In-context learners like TabPFN are promising for biomolecule efficacy prediction, where established molecular feature sets and relevant experimental results can serve as powerful contextual examples. However, their performance is highly sensitive to the provided context, making strategies like post-hoc ensembling of models trained on different data subsets a viable approach. An open question is how to select the best models for the ensemble without access to ground truth labels. In this study, we investigate an uncertainty-guided strategy for model selection. We demonstrate on an siRNA knockdown efficacy task that a TabPFN model using straightforward sequence-based features can surpass specialized state-of-the-art predictors. We also show that the model's predicted inter-quantile range (IQR), a measure of its uncertainty, has a negative correlation with true prediction error. We developed the OligoICP method, which selects and averages an ensemble of models with the lowest mean IQR for siRNA efficacy prediction, achieving superior performance compared to naive ensembling or using a single model trained on all available data. This finding highlights model uncertainty as a powerful, label-free heuristic for optimizing biomolecule efficacy predictions.

LLM web agents now browse and take actions on the open web, yet current agent evaluations are constrained to sandboxed environments or artificial tasks. We introduce BrowserArena, a live open-web agent evaluation platform that collects user-submitted tasks, runs Arena-style head-to-head comparisons, and uses step-level human feedback to surface failure modes. Collecting and analyzing step-level annotations on the agent traces, we identify three consistent failure modes: captcha resolution, pop-up banner removal, and direct navigation to URLs. By constructing targeted datasets to further study these tasks, we discover variations in how different language models navigate these failure modes. We find, for example, that o4-mini deploys a wider variety of strategies to circumvent captcha resolution than other models and DeepSeek-R1 consistently misleads users about pop-up banner closure. Our findings surface both the diversity and brittleness of current web agents. More broadly, our benchmarking methodology provides an approach to evaluating and understanding web agent failure modes at scale.

Reconstructing high-resolution rainfall fields is essential for flood forecasting, hydrological modeling, and climate analysis. However, existing spatial interpolation methods-whether based on automatic weather station (AWS) measurements or enhanced with satellite/radar observations often over-smooth critical structures, failing to capture sharp transitions and localized extremes. We introduce RainSeer, a structure-aware reconstruction framework that reinterprets radar reflectivity as a physically grounded structural prior-capturing when, where, and how rain develops. This shift, however, introduces two fundamental challenges: (i) translating high-resolution volumetric radar fields into sparse point-wise rainfall observations, and (ii) bridging the physical disconnect between aloft hydro-meteors and ground-level precipitation. RainSeer addresses these through a physics-informed two-stage architecture: a Structure-to-Point Mapper performs spatial alignment by projecting mesoscale radar structures into localized ground-level rainfall, through a bidirectional mapping, and a Geo-Aware Rain Decoder captures the semantic transformation of hydro-meteors through descent, melting, and evaporation via a causal spatiotemporal attention mechanism. We evaluate RainSeer on two public datasets-RAIN-F (Korea, 2017-2019) and MeteoNet (France, 2016-2018)-and observe consistent improvements over state-of-the-art baselines, reducing MAE by over 13.31% and significantly enhancing structural fidelity in reconstructed rainfall fields.

OpenAI's o3-preview reasoning model exceeded human accuracy on the ARC-AGI benchmark, but does that mean state-of-the-art models recognize and reason with the abstractions that the task creators intended? We investigate models' abstraction abilities on ConceptARC. We evaluate models under settings that vary the input modality (textual vs. visual), whether the model is permitted to use external Python tools, and, for reasoning models, the amount of reasoning effort. In addition to measuring output accuracy, we perform fine-grained evaluation of the natural-language rules that models generate to explain their solutions. This dual evaluation lets us assess whether models solve tasks using the abstractions ConceptARC was designed to elicit, rather than relying on surface-level patterns. Our results show that, while some models using text-based representations match human output accuracy, the best models' rules are often based on surface-level ``shortcuts'' and capture intended abstractions far less often than humans. Thus their capabilities for general abstract reasoning may be overestimated by evaluations based on accuracy alone. In the visual modality, AI models' output accuracy drops sharply, yet our rule-level analysis reveals that models might be underestimated, as they still exhibit a substantial share of rules that capture intended abstractions, but are often unable to correctly apply these rules. In short, our results show that models still lag humans in abstract reasoning, and that using accuracy alone to evaluate abstract reasoning on ARC-like tasks may overestimate abstract-reasoning capabilities in textual modalities and underestimate it in visual modalities. We believe that our evaluation framework offers a more faithful picture of multimodal models' abstract reasoning abilities and a more principled way to track progress toward human-like, abstraction-centered intelligence.

Multimodal reasoning models have recently shown promise on challenging domains such as olympiad-level geometry, yet their evaluation remains dominated by aggregate accuracy, a single score that obscures where and how models are improving. We introduce MathLens, a benchmark designed to disentangle the subskills of multimodal reasoning while preserving the complexity of textbook-style geometry problems. The benchmark separates performance into three components: Perception: extracting information from raw inputs, Reasoning: operating on available information, and Integration: selecting relevant perceptual evidence and applying it within reasoning. To support each test, we provide annotations: visual diagrams, textual descriptions to evaluate reasoning in isolation, controlled questions that require both modalities, and probes for fine-grained perceptual skills, all derived from symbolic specifications of the problems to ensure consistency and robustness. Our analysis reveals that different training approaches have uneven effects: First, reinforcement learning chiefly strengthens perception, especially when supported by textual supervision, while textual SFT indirectly improves perception through reflective reasoning. Second, reasoning improves only in tandem with perception. Third, integration remains the weakest capacity, with residual errors concentrated there once other skills advance. Finally, robustness diverges: RL improves consistency under diagram variation, whereas multimodal SFT reduces it through overfitting. We will release all data and experimental logs.

Time series forecasting is central to decision-making in domains as diverse as energy, finance, climate, and public health. In practice, forecasters face thousands of short, noisy series that vary in frequency, quality, and horizon, where the dominant cost lies not in model fitting, but in the labor-intensive preprocessing, validation, and ensembling required to obtain reliable predictions. Prevailing statistical and deep learning models are tailored to specific datasets or domains and generalize poorly. A general, domain-agnostic framework that minimizes human intervention is urgently in demand. In this paper, we introduce TimeSeriesScientist (TSci), the first LLM-driven agentic framework for general time series forecasting. The framework comprises four specialized agents: Curator performs LLM-guided diagnostics augmented by external tools that reason over data statistics to choose targeted preprocessing; Planner narrows the hypothesis space of model choice by leveraging multi-modal diagnostics and self-planning over the input; Forecaster performs model fitting and validation and, based on the results, adaptively selects the best model configuration as well as ensemble strategy to make final predictions; and Reporter synthesizes the whole process into a comprehensive, transparent report. With transparent natural-language rationales and comprehensive reports, TSci transforms the forecasting workflow into a white-box system that is both interpretable and extensible across tasks. Empirical results on eight established benchmarks demonstrate that TSci consistently outperforms both statistical and LLM-based baselines, reducing forecast error by an average of 10.4% and 38.2%, respectively. Moreover, TSci produces a clear and rigorous report that makes the forecasting workflow more transparent and interpretable.

As large language models (LLMs) advance, ensuring AI safety and alignment is paramount. One popular approach is prompt guards, lightweight mechanisms designed to filter malicious queries while being easy to implement and update. In this work, we introduce a new attack that circumvents such prompt guards, highlighting their limitations. Our method consistently jailbreaks production models while maintaining response quality, even under the highly protected chat interfaces of Google Gemini (2.5 Flash/Pro), DeepSeek Chat (DeepThink), Grok (3), and Mistral Le Chat (Magistral). The attack exploits a resource asymmetry between the prompt guard and the main LLM, encoding a jailbreak prompt that lightweight guards cannot decode but the main model can. This reveals an attack surface inherent to lightweight prompt guards in modern LLM architectures and underscores the need to shift defenses from blocking malicious inputs to preventing malicious outputs. We additionally identify other critical alignment issues, such as copyrighted data extraction, training data extraction, and malicious response leakage during thinking.

Differential privacy (DP) provides strong protection for sensitive data, but often reduces model performance and fairness, especially for underrepresented groups. One major reason is gradient clipping in DP-SGD, which can disproportionately suppress learning signals for minority subpopulations. Although adaptive clipping can enhance utility, it still relies on uniform hard clipping, which may restrict fairness. To address this, we introduce SoftAdaClip, a differentially private training method that replaces hard clipping with a smooth, tanh-based transformation to preserve relative gradient magnitudes while bounding sensitivity. We evaluate SoftAdaClip on various datasets, including MIMIC-III (clinical text), GOSSIS-eICU (structured healthcare), and Adult Income (tabular data). Our results show that SoftAdaClip reduces subgroup disparities by up to 87% compared to DP-SGD and up to 48% compared to Adaptive-DPSGD, and these reductions in subgroup disparities are statistically significant. These findings underscore the importance of integrating smooth transformations with adaptive mechanisms to achieve fair and private model training.

Foundation Models are designed to serve as versatile embedding machines, with strong zero shot capabilities and superior generalization performance when fine-tuned on diverse downstream tasks. While this is largely true for language and vision foundation models, we argue that the inherent diversity of time series data makes them less suited for building effective foundation models. We demonstrate this using forecasting as our downstream task. We show that the zero-shot capabilities of a time series foundation model are significantly influenced and tied to the specific domains it has been pretrained on. Furthermore, when applied to unseen real-world time series data, fine-tuned foundation models do not consistently yield substantially better results, relative to their increased parameter count and memory footprint, than smaller, dedicated models tailored to the specific forecasting task at hand.

Neural network approaches that parameterize value functions have succeeded in approximating high-dimensional optimal feedback controllers when the Hamiltonian admits explicit formulas. However, many practical problems, such as the space shuttle reentry problem and bicycle dynamics, among others, may involve implicit Hamiltonians that do not admit explicit formulas, limiting the applicability of existing methods. Rather than directly parameterizing controls, which does not leverage the Hamiltonian's underlying structure, we propose an end-to-end implicit deep learning approach that directly parameterizes the value function to learn optimal control laws. Our method enforces physical principles by ensuring trained networks adhere to the control laws by exploiting the fundamental relationship between the optimal control and the value function's gradient; this is a direct consequence of the connection between Pontryagin's Maximum Principle and dynamic programming. Using Jacobian-Free Backpropagation (JFB), we achieve efficient training despite temporal coupling in trajectory optimization. We show that JFB produces descent directions for the optimal control objective and experimentally demonstrate that our approach effectively learns high-dimensional feedback controllers across multiple scenarios involving implicit Hamiltonians, which existing methods cannot address.