Papers

Multi-step causal reasoning requires chaining inferences where each step constrains the next. An early error propagates silently, and a correct answer reached via flawed logic evades outcome-level detection. In specialized domains, teacher LLMs err on intermediate steps, safety constraints restrict cloud distillation, and shifting conditions demand adaptation, leaving self-evolution as the practical route. Naive self-evolution can collapse: outcome-only rewards let the model exploit distributional shortcuts, and weak self-evaluation reinforces spurious paths into stable failure patterns. We exploit a key asymmetry: generating a correct chain is hard, but verifying a single step is easy. Many high-stakes domains admit a deterministic, queryable oracle, a physics simulator or rule engine over codified constraints. It checks asserted steps without teacher-level ability and abstains beyond its rules; it can check what the model asserts, never replace it. This enables CoEvo, an oracle-grounded self-evolution framework where a single model alternates between Proposer and Solver. As Solver, the model generates competing chains; intra-group debate exposes disagreement steps, a proxy for the capability boundary, and the oracle adjudicates them into process-level supervision. As Proposer, the same model constructs progressively harder scenarios inside oracle constraints, steering the curriculum toward deep multi-hop chains. Both roles are updated jointly, so training pressure co-evolves with the model. On industrial, clinical, and legal multi-step causal reasoning benchmarks, CoEvo enables an 8B LLM to sustain self-evolution, surpassing distillation baselines and the strongest proprietary reference on path correctness (82.1% vs. 71.4%). The trained model generalizes to unseen categories and systems, preserving root-cause accuracy.

We present a learning algorithm for inferring threshold Boolean networks (TBNs) with a prescribed set of fixed points. The proposed method employs a custom differentiable loss function that jointly enforces fixed point preservation, penalizes spurious attractors, encourages binary outputs, and promotes sparsity through L1 regularization. Applied to the FOS-GRN model of Arabidopsis thaliana, the approach achieved perfect reconstruction (i.e., all 10 desired fixed points and no spurious ones) in 5 out of 30 independent runs, recovering on average 8.53 $\pm$ 0.90 correct fixed points with no spurious attractors. In contrast, standard methods such as the Perceptron and Logistic Regression recovered up to 10 fixed points but introduced between 8 and 31 spurious ones. An additional analysis varying the sparsity coefficient ($\lambda$) confirmed that the method's performance and the structural properties of the inferred networks remain robust within a practical range (up to 0.01) of regularization strengths. Overall, the results demonstrate the effectiveness and stability of the proposed algorithm in capturing meaningful network dynamics under prescribed dynamical constraints.

Automated training of surgeons is one of the most crucial factors that significantly minimize surgical training risks and expenses. With recent advances in artificial intelligence (AI) knowledge and available data from various surgeries, AI's involvement in surgical training is becoming very promising. It is recommended that at the early stages of AI development, it interferes in the surgery as a third agent alongside the trainer. As trust in AI increases, this process will lead to an AI agent acting as a trainer in the future. The first phase in which AI can intervene in the training process is to suggest an improved surgical path to the trainer. A platform must be constructed in the first step, to accomplish this task and to enhance the movement path of trainee surgeons. This paper introduces this platform along with an annotated capsulorhexis surgery dataset called the ARAS-Farabi dataset. In this research, a deep convolutional neural network is pre-trained with JIGSAWS and ARAS-Farabi surgical datasets that can extract surgical skill characteristics from surgery tool tip motion data. The proposed platform develops a reference model from the feature space of an expert surgeon's movement trajectory and proposes an improved path to enhance the skill of a novice surgeon. An optimization with two loss functions is utilized to create a path that raises the skill level of the novice surgeon's path while simultaneously predicting and preserving his/her intent. The results of this study reveal that, with the assistance of an AI agent, the trainee surgeon's movement path can be enhanced by at least 20 percent while maintaining his intentional objective. In addition to the recommended deep network, various tangible indicators have also been developed in this research to verify the level of trainee improvement.

We introduce Malaria-Instruct, a curated instruction-following dataset derived from the ChEMBL Legacy Malaria corpus for Malaria virtual screening, and conduct a systematic evaluation of five open-source LLMs; Gemma-2 2B/9B, TxGemma-2B/9B, and LlaSMol-Mistral-7B, on a rigorous out-of-distribution data split. Performance was benchmarked against classical ML models (Random Forest, XGBoost) and frontier proprietary models (Gemini 2.5, OpenAI o3) under few-shot conditions. Fine-tuned LLMs substantially outperformed all baselines: TxGemma-9B achieved the highest ROC-AUC ($0.731 \pm 0.005$) and LlaSMol-Mistral-7B the best enrichment factor (EF@1% $\approx$ 4.99). Domain-specific fine-tuning proved categorically indispensable with TxGemma-9B collapsing from ROC-AUC 0.731 to 0.499, under its best few-shot condition, and neither Gemini 2.5 (ROC-AUC $\approx$ 0.53) nor o3 (ROC-AUC $\approx$ 0.59) achieved reliable discrimination without fine-tuning. Biomedical pretraining conferred a measurable advantage at equivalent scale, while chemistry-aware pretraining yielded superior prospective enrichment. Fine-tuned open-source LLMs represent a compelling, resource-efficient paradigm for antimalarial VS, outperforming both classical pipelines and proprietary reasoning models under structurally challenging conditions.

An aggregation pass gives a language model two things: candidate solutions and another chance to answer. We ask whether the candidates improve that attempt, and whether aggregation improves on selecting an existing answer. Our candidate-free control asks the same model to solve the problem again from the problem alone, using the same decoding settings and maximum output-token allowance. On AIME-2025 and HMMT-2025, Qwen3-4B-Instruct-2507 reaches 15.5% accuracy when all four candidates are wrong, compared with 27.7% for a fresh solve. When at least two candidates are correct, the ordering reverses: 88.7% for aggregation versus 59.7% for a fresh solve. With exactly one correct candidate, the benefit remains unresolved under the specified test. Yet plurality voting reaches 96.0% in the multiple-correct regime without another model generation; its accuracy difference from aggregation is inconclusive (raw $p = 0.180$). The all-wrong accuracy loss and multiple-correct gain survive a statistical correction for adding the second benchmark, developed after data collection, and persist when fresh responses are scored by their final answers. These findings concern one 4B model, two mathematics benchmarks, and a single aggregation pass over candidates truncated at their first answer. All-wrong recovery can arise from solving the problem afresh. The fresh-solve control tests whether candidates improve accuracy. Comparing with selection asks whether the additional generation improves on an existing answer.

We address two open questions in streaming PCA via Oja's algorithm: sharp operator-norm convergence for general rank under sub-Gaussian data, and distributional inference for the resulting subspace estimator. Existing convergence analyses, even in the rank-one case, either assume bounded data or leave non-vanishing remainder terms that prevent adaptation to a polynomially vanishing tail spectrum, while existing distributional results are confined to the rank-one case. Our convergence theory removes these remainder terms and yields a sharp rate. In the dense-tail spiked covariance regime, this rate matches the minimax rate up to logarithmic factors. More generally, we prove a matching lower bound, up to logarithmic factors, across both dense-tail and sparse-tail regimes under a mild nondegeneracy condition. The analysis yields a linearization of Oja's iterates, which in turn enables a high-dimensional Gaussian approximation for the general-rank subspace estimation error with an explicit limiting covariance. We also establish a row-wise Gaussian approximation over convex sets for the aligned difference, recovering prior rank-one results as special cases. For practical inference, we develop an online multiplier bootstrap algorithm and prove its consistency. Beyond streaming PCA, our techniques contribute to Gaussian approximation and bootstrap inference for nonconvex stochastic approximation.

Tool-using large language-model agents can complete a task while exercising authority that the user did not grant or the task does not need, causing excess-authority errors. Traditional permission gating systems alone for validating agent environments are insufficient. We study whether post-training can teach a 4B-parameter model to choose task-conditioned authority in executable terminal and Model Context Protocol (MCP) environments to complement those measures. We propose a framework where each action is audited before execution and again from observed effects along six dimensions of risk. This auditing is conducted using deterministic verifiers that score completion, evidence, exact state, prohibited attempts, and safe success. In conjunction with predefined task-specific sufficient-authority envelopes, we determine task-specific excess privilege values for trajectories, which are then optimized for in post-training. We find that after training using this framework on Qwen3.5-4B over 1,500 tasks, the selected seed reaches 98.48% safe success across 2,896 evaluation episodes spanning all 500 held-out tasks, compared with 64.36% for the base policy, and reduces excess-authority error events from 4.56% to 0.79%. Furthermore, external tests show capability retention and prompt-directed improvement. A 400 task continuation study also found evidence of generalization, reducing excess-authority events by 6.99 percentage points while maintaining previous capabilities. We conclude learned restraint through least-privilege aware post-training is therefore useful as an additional control layer for tool-using agents in executable terminal and MCP environments, but it does not replace permission gates and sandboxing.

Coupled-cluster theory defines the accuracy standard for molecular electronic-structure properties but scales too steeply for routine application, whereas density-functional theory is affordable yet systematically biased. We resolve this trade-off with a single equivariant network, HARP (Hamiltonian Read-out for Properties), that predicts an effective one-electron Hamiltonian from one inexpensive B3LYP/def2-SVP calculation and derives a broad suite of properties from it (energy, optical gap, dipole, quadrupole, polarizability, Mulliken atomic charges, and Mayer bond orders) at coupled-cluster accuracy across nine main-group elements, including the under-served phosphorus, sulfur, and chlorine chemistries. The model is trained on a new in-house dataset of multi-property labels computed at the CCSD(T) level for all nine elements. On a held-out test set, it reduces the error of every property by a factor of 3.8 to 270 relative to semi-local, hybrid, and double-hybrid DFT (referenced to composite CCSD(T)/cc-pVTZ), while adding only ~0.1 s wall time per molecule, delivering coupled-cluster-quality predictions at the cost of a single DFT calculation. Critically, deriving every property from a predicted Hamiltonian rather than pooling per-atom features builds the correct size-scaling into the model architecture: on pi-conjugated oligothiophenes it matches finite-field CCSD polarizability to ~1% and the EOM-CCSD optical gap to ~3% at the largest sizes where those references remain affordable (44 and 37 atoms, where a single CCSD field point already costs ~500x the model's entire inference) and extrapolates the corrected trends to 58-atom chains, a regime where pooling-based architectures fail by construction. Accurate extrapolation is therefore set by the model's inductive bias rather than by the training data.

Persistent acoustic monitoring can detect machine faults without physical contact, but always-on inference is constrained by power, latency, and deployment complexity. We demonstrate autoencoder-based acoustic anomaly detection on an Intel Loihi 2 neuromorphic processor under clean and noisy conditions. Log-mel features are computed off chip; normalization, autoencoder inference, L1 reconstruction scoring, and thresholding run on chip. In a clean, microphone-position-invariant ToyADMOS ToyCar benchmark, the on-chip model achieves 0.9959 AUC and 0.9785 standardized pAUC at maximum false-positive rate 0.1. In the DCASE 2026 Task 2 ToyCar noisy benchmark, the model achieves source AUC 0.7990, target AUC 0.6466, and pAUC 0.6426, exceeding reported baseline metrics. Power profiling on a 16-chip Loihi 2 VPX system shows real-time throughput with 0.0406$\unicode{x2013}$0.0426 mJ dynamic energy per sample, two orders of magnitude lower than both a CPU and GPU. These results support neuromorphic acoustic anomaly detection as a practical candidate for low-power, persistent machine monitoring.

Vision-language models (VLMs) have demonstrated remarkable zero-shot capabilities yet remain sensitive to real-world distribution shifts during inference. Although significant efforts are devoted to adapting VLMs at test time, they rely heavily on noisy pseudo-labels predicted directly from raw embedding similarities during inference, which are unreliable under distribution shift and mislead the adaptation. To avoid noise amplification, existing works craft coarse-grained surrogate objectives during adaptation, which fail to explicitly model sample-level relationships across different modalities, creating objective mismatch with inference, thus leading to marginal performance improvement. In this work, we aim to bridge the detached objectives of inference and adaptation for VLMs, and propose a principled VLM TTA method called \algname. For VLM inference, we formulate the zero-shot image classification task as a cross-modal alignment problem encoded via a Wasserstein OT formulation, providing robust pseudo-labels at the sample-level to effectively adapt VLMs. For VLM adaptation, we adopt a soft-label InfoNCE loss to adapt VLMs based on the OT-induced pseudo-labels, leveraging fine-grained supervisions to explicitly model relationships of individual image-text pairs via contrastive learning, which empowers accurate inference at the same granularity. Moreover, we theoretically reveal that the InfoNCE loss can be neatly reformulated as a Wasserstein OT formulation, thereby unifying the objectives of the inference and adaptation of VLMs to achieve their mutual benefits. Extensive experiments demonstrate the effectiveness and efficiency of our methods, outperforming the best-performing methods by up to 7% with state-of-the-art efficiency.

Whether input-dependent ("dynamic") combination of a regression model pool beats the best static blend depends on the shift and is rarely known before deployment. Can a small labeled target-domain probe tell us when reallocating trust across regions of the input space will pay off? We answer this with $\widehat{D}_{\mathrm{CF5}}$, which estimates from the probe the cross-fitted gain of the regionwise convex combination over the best static convex blend: the realizable value of deciding, region by region, whom to trust. Across a frozen suite of 12 dataset-shift pairs (spatial, temporal, domain, feature-cluster), $\widehat{D}_{\mathrm{CF5}}$ predicts realized regionwise test gains with dataset-level Spearman $+0.98$ (95% CI $[+0.83, +1.00]$; $p=5\times10^{-5}$), including two cases overturning preregistered expectations. The relationship holds in a 16-pair sensitivity analysis (Spearman $+0.83$), whereas alternative probe diagnostics reach at most $+0.66$. This contrast isolates regional trust reallocation: correlation is $+0.98$ for regionwise-convex gain, but $+0.01$ for smooth covariate-dependent stacking after affine correction. A controlled generator shows dynamic gains arise from the interaction of shift heterogeneity and local competence, increase with shift severity, and become realizable between 128 and 256 probe labels in the tested grid. The Probe-Validated Ensemble Selector chooses among a static affine stacker and dynamic realizers, deploying a candidate only when a held-out lower confidence bound clears the static-convex floor. In a preregistered prospective batch, it matched or improved the floor in all 12 runs; two deployments reduced test risk by 11% and 16%, while the gate rejected a candidate whose un-gated deployment incurred $>30\times$ the static loss. We release OpenRegShift, a reproducible evaluation harness for regression ensembles under distribution shift.

Neural networks trained on nonstationary tasks frequently lose the ability to fit new targets, a phenomenon referred to as loss of plasticity. We identify a novel source of plasticity loss due to the growing anisotropy of weight matrices' singular values during training, and analyze this phenomenon both empirically and theoretically. To mitigate this issue, we introduce SingularClip, a procedure that periodically clips the singular values of all weight matrices. We show that SingularClip performs strongly against baselines across a range of tasks in both continual supervised learning and deep reinforcement learning.

Federated multimodal models often assume every site has every modality, although hospitals differ in access to EHRs, chest radiographs, and ECGs. We study this setting on a MIMIC-derived respiratory deterioration task with simulated FL clients and introduce FedCoRe (Federated Cross-Modal Representation Completion). FedCoRe learns representation- or logit-space corrections rather than generating synthetic ECGs or CXR images. When a client observes a modality that may be missing at deployment, it evaluates the same example with and without that modality to obtain paired supervision. Only clients with such pairs update the completion module, and validation may retain the unchanged prediction. We freeze the trained multimodal predictor during evaluation so that measured differences come only from completion. Hiding ECG reduced AUROC by about 0.085; paired-example FedAvg restored 0.0415 AUROC, or 49.0% of the lost performance. We therefore report two distinct effects: paired-example FedAvg partially recovers the missing-ECG gap, while validation-selected completion is a task-specific classifier-logit correction rather than literal ECG recovery. For CXR, effect-aware completion recovers 52.8% of the loss in a controlled test where CXR is hidden. Paired-example FedAvg transfers part of this effect, but validation keeps the no-completion baseline for deployment cases whose inputs lack CXR. Thus, FedCoRe should be read as a validation-gated completion/correction framework: it can recover missing-modality signal in supported settings, but it should be deployed only when paired examples and validation evidence support that modality.

LLM-as-judge evaluation reduces response quality assessment to a single holistic A/B preference choice, providing no mechanism to isolate which quality dimensions drove the preference or distinguish model errors from genuine label ambiguity. We propose SESSE (Sketch, Expand, Sort, Summarize, Evaluate), a training-free framework that decomposes holistic judgment into structured sub-questions mined directly from the judge's own error cases; requiring no oracle responses, task-specific rubrics, or fine-tuning. On RewardBench (n=1,000), SESSE achieves near-parity with the chain-of-thought baseline and is competitive with RISE-Judge-32B (92.7%), a fine-tuned specialist, while remaining fully training-free. Per-criterion vote evidence provides an interpretable audit trail for diagnosing label ambiguity and judge failure modes unavailable from a single holistic output token.

An increasing number of scholars use AI to measure variables they subsequently include in downstream analyses. Although AI-measured variables are often analyzed as if observed without error, ignoring prediction errors in automated measurement leads to substantial bias and invalid confidence intervals in downstream analyses, even if AI measurement accuracy is high, e.g., above 90%. Existing solutions, such as design-based supervised learning and prediction-powered inference, combine error-prone AI-based measurements with gold-standard labels, which may be costly and difficult to obtain in some application areas. In this paper, we propose debiased inference with multiple imperfect measurements (DMM), a framework that combines multiple error-prone AI measurements to enable valid downstream inference without gold-standard labels. Building on the established results on CP decomposition, DMM assumes that these measurements are independent conditional on the latent true label and observed unit-level features, such as text features represented by embeddings. This framework allows for unknown misclassification rates to vary across annotation methods (e.g., large language models) and across units of annotation (e.g., texts). Under this assumption, we use semiparametric inference theory to prove that the DMM estimator is consistent and asymptotically normal, enabling valid inference for a wide range of downstream statistical analyses common in the social sciences. Our simulation results show that DMM yields valid inference and that adding accurate, though imperfect, measurements can improve efficiency. Focusing on common applications of large language model annotations, we also develop diagnostics to assess the conditional independence assumption.

The extraction of structured information from unstructured documents represents a critical component of digital transformations in all sectors. While proprietary solutions dominate commercial applications, a rapidly growing ecosystem of open-source Optical Character Recognition (OCR) engines, Large Language Models (LLMs), and Vision-Language Models (VLMs) offers accessible alternatives. However, systematic evaluations on realistic, multi-step extraction pipelines remain scarce. Responsible usage of such extraction tools require comprehensive evaluations on realistic tasks, especially as these solutions will be key components of applications in the public sector that the EU AI act categorizes as high risk. To address this gap we present a comprehensive benchmark assessing the end-to-end performance of open-source systems on a complex real-world document processing task classified as high risk: Student applications for an international study program. We conduct a comprehensive empirical evaluation with state-of-the-art OCR engines, LLMs and VLMs. Our results reveal that while VLMs generally outperform OCR+LLM pipelines, even state-of-the-art open-source models struggle to handle such tasks reliably in zero-shot settings. Only 4 of 35 configurations achieved F1 scores above 0.5, with the best OCR+LLM pipeline matching top VLM performance, though most OCR+LLM combinations performed substantially worse. Roughly 75% of all configurations scored below 0.25. Model scale influences performance, yet the relationship is non-linear: substantially larger models do not guarantee proportionally better results. Input quality, particularly the structural preservation of OCR output, emerges as a critical factor independent of downstream model capability.

Stochastic mean-field control (MFC) provides a fundamental framework for coordinating large populations of interacting agents under uncertainty, with a wide range of applications. Existing numerical and deep-learning methods solve one MFC problem instance at a time and must be re-optimized whenever the task changes. In this work, we formulate stochastic MFC as an operator-learning problem and develop, to the best of our knowledge, the first mesh-free, self-supervised neural operator for stochastic MFC. The main challenge is that the diffusion term in the controlled Fokker–Planck equation precludes deterministic transport-map representations. We address this challenge by combining the probability-flow ODE with an invertible normalizing-flow-based transformer, which recasts the dynamics as a deterministic continuity equation and enables closed-form score evaluation through the exact inverse and analytical log-determinant of the normalizing flow, with $\mathcal{O}(d)$ cost per particle for networks of fixed size. Through transformer-based in-context learning, task prompts, represented by compact distribution parameters or raw particle clouds, condition the transport map, enabling a single pretrained operator to solve unseen tasks in one forward pass. The resulting Normalizing Flow Invertible Solution Transformer (NFIST) is trained end-to-end by minimizing the stochastic control objective directly, requiring no precomputed numerical solutions for training. We further prove the consistency of the proposed operator-learning formulation with task-by-task optimization. Numerical experiments on stochastic optimal control, Schrödinger bridge, systemic-risk control, and obstacle-avoiding path planning demonstrate effective zero-shot generalization while substantially reducing the computational cost of solving large families of stochastic MFC problems.

Background. Advances in agentic systems are simultaneously, and rapidly, saturating benchmarks. Despite this often discussed phenomena, benchmark scores remain difficult to interpret due to the lack of control and characterization of task difficulty. More specifically, we currently have little understanding of what makes one task harder than another, and to what extent task difficulty is predictable from static task properties. Aims. We propose a measurement framework to investigate and systematically quantify what structural properties of software tasks correspond to agent success rates for issue resolution tasks. Method. We conducted a large scale empirical study on CoderForge-Preview, the largest open dataset of coding agent trajectories to date, by extracting features across task patch, repository and prompt. We evaluated the predictive power of each feature against task outcomes using ensemble methods, SHAP attribution, and effect size analysis. Results We found that task difficulty is substantially predictable from static features (AU C = 0.863) and is largely driven by patch fragmentation and repository scale. Prompt linguistic features become visible among top contributors for tasks in the mid-band, revealing a layered structure of difficulty. Conclusion. The difficulty of an issue resolution task is encoded in its structure. This enables static, pre-hoc difficulty estimation and lays the groundwork for difficulty-controlled benchmark construction for evaluation of agents.

Metasurfaces have revolutionized the development of photonic devices by enabling unprecedented precision in light manipulation. However, their design processes are often constrained by computationally expensive simulations and complex high-dimensional design spaces. Although deep learning has accelerated the design process by serving as a surrogate model, it remains constrained by task-specific architectures and lacks universal reasoning capabilities. This review surveys how Large Language Models (LLMs) are adding semantic interfaces, code generation, and tool orchestration to established numerical nanophotonic workflows. We first outline the development from classical neural networks to transformer-based models and their applications in nanophotonic design. We then review the emergence of LLM-related methods in nanophotonics and organize them into two operational modes: surrogate models that treat structure-spectrum mapping as a language task, and agentic systems that have been demonstrated to generate code, orchestrate selected simulation steps, and support closed-loop optimization. Furthermore, to identify future cross-disciplinary opportunities, we briefly explore applications of LLMs in research fields such as materials science and wireless communications. This review concludes by looking ahead to the next generation of multimodal foundation models with physical perception capabilities. In this vision, artificial intelligence is evolving from passive tools into active collaborators, participating in autonomous scientific discovery.

OpenAI Privacy Filter is a compact, bidirectional token-classification model for detecting and redacting personally identifiable information (PII) and secrets in unstructured text. The model is derived from an autoregressively pretrained checkpoint and converted into a bidirectional, banded-attention classifier that labels an input sequence in a single forward pass. A constrained Viterbi decoder produces coherent spans across eight privacy categories and exposes configurable operating points for precision-recall tradeoffs. Privacy Filter has 1.5 billion total parameters, 50 million active parameters per token, and a 128,000-token context window. It is designed for efficient local deployment and domain-specific fine-tuning. Privacy Filter is intended as a configurable data-minimization component within layered privacy workflows, not as an anonymization or compliance guarantee.

On-policy self-distillation (OPSD) trains a student on its own responses using token-level supervision from the same model conditioned on privileged reference information. We investigate whether performance gains from OPSD show that the student learned the information in the reference or instead reflect recovery of reasoning behavior already present in the base model. We perform OPSD experiments on science and mathematics datasets using Qwen3 models ranging from 1.7B to 8B. Our analysis framework separates the supervision induced by the reference from the supervision provided by the teacher without the reference and measures how each aligns with changes in the student's predictions. The correct reference does not provide a consistent performance benefit across teacher generation modes, model sizes, and training datasets. Students can improve without the correct reference, and a solution from another problem can outperform the correct solution on several mathematical reasoning benchmarks. The student's predictions align more strongly with the base model's thinking behavior than with the supervision induced by the reference, but controls constructed from other problems reproduce much of both alignments. Moreover, stronger alignment attributable to the correct reference does not reliably coincide with a greater performance benefit from the reference. Performance gains and distributional alignment alone therefore cannot determine how privileged reference information contributes to student learning in OPSD.

Today most media content is consumed based on algorithmic recommendations. Evidence suggests that this can lead to politically biased media consumption patterns. Automated extraction of political agendas from texts can reveal and analyze political biases in online media – and thus help fostering politically unbiased media consumption. Here we employ modern political text analysis methods demonstrating the potential of automated fine-grained political bias analysis in online media. We conduct an analysis of political content in German language online media during the period 2019–2022, encompassing several million articles and tweets covering events with profound societal impact globally and nationally, the COVID-19 pandemic and the beginning of the war in Ukraine. Our analysis identifies thematic similarity between national (German and Swiss) reporting, particularly for categories driven by international events. We also find divergences emerging in domestically influenced categories, reflecting differences in national policies and institutional structures. A comparison of newspaper and Twitter discourse reveals that both media converge around a shared core during the pandemic, yet differ in intensity and temporal dynamics. Newspapers exhibit more stable political content, while Twitter reacts through short-lived event-driven spikes. These findings indicate that international crises act as a powerful synchronizing force on political content in classical media, temporarily overriding both national and media-form differences. Our automated political analysis empowers citizens by rendering political agendas in online media transparent. This transparency also enables media outlets to bridge the gap between algorithm-driven echo chambers and a more informed, balanced public discourse.

Traffic signal control is a challenging sequential decision-making problem requiring a balance between throughput, vehicle delay fairness, signal stability, predictable switching, and emergency vehicle prioritisation. Existing reinforcement learning (RL) approaches often rely on predefined objectives and offer limited adaptability to changing operational priorities. They may also generalise poorly across geometrically equivalent intersections. To address these limitations, SIGMA (Symmetry-aware, Intelligent, Geometric, Multi-faced Adaptive traffic control) is proposed as a multi-faced RL framework augmented with a large language model (LLM) for dynamic objective adaptation and orientation-invariant policy learning. Natural language emergency instructions are translated into priority vectors that guide a multi-faced actor-critic controller. Symmetry-aware rotational augmentation supports transfer across homogeneous four-way intersections, while an offline-to-online learning strategy enables stable initialisation and adaptation to evolving traffic conditions. Structural properties concerning emergency service levels, graceful degradation under LLM failure, and sensitivity to traffic demand perturbations are established, alongside bootstrap-based statistical validation. Experiments in SUMO across four urban intersections modelled on Kolkata traffic scenarios compare SIGMA with fixed-time, actuated, and Deep Q-Network controllers. Reductions in average waiting times, emergency waiting times, and queue lengths are observed, alongside improved traffic throughput. Empirical reliability and ablation analyses further assess robustness to component failures and geometric orientation. SIGMA integrates language-guided objective adaptation, multi-faced policy learning, and statistical reliability assessment into an adaptive framework for intelligent traffic control.

Serving a 235B-parameter Mixture-of-Experts (MoE) model on a single 8 GB GPU is bottlenecked not by compute but by memory bandwidth: decode must stream each token's active experts from whichever tier holds them, and on consumer hardware most experts sit on an SSD far slower than RAM. We quantify this bandwidth wall on Qwen3-235B (Q4_K_M, 134 GB): measured decode is 0.44 tok/s warm, matching a bytes-per-token / bandwidth model, while a batching scheme that should amortize one disk sweep instead collapses at batch 32 from paging thrash. We build llama-moe-trace, a zero-surgery router-telemetry tool, and measure routing on Qwen3-30B: adjacent-token expert reuse is 2.0x chance, 95% of traffic uses 52.5% of experts, and an LRU cache of 13.4% of experts serves 66% of requests. We then ask whether cacheability is trainable: we pre-register training of 137M MoE language models with auxiliary locality and domain router losses, under joint criteria on cache-miss reduction and perplexity. The mechanism works (misses down up to 60%; a 99% static-pin hit rate) but every configuration fails the pre-registered <=1% perplexity gate – miss reduction and quality are tightly coupled. Concurrent StickyMoE reports the same loss as near-free on single-domain sub-25M models; on multi-domain 137M we find the tax real. Our contribution is this pre-registered, stricter-criterion, multi-domain evaluation plus edge-serving measurements. A 340M rung shows the tax does not shrink with scale (it rises slightly). We further show training-free cache-aware rerouting stacks with trained locality – together ~80% miss reduction at <=3.4% perplexity at both sizes, far cheaper than either alone – while domain-primed prefetching does not help. All code, traces, and the pre-registration are released.

Large Language Models (LLMs) are being increasingly used in everyday applications. A major challenge in the context of LLMs or Artificial Intelligence (AI) in general is to ensure privacy when using them, meaning that personally identifiable information (PII) is removed from any text that enters an LLM. These challenges have become more urgent with novel EU legislation. Uncertainty around LLM usage with respect to privacy concerns in EU countries can be a major blocker for the speed of innovation and transfer from research to applications. Here we present Redakto, a tool that can be used for anonymizing text prior to feeding it to an LLM or other downstream text processing. We provide state-of-the-art functionalities for both redaction of PII but also when used for pseudonymization. These functionalities are exposed such that they can easily be used by end-users, through the Redakto web application, and by developers and researchers, via REST APIs and model context protocol (MCP) hooks. The implementation is fully open source, requires modest compute resources, and can be readily deployed on local hardware. In contrast to prior work and in order to better assess the quality of the anonymized texts, we conduct extensive empirical evaluations on textual data from legal and medical domain with respect to both privacy and utility of the redacted texts. Our empirical results demonstrate that the texts anonymized with different redaction strategies achieve utility scores on par with the original texts, suggesting that anonymization with Redakto can be used for LLM tasks without substantial negative impact for the tasks we explored.

Fine-grained wildlife re-identification remains a challenging area in research. Current state-of-the-art approaches apply a detection and re-identification pipeline. We propose a one-stage end-to-end detection and re-identification model that performs identity searching within the latent space. We adopt DINOv2 for robust spatial geometry and MegaDescriptor for wildlife re-identification. We enhance latent queries with prompt re-identification features. A detection decoder queries the scene latent space to establish object boundaries around the target identity. Preliminary findings reflect a competitive mean average precision score of 30.584% compared to the state-of-the-art two stage approach of 44.89%. Qualitative results depict effective bounding and identification of animal identities.

Recent work has shown that representational alignment between biological and artificial neural networks is asymmetric: model representations predict neural responses much better than neural responses predict model representations. This asymmetry raises the question of whether representational geometry contributes to bidirectional representational alignment. We hypothesized that steering representational geometry during training can systematically influence bidirectional alignment. To test this hypothesis, we developed a computational framework that integrates spectral regularization with bidirectional predictivity analyses. As an initial demonstration, we evaluated our framework using self-supervised contrastive vision models. Steering the spectral geometry of the learned representations substantially increased reverse predictivity with modest reductions in forward predictivity, yielding a 55% relative improvement in bidirectional predictivity. These improvements were accompanied by reduced effective dimensionality and a reorganization of the shared representational subspace, within which forward and reverse predictivity became approximately symmetric at intermediate spectral exponents. Overall, these findings demonstrate that representational geometry can be systematically steered to modulate bidirectional representational alignment between biological and artificial neural networks.

Large language models fail on multi-step compositional reasoning, but measuring that failure is hard, because new models are trained on the benchmarks used to evaluate them. A fixed test set becomes a memorisation check soon after release. Constructive benchmarks avoid this by generating instances on demand. We introduce ClosureBench, a constructive benchmark for graph-relational logical reasoning. Each task is built from explicit primitives (reachability, degree, set operations, connectivity, aggregation), and its reference answer is computed by executing code that implements that logic exactly. Ground truth is therefore verified, and the supply of fresh instances is unlimited. The benchmark spans 26 task categories at three compositional levels, with three independent difficulty axes: graph size, edge density, and query depth. We evaluate models from 1.5B open weights to frontier systems (o3, GPT-4.1, Gemini 2.5, Claude Sonnet 4). Accuracy falls as graph size and query depth increase, and the two axes interact. The difficulty does not lie in the surface form, since it persists when the graph is given as a JSON edge list or an adjacency matrix rather than prose, nor in the reasoning rule, which models state correctly. It lies in carrying that rule out over the graph across many steps. A 4B model fine-tuned to emit verified programs instead of answers stays nearly flat across compositional levels, while every frontier model degrades. o3 falls from 96% on atomic queries to 82% on the most compositional; the 4B model holds at 93% at a fraction of the token cost. The program offloads multi-step execution to a runtime, and the model's remaining errors are almost entirely misread edges. Constructive generation also supports a direct memorisation check, comparing accuracy on seen and fresh instances.

Estimating the difference of two Stein's score functions is a fundamental problem in generative modeling. In particular, score differences arise naturally in transfer learning, where the score difference provides the mechanism for adapting a pre-trained model to a new target distribution, and in diffusion model-based post-training methods such as discriminator guidance. Existing estimators for score differences in these settings either lack of statistical consistency or are difficult to scale up in high-dimensions. We propose a statistically consistent and scalable estimator for score differences based on Sobolev regularization, which plays a crucial role in ensuring consistency and stablizing the training in the small-sample regime. Mathematically, we establish a convergence rate of $O(n^{-\frac{s-1}{d+2s-2}})$ where $d$ is the dimension and $s$ denotes the smoothness of the underlying densities, and provide a minimax lower bound of $\tilde{\Omega}(n^{-\frac{2(s-1)}{d+2s}})$ (in mean-squared error). Empirically, our estimator exhibits significantly improved stability in small-sample regimes compared to existing methods. We demonstrate its effectiveness on real-world tasks, including transfer learning for ECG signal generation, where it substantially outperforms non-regularized score difference estimators in downstream classification performance.

In this work, we investigate whether the latent representations learned by a Deep Belief Network (DBN) and a Bidirectional Gated Recurrent Unit (Bi-GRU) can discriminate among four dynamically distinct trajectory types in the three-state majority vote model (MV3): approach from disorder, approach from order, departure to disorder, and departure to order. The DBN, pre-trained in an unsupervised manner on static equilibrium samples via a Gaussian-Bernoulli Restricted Boltzmann Machine input layer and architecture $784 \to 4096 \to 225 \to 81$, encodes each lattice snapshot into an 81-dimensional latent vector. A t-SNE analysis of the DBN latent space reveals only partial separation of the four trajectory types, reflecting the fact that a model trained on static configurations cannot fully resolve directional temporal structure. A two-layer Bi-GRU classifier, trained on sequences of DBN-encoded snapshots of length $T = 50$, achieves near-perfect separation of all four trajectory types in its hidden state space, as confirmed by t-SNE visualization on both training and test sets. Furthermore, a sliding-window application of the trained Bi-GRU to continuous MV3 dynamics demonstrates its ability to sense the system's current dynamical regime in real-time. These results establish a principled hierarchical architecture for detecting and classifying critical transitions in agent-based opinion dynamics models.