system acts on plus the semantic relation under which an output counts correct, and weakening either hides base-classifier error. We develop a decision-contract theory: an error-conservation law showing error is only reassigned among harmful automation, human deferral, and semantic masking; a label-free singleton capacity certifying structural incapacity, with a risk-feasible refinement separating recoverable threshold misalignment from risk-constrained incapacity; and a non-degenerate actionability certificate excluding all-abstain solutions by construction. We instantiate this on ATT&CK-aligned alert triage for LLM-based intrusion detection, the setting that exposed the vacuity failure. Across 3 IDS datasets, 6 LLMs, and 4 error-rate thresholds, empirical false-attribution risk stays at or below target in 90.3% of configurations, with 83.4% mean correct automation. The capacity diagnostic explains every low-utility configuration; its refinement separates genuine misalignment from risk-constrained incapacity, confirmed by an exhibited alternative threshold; a training-stability re-run finds no confirmed structural-incapacity instance; and real fine-grained attack-subtype labels confirm the coarsening-transfer identity under a genuine many-to-one map, with small but non-zero masking mass.

Neural combinatorial optimization (NCO) relies on parallel solution sampling for training, yet existing methods fail to fully exploit the rich information latent in a co-sampled solution group. Preference-optimization methods anchor on the single best solution and discard fine-grained quality and structural signal from all other peers-a failure we term gradient signal polarization. Mean-based baselines instead weight peers uniformly, so structurally near-identical peers flood the baseline with redundant information and keep gradient variance high-a failure we term baseline redundancy. We propose SSPO (Structure-Aware Similarity-Weighted Preference Optimization), which scores all $B$ sampled solutions jointly through a dissimilarity-weighted leave-one-out baseline: structurally distinct peers receive higher weight, resolving both failures in a single mechanism. The baseline uses zero-parameter, problem-adaptive solution embeddings built from the encoder's existing node representations. Experiments on TSP, EFL, and JSP benchmarks show consistent gains over prior best-anchor and uniform-weight baselines. A direct comparison against uniform RLOO on TSP and EFL confirms that structure-aware weighting is the primary driver of improvement. The SSPO-trained EFL policy has been deployed in a production facility-location system at JD$\mathord{.}$com, confirming practical viability at scale.

Deep learning detectors for anomalies in dynamic graphs have reached strong accuracy, yet they remain opaque: when an edge is flagged, the analyst receives a score but no reason. This opacity is untenable in the cooperative, regulated information systems where such detectors are deployed, where automated decisions must be auditable and trustworthy. We address this gap for AddGraph, the foundational GCN+GRU framework for edge-level anomaly detection in dynamic graphs, which to our knowledge has never been equipped with any form of explainability. We present a strictly post-hoc explainability framework, X-AddGraph, built on a Dual Spatial-Temporal Attribution (DSTA) mechanism whose three components are each aligned with one of AddGraph's architectural modules: a gradient-based relevance attribution over the current adjacency structure (spatial), a direct reading of the contextual attention weights already computed during inference (short-term temporal, at zero additional cost), and a gradient rollback through the recurrent hidden states (long-term temporal). Because the detector is frozen, detection performance is preserved exactly (Delta AUC = 0, verified empirically to ten decimal places). On the UCI Message benchmark, our trained AddGraph baseline reaches an average per-snapshot AUC of 0.8705, exceeding the originally published result; X-AddGraph reproduces every score identically while adding explanations where none existed. Evaluated across four edge populations - confident true positives, low-confidence true positives, false positives, and random samples - the long-term attribution identifies historical snapshots carrying significantly more counterfactual signal than random selection (0.127 vs. 0.074), a capability that no spatially-blind explainer can provide. We release our implementation for full reproducibility.

We formulate generative modeling as a path integral in which flow-based, diffusion-based, variational, and adversarial models arise as different evaluation principles for a single master action. Its Martin-Siggia-Rose-Janssen-de~Dominicis (MSRJD) form separates free from interacting probability flows and opens them to diagrammatic perturbation theory. The expansion yields a one-loop correction to deterministic samplers at no stochastic-sampling cost, which we validate on solvable and nonlinear drifts, where it reduces a 53 % tree-level error to 1.6 %. Imperfect learned scores enter as insertions and yield a response-weighted score-matching objective, and symmetry-equivariant drift design becomes an operator expansion with EFT power counting.

Multi-AUV ad-hoc network-based target tracking requires networked autonomous underwater vehicles (AUVs) to cooperatively track maneuvering targets under constrained acoustic communication, dynamic topology, and uncertain ocean disturbances. Although multi-agent reinforcement learning (MARL) enables decentralized coordination through centralized training, existing methods suffer from high-dimensional joint state-action modeling, noise-sensitive policy generation, leading to unstable training and degraded tracking. To address these issues, we propose VGG-MADiffRL, a value-gradient-guided multi-agent diffusion RL algorithm, and MDCA, a diffusion?based hierarchical control architecture. Leveraging underwater mission characteristics, we model sonar detection mechanisms and ocean current disturbances, formulating cooperative tracking for multi-AUV ad-hoc networks as an MDP. The proposed MDCA constitutes a three-tier closed-loop control framework: a global intelligent control layer, a local online training layer, and a physical action execution layer. This structure enables synergistic optimization across task allocation, local decision processes, and execution feedback. Within MDCA, the local online training layer is the policy learning framework; VGG-MADiffRL builds on diffusion policies and incorporates value gradients to guide action generation in the reverse denoising process, steering the generated actions towards higher expected returns. It employs twin value networks with joint optimization and soft target updates to mitigate overestimation and training oscillations, promoting more stable convergence. Experimental results show that VGG-MADiffRL consistently achieves faster convergence, higher tracking accuracy, and smoother training dynamics in cooperative tracking scenarios, validating its effectiveness and practical engineering value in dynamic underwater settings.

Transformers owe much of their strong long-context retrieval capability to a token-level memory that grows with context length. This flexibility, however, incurs a quadratic computation complexity during training and a key–value cache that grows linearly during autoregressive inference. Recurrent alternatives offer efficient decoding by compressing the entire history into a fixed-size state, but often underperform on recall-intensive tasks since earlier associations usually get overwritten by subsequent updates, and only the most recent contextual information is retained. In this paper, we introduce Memory-Anchor Routing across Context History (MARCH), a network architecture that effectively scales state-space models beyond a fixed-size dimension, while maintaining computational efficiency over long-sequences. MARCH periodically caches cumulative recurrent-state checkpoints as state anchors and associates each anchor with a compact, content-conditioned anchor key. This lets MARCH maintain a memory bank, which can grow as context length increases, providing a controllable trade-off between historical resolution and memory cost. At each token, MARCH produces an anchor query to attend all causally available state anchors, and the output is calculated as an attention-style aggregation over all historical anchors along the current state. We show that after standard pretraining, MARCH consistently outperforms multiple linear attention variants across commonsense reasoning, LongBench, and in-context retrieval. These results demonstrate that content-routed state caching substantially strengthens recurrent long-range memory while preserving its native computation path.

This study focuses on developing an AI-supported prototype for multiperspective interest rate forecasting that combines classical econometric models with modern artificial intel-ligence methods. Tested in a major European bank, the system enables more precise and flexible prediction of interest rate developments, supporting strategic decision-making in Asset-Liability Management (ALM). It integrates topic modeling, sentiment analysis, econometric forecasting, and market-based analyses within an interactive platform. Leveraging AI to analyze large volumes of financial documents and market data enables the identification of monetary policy trends and sentiment signals at an early stage. The core econometric model is a Bayesian vector autoregression (BVAR) that enables simulation-based scenario analyses to evaluate economic developments from multiple perspectives. The system's innovation lies in its integration of several forecasting approaches that consolidate previously separate information sources and present them transparently and interpretably. Financial analysts and risk managers thus gain a better basis for making decisions, allowing them to assess interest rate risks more accurately and manage market movements more proactively. While the prototype demonstrates how AI can transform interest rate management in banking, further development is required to optimize real-time data integration and regulatory compliance. Even at this stage, the study shows that multi-perspective, AI-driven forecasting provides substantial added value for banks by increasing transparency, strengthening evidence-based decision-making, and improving risk management.

Two free satellite signals carry real information about glacial-lake outburst risk in the Nepal Himalaya: radar interferometry sees a moraine dam slowly sagging, and satellite weather marks the weeks when a primed lake is under stress. A companion feasibility study found that deformation indicates which lake is destabilizing and weather indicates when it is at risk, but proposed no predictive model. To address this gap, we propose and evaluate models that predict which site is susceptible and when a trigger arrives. We test three related hazards on free data alone: large moraine- and ice-dammed bursts, rainfall-triggered landslides, and smaller floods from ponds on and around a glacier. Each hazard gets two questions, never blended. Using 589 dated outbursts from HMAGLOFDB and several thousand catalogued landslides, we match each event against similar but unfailed sites, and hold every model to a strong simple baseline under spatial cross-validation that withholds whole map tiles, so no model succeeds by recognising a trained-on neighbourhood. Antecedent weather times the trigger at ROC 0.73 for big bursts, 0.83 for landslides, and 0.82 for small floods. Terrain ranks susceptibility only in part: scored naively it appears near 0.9, largely because catalogued failures cluster in wetter ranges; matched against comparable nearby sites the honest figures are 0.76, 0.71, and 0.54 (no better than chance). The burst signal holds within single regions, reaching 0.89 in Nepal alone. Five deep-learning models do not decisively beat a simple gradient-boosted baseline. Three score marginally higher on landslides, a hint too small to confirm. For the lake hazards the baseline wins outright, reproduced by a three-rule decision tree on ruggedness and monsoon rainfall. We close with a ranked Nepal watchlist, a prioritisation aid, not a prediction, and note where free data reaches its limits.

Large language models (LLMs) have achieved remarkable breakthroughs across various applications. However, their architectures remain inefficient in pretraining due to two main limitations: (i) self-attention lacks an explicit inductive bias for locality, leading to redundant modeling of sequence-internal local information; (ii) mixture-of-experts (MoE) implicitly couples knowledge storage with computational pathways, hindering flexible access to sequence-external global knowledge. To overcome these limitations, we propose LoKiFormer, a novel LLM architecture that augments the standard decoder with two dedicated modules: 1) Local Fusion Attention (LFA), which incorporates a convolutional fusion to attention, explicitly capturing local patterns and allowing the attention to operate on more informative representations; 2) Knowledge Memory Module (KMM), which introduces a parametric key-value memory that explicitly stores global knowledge in addressable slots, decoupling storage from computation and enabling direct knowledge retrieval. Together, these modules enable LoKiFormer to achieve more efficient and effective integration of information at both levels. Experimental results show that LoKiFormer converges 1.33x faster in pre-training than baseline models, underscoring its superiority over existing LLM architectures.

Recent work on distillation transfers the capabilities of large models to smaller ones often by updating the latter's parameters, through teacher forcing, on-policy distillation, and related training-time methods. In this paper, we ask whether such transfer can instead occur at test time. We study strong-to-weak scaffolding: whether a stronger builder model can construct inference-time harnesses that help a weaker target model solve tasks more reliably without any parameter updates. Using four representative Theory-of-Mind benchmarks, each builder model uses 5% of the data as a validation set to iteratively refine its harness over multiple rounds, after which the finalized harness is evaluated on the full test set. Empirically, this form of test-time capability transfer is highly effective, nearly doubling average target-model performance from 0.49 to 0.91. Our analysis shows that the gains come primarily from offloading unstable model reasoning into deterministic code, benchmark-specific routing, and strict answer-format enforcement, rather than from encouraging the target model to reason more extensively or sample more broadly. We further find that builder-model reasoning effort improves harness quality monotonically, platform effects are modest relative to the builder model's own capability, and weaker target models receive the largest gains. These results suggest that inference-time harness design is an important complement to conventional training-time distillation, enabling strong models to transfer cognitive structure to weaker models without retraining.

Safe offline RL typically assumes access to dense per-step cost annotations, but in practice supervisors provide only trajectory-level stop-feedback: a binary signal at the first unsafe transition, with no per-step attribution. We frame this as a temporal credit assignment problem and propose the Redistribution-based Cost Inference (RCI) framework, which converts sparse stop-feedback into dense per-step costs via return decomposition, then trains a constrained offline policy on the augmented dataset. We show that return-equivalent redistribution preserves the feasible policy set and the optimal Lagrangian in a CMDP, establishing that the transformation is lossless in theory while yielding better-conditioned cost critic learning in practice. Experiments on highway driving and robotic manipulation demonstrate substantially lower violation rates than sparse and classifier-based baselines, with robustness to heterogeneous dataset compositions and label noise.

We present a formal process to enable non-experts to instantiate and iterate on human-aligned reward functions, i.e. reward functions that adhere to a given preference ordering over trajectories. Given a task described in natural language, our process produces a linear reward function in three steps: distill the task's objectives into a set of fundamental objectives and derive measurable outcome variables that capture those fundamental objectives, select a causally representative subset of outcome variables as the reward terms, and fit weights to those reward terms via preference elicitation. Our contributions describe the first step and formalize the latter two steps. The first is a guided workflow for deriving outcome variables. The second is a reduction of reward term selection to minimum-cost partial cover on a causal DAG, solved in polynomial time via max-flow. The third is a geometric framing of weight fitting as a convex feasibility problem iteratively narrowed by preference queries, solved by existing separation oracle methods. To the best of our knowledge, this is the first reward-design method that maintains a deterministically conflict-free feasible weight region, narrowed to a desired tolerance via a separation oracle with O(n log \kappa) preference queries.

Global weather reanalyses and forecasts resolve the evolving atmospheric state on coarse grids, but site-specific applications require predictions at arbitrary locations where near-surface conditions also depend on unresolved terrain and land-surface properties. Existing probabilistic downscalers address this gap using hand-crafted topographic and surface descriptors. We ask instead whether Earth observation foundation models can provide transferable subgrid surface representations for probabilistic weather downscaling. We augment a convolutional conditional neural process (ConvCNP) that downscales coarse ERA5 reanalysis fields at ~25 km resolution with a learned local surface descriptor, obtained by compressing a patch of TESSERA embeddings at 10 m resolution. Although these embeddings summarize annual surface conditions, they improve downscaling by encoding persistent surface properties that capture a location's departure from the coarse-grid atmospheric state. Across five climatically diverse regions, the embedding improves point and probabilistic skill at stations held out in both space and time, overall improving CRPS skill by 11.5% for 2 m temperature and 6.2% for 10 m wind speed relative to a topography-only ConvCNP baseline. A hand-crafted descriptor incorporating richer surface information than topography alone captures comparable persistent subgrid signal but yields far smaller predictive gains than the learned embedding representation. These improvements persist when forecasts from the Aurora AI model replace ERA5 reanalysis fields and when predicting at newly deployed weather station networks. To our knowledge, this is the first evidence that long-timescale Earth observation embeddings can support short-timescale weather downscaling where subgrid departures are systematically structured by persistent surface properties.

Financial forecasting models are typically developed in full precision, yet production deployment often requires low-precision inference to reduce memory and computational cost. Post-training quantization (PTQ) enables such deployment without retraining. However, reliable activation quantization requires calibration: activation ranges are estimated from historical data before deployment and then remain fixed during future inference. The importance of this deployment choice for financial forecasting remains poorly understood. We present a systematic study of activation calibration for PTQ in cross-sectional volatility forecasting on the S&P 500. Our evaluation covers seven representative neural architectures, eight walk-forward test years (2018-2025), and 560 trained models. We find that activation calibration has little effect at 8 bits but becomes the primary determinant of predictive performance at 4 bits. Under default absolute-maximum (abs-max) calibration, static 4-bit quantization of both weights and activations removes 11-62% of the full-precision mean information coefficient in affected architectures. Replacing abs-max with percentile calibration recovers 53-94% of this degradation in the four most affected architectures. The preferred activation range also varies across market periods. Narrow ranges improve resolution under typical market conditions but lose part of their advantage when test-period market dispersion exceeds the calibration history. These findings show that activation calibration is a first-class deployment decision for reliable 4-bit PTQ in financial forecasting. When substantial degradation remains, 8-bit activations or weight-only 4-bit quantization provide more robust deployment choices.

Multi-agent reinforcement learning for human-AI interaction typically relies on a single large language model to simulate user behavior. We show that this approach systematically fails to generalize, and trace the failure to simulator collapse: because the simulator LLM is mode-collapsed, an LLM policy trained against it overfits to narrow strategies that exploit the simulator's dominant mode, and such a policy transfers poorly to unseen simulators and real users. We formalize this collapse theoretically and propose two complementary solutions, one at inference time and one at training time. The inference-time solution, Verbalized Sampling, broadens the simulator's behavior by sampling from a verbalized response distribution, reducing mode collapse. The training-time solution, Co-Training, jointly optimizes the policy against a population of trainable simulators, preventing it from overfitting to any single simulator's mode. We validate both solutions on three multi-turn benchmarks: Persuasion for Good, $\tau^2$-bench, and CooperBench. Verbalized Sampling improves held-out success by up to 9% over single-simulator RL, and Co-Training pushes gains further to 14%; the human study shows similar gain on real users. Both solutions preserve the policy diversity that collapses under single-simulator RL. To support further work in this direction, we release SCOPE, an open-source framework for Population Co-Training multi-agent RL. More broadly, our results suggest that the diversity of the training environment, not only the policy, is critical to the generalization of multi-turn RL to real-world deployment.

Financial volatility is regime dependent, yet incorporating regime information into neural networks can also destabilize training. This paper asks where such information should enter a neural cross-sectional volatility forecasting model. We study five-day realized-volatility forecasts for 1,027 U.S. equities using a rolling walk-forward evaluation framework in which information, model capacity, hyperparameter tuning, and random seeds are matched across architectures. We propose RG-ResMoE, a regime-gated residual mixture-of-experts architecture in which regime information is used only for expert routing rather than for direct forecasting. The base predictor models volatility from stock features, while a gating network uses regime state variables to route residual corrections. RG-ResMoE consistently outperforms a capacity-matched MLP in both forecasting accuracy and training stability in the main U.S. study. Similar gains are observed on an independent Japanese panel. The integration pathway is decisive: appending the same regime variables directly to the forecasting input degrades both predictive performance and training stability, whereas restricting them to the routing gate improves accuracy and Value-at-Risk calibration. Hard routing consistently underperforms soft routing. The results suggest that, in compact neural volatility forecasting models, the primary value of mixture-of-experts models lies less in increasing model capacity than in controlling how nonstationary regime information influences prediction.

We study adversarial combinatorial bandits with $m$-set actions, where at each round the learner selects $m$ out of $d$ items and observes only the aggregate loss of the selected items. The resulting action set contains $K=\binom{d}{m}$ elements and can therefore be exponentially large. Nevertheless, the loss of every action is determined by the same $d$-dimensional vector of item losses. We propose a computationally efficient algorithm that exploits this structure without explicitly enumerating the action set. Against adaptive non-anticipating adversaries, it guarantees, with probability at least $1-\delta$, regret against the best fixed action of \[ R_T = O\left(\sqrt{dT\log(K/\delta)}\right). \] This matches the high-probability regret bound of the finite-action EXP3-KW algorithm of Zimmert and Lattimore, whose direct implementation may require exponential space. Our algorithm instead represents each sampling distribution with $d$ parameters and runs in polynomial time without enumerating the action set. Thus, it resolves the open problem posed by Maiti et al. We complement this upper bound with a matching high-probability lower bound. For all sufficiently small $\delta$, every randomized policy admits a deterministic adaptive non-anticipating adversary for which, with probability at least $\delta$, \[ R_T = \Omega\left(\sqrt{dT\log(K/\delta)}\right). \] Thus, the rate is minimax optimal up to universal constants in this regime. In particular, setting $m=1$ proves that the $\log K$ for ordinary $K$-armed bandits against adaptive non-anticipating adversaries is unavoidable, closing the remaining $\sqrt{\log K}$ gap between confidence-tuned upper and lower bounds left by Gerchinovitz and Lattimore.

Treating patients with combinations of drugs reduces the risk of resistance to any individual drug. Finding effective combinations is difficult because the large search space makes combinatorial screens prohibitively expensive, time consuming, and often technically infeasible. Predictive models can fill this gap, yet existing methods typically require molecular profiling of each sample and per-cohort training, limiting their applicability when time and tissue are scarce. To address this challenge, we introduce ScreenShot, a hierarchical transformer pretrained on 40 drug screening datasets covering 3,700 drugs and 6,000 biological samples, whose architecture mirrors the nested structure of screening data. Given a few-shot context of observations from a new patient, ScreenShot predicts the response of the sample to combination therapies through in-context learning, operating directly on functional measurements with no fine-tuning and no molecular profiling. On four held-out datasets, ScreenShot outperforms all baselines in both prediction accuracy and identification of selectively effective treatments. ScreenShot's internal representations are directly useful for experimental design: we use them to drive a weighted k-means++ active learning strategy that selects which experiments to run, achieving the same hit detection as uniform screening with a third of the budget. Source code and interactive dashboard: https://github.com/tansey-lab/screenshot.

Recent research in machine and deep learning has shown the potential of learningbased motion planning approaches to improve the driving behavior of automated vehicles, especially in complex environments. However, their complex nature and lack of transparency can hinder explainability and trustworthiness and complicate safety assurance. Motivated by these challenges, we propose a hybrid planning architecture that combines the advantages of machine learning with the verifiability and the determinism of classical approaches. Specifically, we developed a deep neural network to interpret complex traffic scenes and propose driving behavior, while an optimization-based supervision layer validates this proposal and enforces explicit drivability and safety constraints. We evaluate the learned planner's driving behavior in open-loop studies on real-world urban data, discuss system integration aspects for stable closed-loop operation, and report results from real-world deployment on our research vehicle karl..