Open-ended scientific discovery with large language models (LLMs) increasingly operates as a long-horizon loop of hypothesis search and verification, where a reward signal guides which hypotheses to test next. A notable recent example is AutoDiscovery, which uses "Bayesian surprise" - the belief shift an LLM undergoes after observing evidence for a hypothesis - as both a discovery metric and a reward for search. We first observe that AutoDiscovery treats surprisal as a static quantity, while surprisal in human reasoning is non-stationary - it is defined relative to beliefs that evolve with experience, a prerequisite for continual scientific discovery. We address this mismatch with evidence-informed LLM beliefs: priors updated with evidence from previous hypotheses to compute non-stationary surprisal for new hypotheses. We compare in-context belief-updating mechanisms and find that embedding-based retrieval-augmented generation over prior discoveries best anticipates eventual posteriors, identifying 37.5% of static surprisals as spurious. We then modify search to avoid these spurious rewards and prioritize hypotheses that remain surprising under non-stationary beliefs. Concretely, we introduce two complementary changes to the original search procedure: belief-update filtering and diversity maximization. Across five discovery domains, our method increases accumulated non-stationary surprisal by 30.62% on average compared to the original search procedure, demonstrating that continual scientific discovery with LLMs requires not only better belief measurement but also search procedures that avoid redundancy and encourage diversity. less

Large language models (LLMs) are increasingly used to take actions in the real world and support human decision-making, yet most agents rely on parametric knowledge, fixed post-training data, retrieval, or search. This paradigm breaks down in novel domains and for sophisticated queries that cannot be answered from prior knowledge alone. Knowing the laws of physics, for instance, does not by itself enable LLMs to answer queries or complete long-horizon tasks in a complex physical system. To address this, we introduce Hierarchical Experimentalist Agents (HExA), an in-context self-improvement framework to learn from active experimentation. HExA iteratively designs and refines query-relevant experiments, learns a reusable library of composable skills from experience, and integrates experimental evidence to answer queries or take actions. HExA is training-free, compatible with any black-box model, and does not require external supervision, oracles, or offline data. To evaluate active experimentation, we introduce Interphyre, a tool-calling benchmark built on the PHYRE 2D procedural physics environment, where agents propose interventions and test hypotheses through simulation APIs. Experiments show that current LLM agents struggle in these settings, especially on the hardest levels of Interphyre. Claude Sonnet 4.6 achieves only 2% success, while HExA improves the same model to up to 77% success. HExA also improves open-weight models and outperforms agentic baselines such as ReAct and Reflexion. Moreover, using only skills learned from easier levels and transferred without active experimentation, HExA achieves 44% success, demonstrating the reusability and generalization of its learned skills. Overall, HExA shows that learning through active experimentation can help agents discover useful knowledge, acquire reusable skills, and make efficient progress on novel long-horizon tasks. less

On-policy self-distillation (OPSD) trains a reasoning model on rollouts sampled from its own policy by matching a privileged teacher that also sees verified reference solutions. Existing OPSD objectives supervise only the output distribution, so privileged context affects training through a token-level divergence without directly supervising the internal computation that produced that distribution. We propose Privileged Hidden Flow (PHF), which additionally distills how a privileged teacher's hidden states move along the same rollout. Rather than forcing each student hidden vector to match the teacher vector at the same token position, PHF aligns token-to-token transition directions and trajectory geometry over selected generated positions. The all-layer recipe also includes an adjacent-layer relation computed from these same transitions, without pointwise hidden-state imitation. Under the same 100-step training schedule, PHF improves the Average@12 aggregate over our reproduced OPSD baseline on Qwen3-1.7B, 4B, and 8B, with observed gains of about +2.2, +1.5, and +1.7 points. The transport objective is exactly invariant to shared trajectory offsets; its local geometry term is also invariant to orthogonal transformations of transition directions. Ablations distinguish the fixed PHF recipe from pointwise hidden-state matching, single-channel transition losses, and layer-subset choices, supporting PHF as a compact hidden-flow extension to OPSD. less

Sparse autoencoders (SAEs) have become a leading tool for interpreting the representations of vision foundation models, decomposing their polysemantic activations into a larger set of sparse, more monosemantic features. The Top-$k$ SAE, a now-standard variant, enforces sparsity architecturally through its activation function, retaining only the $k$ most active latents per input. Because it was designed precisely to avoid the $\ell_1$ penalty used by earlier SAEs and its known drawbacks, it has not been combined with an explicit sparsity regularizer, despite retaining limitations of its own, such as a budget $k$ that is fixed regardless of input complexity and a tendency to overfit to the training value of $k$. We introduce two sparsity regularizers compatible with the Top-$k$ architecture, both acting on the activations before the Top-$k$ selection: an $\ell_1$ penalty on the unselected (off-support) units, and a scale-invariant $\ell_1/\ell_2$-ratio penalty that concentrates the code onto fewer effective units. Both penalties are applied only to the batch-active units, those selected by the Top-$k$ operator at least once within the batch. Across two datasets, three vision foundation models, and a range of $k$, both regularizers consistently improve monosemanticity at no cost to reconstruction quality. The $\ell_1/\ell_2$ penalty further concentrates information into fewer latents, making reconstruction more robust to the inference-time choice of $k$ and improving small-budget linear probing. Our central finding is that hard architectural sparsity and soft sparsity regularization are complementary rather than mutually exclusive. less

AI nudification uses generative models to create synthetic non-consensual sexually explicit imagery (SNEACI) of real individuals. Prior work has examined dedicated nudification platforms and model repositories, finding that most targets are female celebrities. However, the anonymous content community, where SNEACI is actively requested, generated, and exchanged, remains unexplored. In this work, we present a large-scale study of AI nudification in the wild, identifying 24,105 SNEACI items. We find a significant shift in target demographics: non-celebrity individuals now account for 55.8\% of targets, compared to only 4.7\% in prior studies, indicating that AI nudification has expanded from targeting public figures to increasingly harming individuals within users' own social circles. Meanwhile, open-source models dominate production, with Stable Diffusion family generating 42.7\% of images and Wan generating 66.5\% of videos, all driven by thousands of shared fine-tuned models and accessible tutorials. Yet the ecosystem runs on a small cohort of active producers, with the most prolific producing 780 items, drives community engagement, shapes target demographics, and disseminates technical knowledge that lowers barriers for new producers. Our work provides an empirical understanding of how AI nudification operates in the wild, revealing the mechanisms that sustain this ecosystem and highlighting the urgent need for interventions in platform governance, technical safeguards, and affected individual protection. less

Modern AI systems are increasingly deployed under non-stationary computational, demographic, and operational conditions in which static resource allocation strategies degrade both predictive performance and human-centric properties such as fairness and explainability. This paper presents AURORA-AI, an Adaptive Utility-driven Resource Orchestration framework for Resilient AI that unifies Hamilton-Jacobi-Bellman feedback control, Lyapunov-based stability monitoring, and a fairness-aware composite utility into a single closed-loop policy.The framework continuously redistributes computational budget across a population of heterogeneous AI models so that the global utility, defined jointly over predictive performance, demographic parity, cost, latency, robustness, and interpretability, remains maximised under disruption. The framework is evaluated in a stress-rich discrete-time simulation that concurrently injects demographic bias shocks, gradual concept drift, and abrupt black-swan disruptions, and is compared against five established controllers including Static, Round Robin, Greedy, LinUCB, and a deep reinforcement-learning agent based on Proximal Policy Optimisation. AURORA-AI achieves immediate recovery from the black-swan event compared to eighty-eight time steps for the Static baseline and twenty-two for Proximal Policy Optimisation, lifts the alpha-quantile and the super-quantile by twenty-nine and twenty-five percent respectively, simultaneously reduces the mean and maximum demographic parity gap, and increases the fraction of Lyapunov-stable operating steps. These results indicate that fairness-aware adaptive orchestration grounded in stability theory is a practical and theoretically motivated path toward resilient human-centric AI deployment. less

Evaluating LLM outputs remains a major bottleneck in NLP: human evaluation is expensive and slow, lexical metrics correlate poorly with human judgments on open-ended generation, and holistic LLM judges often produce opaque scores that are hard to debug. We propose BINEVAL, a framework that decomposes evaluation criteria into atomic binary questions and aggregates the resulting verdicts into interpretable, multi-dimensional scores. Given a task prompt, a meta-prompt generates fine-grained evaluation questions, and an LLM answers them independently for each output, yielding transparent question-level feedback together with calibrated overall scores. This decomposition makes evaluation easier to inspect, easier to diagnose, and directly usable for prompt improvement. Across SummEval, Topical-Chat, and QAGS, BINEVAL matches or outperforms strong baselines including UniEval and G-Eval, with especially strong results on factual consistency benchmarks such as QAGS. Beyond competitive correlation with human judgments, BINEVAL better matches human score distributions and avoids the ceiling effects common in prior LLM judges, leading to better discrimination between borderline and clearly flawed outputs. We further show that the same question-level feedback supports iterative prompt optimization, improving evaluator prompts on summarization and generation prompts on IFBench under both self-update and cross-model update settings. Overall, BINEVAL provides a task-agnostic, training-free, and interpretable evaluation framework that combines strong empirical performance with practical diagnostic and optimization value. less

Earth Observation (EO) forecasting aims to predict future Earth surface dynamics from satellite observations under changing meteorological conditions. In this paper, we view this task as a partially observed, weather-driven world modeling problem, in which weather acts as a conditioning signal, while forecasting remains uncertain due to sparse observations and unobserved land-surface states. However, existing methods do not fully capture this setting: deterministic models collapse uncertainty into a single future prediction, while diffusion-based methods typically treat weather variables as undifferentiated conditioning signals, and existing benchmarks focus mainly on reconstruction accuracy rather than whether forecasts respond correctly to changed weather forcing.We introduce EO-WM, a video diffusion transformer for multispectral EO forecasting. EO-WM incorporates a physically informed conditioning framework that represents meteorological forcing through a climatological baseline, weather anomalies, and cumulative physical stress signals. Specifically, it separates baseline and anomaly through distinct conditioning pathways, and accumulates anomalous forcing over time to capture sustained heat and drought stress. To evaluate weather-response behavior beyond standard metrics, we introduce two diagnostic benchmarks: an Extreme Summer Benchmark for severity-aware prediction of vegetation degradation under extreme weather, and a Seasonal Matched-Pair Benchmark for testing response fidelity under changed weather forcing. Experiments show that EO-WM reduces the error in predicted Normalized Difference Vegetation Index (NDVI) decline amplitude by a relative 5.63% and improves directional hit rate by a relative 7.80%, while remaining competitive on standard pixel-level metrics. The benchmarks and model will be made open-source at https://github.com/Luo-Z13/EO-WM. less

Mechanistic epidemiological models are widely used to support infectious disease forecasting and public-health decision making. Bayesian calibration of such models is commonly performed using Markov chain Monte Carlo (MCMC), which can become computationally expensive for high-dimensional nonlinear systems and repeated near-real-time analyses. Here, we investigate simulation-based inference (SBI) using neural posterior estimation as a scalable alternative for Bayesian calibration of a mechanistic SECIR epidemiological model using COVID-19 intensive care unit (ICU) occupancy data from Germany during 2020. We compared SBI and MCMC across multiple epidemic phases using both 31-day inference windows and a substantially more challenging 201-day reconstruction problem involving multiple transmission change points. Posterior agreement was evaluated quantitatively using Wasserstein distances and Kullback-Leibler divergences together with posterior predictive checks. Across the 31-day windows, SBI recovered posterior distributions in strong agreement with MCMC while accurately reproducing observed ICU trajectories. In the 201-day setting, SBI preserved the dominant posterior structure despite increased uncertainty. SBI, by combining CPU and GPU resources, substantially reduced computational runtime compared with MCMC, which was restricted to running on CPUs. Whereas MCMC required approximately 1000 seconds for the 31-day inference problems, SBI achieved comparable posterior and predictive performance in approximately 60-70 seconds on a single GPU. For the 201-day inference problem, SBI required an average of 157 seconds, while the MCMC runs took over 19,000 seconds. Our results demonstrate that SBI provides a rapid and computationally efficient framework for Bayesian calibration of mechanistic epidemiological models, supporting repeated near-real-time inference and rapid outbreak analysis. less

Digital twins have emerged as a promising paradigm for personalized healthcare, enabling modeling of individual behavior and health trajectories. In cognitive health, early detection of Mild Cognitive Impairment (MCI) remains challenging, where language and conversational patterns serve as non-invasive biomarkers. In this work, we propose a language-based digital twin framework that leverages large language models (LLMs) to mimic the conversational behavior of elderly individuals by incorporating stylometric cues and contextual metadata. To evaluate fidelity and cognitive consistency, we introduce a multi-head conditional variational autoencoder (cVAE) that jointly measures reconstruction quality and predicts cognitive scores. Experiments on the I-CONECT dataset show that the digital twin preserves identity-specific characteristics and achieves reconstruction and MoCA prediction errors comparable to real data, while outperforming baseline GPT-generated responses. These results highlight the potential of language-based digital twins as a scalable and non-invasive approach for personalized and continuous cognitive health monitoring. less