Background: The widespread adoption of electronic health records (EHRs) has generated large-scale repositories of highly sensitive clinical information, emphasizing the need for robust anonymization strategies to enable secondary use for research while safeguarding patient privacy. Conventional rule-based and machine learning approaches for deidentifying medical text face limitations with the linguistic complexity, variability, and context dependence inherent to clinical documentation. Recent advances in large language models (LLMs), combined with emerging quantum computing paradigms, present novel opportunities to enhance the accuracy, scalability, and resilience of health care data anonymization. Objective: This study aims to evaluate the efficacy of LLM-based and quantum-enhanced hybrid architectures for medical text anonymization, assessing the effectiveness and computational efficiency across multiple entity types in Portuguese clinical notes. Methods: We constructed a gold-standard corpus of 1000 Portuguese outpatient clinical notes, manually annotated by 5 trained researchers for 5 protected-entity categories: patient names, dates, identifiers, organizations, and geographic locations. Four anonymization strategies were evaluated: 2 stand-alone LLMs (Llama-3.1-8B-instruct and Llama-3.3-70B-instruct) and 2 quantum-enhanced hybrid models (Dynex-QML with 8B and 70B base models) incorporating quantum optimization via Quadratic Unconstrained Binary Optimization (QUBO) formulations. The quantum-enhanced approach transforms the final attention layer of the LLM into a global constraint satisfaction problem solved via neuromorphic quantum annealing. Model performance was measured on a held-out test set of 500 notes using precision, recall, and -score metrics. Computational efficiency was quantified through end-to-end processing time. Results: The quantum-enhanced Dynex-QML-70B model achieved the highest overall performance with a macro-score of 0.855 (95% CI 0.823‐0.880), outperforming the stand-alone Llama-3.3-70B (0.726, 95% CI 0.704‐0.747), Dynex-QML-8B (0.733, 95% CI 0.709‐0.756), and Llama-3.1-8B (0.602, 95% CI 0.588‐0.615). Compared with Llama 3.3 70B, Dynex-QML (Llama 70B) improved macro-score by 0.128 (95% CI 0.091‐0.163; empirical 2-sided bootstrap

arXiv:2609.12578v1 Announce Type: new Abstract: A compact controller can coordinate stronger experts by selecting whom to consult, formulating requests, and integrating their responses. We study whether learning from both the controller's decisions and the experts' reasoning and code improves its generation after expert removal. We introduce \textsc{Rivet} for \emph{collaboration internalization}: expert-augmented reinforcement learning applies a shared outcome signal to controller decisions and returned expert spans, and verified trajectory internalization consolidates complete successful interactions through format-aware supervised training. The deployed controller generates reasoning, code, and interaction structure with local Python execution and no external LLM. Across seven competition-mathematics benchmarks, RIVET-1.7B and RIVET-4B achieve average accuracies of $28.25\%$ and $44.16\%$; Stage~II improves RIVET-4B's accuracy after expert removal by $6.49$ points, and GPQA-Diamond results provide evidence of generalization to scientific reasoning. Ablations show gains from ordinary trajectory supervision and additional format weighting, supporting the effectiveness of training on the content and structure of verified collaborations.

arXiv:2609.12704v1 Announce Type: new Abstract: A century of psychology has found that the trait words people use to describe one another vary, but the relational structure among those traits, which ones go together and which oppose, is strikingly consistent across raters and cultures. We test whether the LLM (Qwen 2.5-7B-Instruct) reproduces this structure in its internal trait representations. From millions of crowd-sourced personality ratings of fictional characters, we build a human implicit-personality matrix over hundreds of traits; from contrastive model activations, we build a matching matrix over the same traits. The two relational structures align strongly (Mantel r = 0.77), and the agreement holds trait by trait as well as in aggregate. Two dominant axes of the model's trait representations recover the social and intellectual dimensions long known to organize human personality impressions, social warmth and intellectual competence. On held-out dialogue, projecting model activations onto these directions yields personality profiles that agree with human ratings. This work establishes a framework that enables comprehensive, human-grounded comparison between internal model trait geometry and the shared structure of human personality impressions.

arXiv:2609.13009v1 Announce Type: new Abstract: Low reported scores on leading physics benchmarks, including those featured in the Artificial Analysis Intelligence Index (2026), suggest that frontier language models still struggle with advanced physics, a demanding test of their scientific reasoning and quantitative problem-solving abilities. Yet this impression does not always align with domain experts' experiences using these models in their work. We revisit these reported findings by evaluating frontier models on six widely used physics benchmarks and auditing them with experts, focusing on text-only problems with verifiable final answers. For each subfield of physics, faculty and graduate researchers with relevant expertise carefully review problem statements, reference solutions, and model responses to distinguish genuine model errors from grader errors, incorrect reference solutions, and ambiguous or underspecified questions. Most audited cases initially evaluated as incorrect reflect these benchmarking issues rather than errors in the models' physics reasoning. We then ask experts to address these benchmarking issues by correcting erroneous reference solutions and repairing or excluding flawed questions. We find that GPT-5.6-Sol's measured mean@4 rises from 47.3% to 78.7% on HLE-Physics and from 61.0% to 87.2% on CMT-Benchmark, while its corrected pass@4 reaches 94.4% on the 54 retained CritPt challenges. Corrected scores are computed on the retained evaluation subsets following expert review. Scores on the audited subsets of UGPhysics, PRISM-Physics, and PHYBench also rise substantially after correction. These findings suggest that current benchmarks substantially understate frontier models' ability to solve well-posed physics problems. Near-saturation on these closed-ended tasks highlights the need for more demanding, expert-validated evaluations.

arXiv:2609.12278v1 Announce Type: cross Abstract: World models let agents plan by predicting the consequences of their actions, but changes in the environment can make them inaccurate. We study the problem of adapting a world model to an unknown test-time environment, drawn from a known environment family, using only a few episodes of interaction. Existing approaches trade off computational cost against expressivity, i.e., the range of models a method can produce. For example, in-context learning is computationally cheap but limited in expressivity, and gradient-based adaptation is expressive but computationally expensive. We present CLAW (Context-conditioned Low-rank Adaptation of World models), which addresses this tradeoff by using a hypernetwork to generate low-rank (LoRA) adapters at test time. During pretraining, we simulate adaptation to a variety of environments and jointly train the hypernetwork and base world model. At test time, we freeze the base model and use a forward pass of the hypernetwork to generate adapters from a small batch of test-time transitions. We evaluate CLAW in locomotion and manipulation environment families that vary in dynamics, embodiment, and reward. We show that, using only seconds of test-time data, CLAW outperforms gradient-based adaptation and in-context learning during online adaptation. We also show that CLAW avoids overfitting in data-scarce regimes, that its advantage comes from the expressive adapters rather than context conditioning, and that pretraining the hypernetwork jointly with the base model outperforms training it post hoc.

arXiv:2609.12419v1 Announce Type: cross Abstract: Reinforcement learning with verifiable rewards is typically performed on-policy, keeping training data close to the current policy but limiting learning to trajectories that the policy can discover itself. Off-policy methods such as supervised fine-tuning, on the other hand, can leverage external knowledge beyond the base model's capabilities, but may suffer from large distribution shift. The key challenge is thus to expand exploration without sacrificing learnability. In this work, we introduce Minimal Intervention Reinforcement Learning (MInTRL), which expands the exploration frontier through sparse, local interventions in otherwise on-policy rollouts. During generation, a judge-intervention policy periodically reviews the current policy's output, replaces erroneous suffixes with short corrections, and immediately returns control to the policy. During training, MInTRL adopts a sequence-level advantage-regression objective that eliminates the need for importance sampling. We show that sparse, local interventions can substantially improve coverage beyond finite-budget on-policy sampling while preserving the overall on-policy nature of the resulting trajectories. Across math and code benchmarks, MInTRL consistently outperforms standard on-policy and off-policy baselines. Ablations show that MInTRL remains effective with self-intervention and across different judge policies, while performance peaks at moderate intervention intensity, highlighting the importance of intervening minimally. These results establish minimal intervention as an effective paradigm for enhancing on-policy RL.

arXiv:2609.12607v1 Announce Type: cross Abstract: Large Language Model alignment typically relies on learned proxy reward models, which significantly increase the memory footprint during training and are notoriously prone to instability and reward hacking. While offline methods like Direct Preference Optimization (DPO) bypass the reward model, they lose the ability to perform online exploration. If no optimization constraints are applied, this can lead to format collapse in bounded, continuous spaces. To resolve this, we propose Direct Preference Density Alignment: An alternative framework that removes the need for a learned proxy reward model while strictly preserving the benefits of online reinforcement learning. We leverage large-scale user data (approximately 90,000 samples) to construct non-parametric preference density maps, establishing an empirical reward surface. In addition to removing the reward model, Direct Preference Density Alignment enables the combination of the online structural grounding of Group Relative Policy Optimization (GRPO) with the targeted offline refinement of DPO. We show that this GRPO+DPO combination achieves the highest performance, and in a blind audio equalization listening test, enables a 1.5B-parameter model to achieve perceptual parity with a carefully prompt-engineered GPT-4o mini baseline, using only a fraction of the inference compute.

arXiv:2609.12841v1 Announce Type: cross Abstract: Mapping observed system behavior to standardized frameworks like MITRE ATT&CK is essential for threat-informed defense, but remains largely manual. Existing automated methods depend on Cyber Threat Intelligence reports, which offer only retrospective accounts of attacks. Low-level telemetry, i.e. kernel-level system calls, instead provides evidence of adversary behavior, yet its volume and complexity have limited its use for automated mapping. We present a methodology that collects kernel-level events via eBPF, correlates attacker commands into a provenance graph, and derives compact graph representations suitable for LLM-based reasoning. These representations are mapped to the MITRE ATT&CK framework using both pure LLM prompting and retrieval-augmented generation (RAG) grounded in the ATT&CK knowledge base, producing ranked technique candidates along with supporting rationales. We implement this methodology as an end-to-end pipeline, named Trace2ATT&CK and evaluate it on 347 Linux Atomic Red Team tests using locally deployed open-weights LLMs. RAG consistently improves ATT&CK mapping performance over pure prompting, while provenance graph substantially outperforms raw telemetry. These results show that local inference over graph-based behavioral descriptions can make automated ATT&CK mapping from kernel-level telemetry operationally viable, without compromising data confidentiality.

arXiv:2609.12860v1 Announce Type: cross Abstract: Computed tomography (CT) and positron emission tomography (PET) provide complementary anatomical and functional information for cancer diagnosis and treatment planning. However, the widespread use of PET is limited by high radiation exposure, elevated costs, and restricted availability. To address these limitations, deep learning-based CT-to-PET translation has emerged as a promising approach for synthesizing PET-like information directly from CT images, although accurately modeling the large cross-modal gap remains challenging. In this work, we propose a 3D CT-to-PET translation framework based on latent Brownian Bridge Diffusion (BBDM). The method consists of two stages. First, a Variational Autoencoder (VAE) is trained on paired CT-PET patches, integrating contrastive learning to improve latent alignment between anatomical and metabolic representations. Second, a BBDM is trained in the latent space to translate CT latent representations into their corresponding PET counterparts. The translated PET latents are then decoded and stitched to reconstruct the final 3D PET volume. We evaluate the proposed approach on two publicly available datasets. Quantitative results based on image fidelity and lesion-level PET-specific metrics demonstrate improved performance compared with competing methods. In particular, the proposed approach improves PET signal fidelity, better preserves clinically relevant uptake patterns, and shows improved performance in preserving small-lesion metabolic activation, paving the way for virtual imaging applications.

arXiv:2609.13031v1 Announce Type: cross Abstract: With the recent rise and adoption of tabular foundation models, optimizing their inference performance becomes an emerging field for efficiency research. While the models are architecturally similar to transformer-based large language models (LLMs), the size and serving patterns differ significantly. We show that the focus should be on the attention calculation and less on weight or KV cache quantization, which are more popular in LLMs. We develop a quantization strategy for queries, keys, and values to FP8 and use explicit FP8 matrix multiplication instructions to speed up the attention calculation. We find that it is crucial to align the quantization error in the test rows with the quantization error in the training rows, as otherwise the accuracy drops drastically. Our Triton kernel achieves a speedup up to 1.7x over regular 16-bit kernels, and we show that on TabPFN-v3 and TabICLv2 there is no relevant accuracy loss across TabArena and BeyondArena.

arXiv:2609.13043v1 Announce Type: cross Abstract: Robust medical image segmentation across imaging modalities is challenging because of large differences in appearance and intensity distributions. Models trained on a single modality often show substantial performance drops when applied to unseen domains. In this work, we develop a unified 3D pancreas segmentation framework that applies domain-adversarial learning to 4,604 heterogeneous CT and MRI scans to learn anatomical representations. A shared nnU-Net encoder-decoder is trained for whole-pancreas segmentation, with a latent domain discriminator encouraging CT-MRI feature alignment. The learned encoder is subsequently transferred to pancreatic head-body-tail segmentation using limited MRI-only subregion annotations. An average Dice score of 87.31% on the in-distribution test set and Dice scores ranging from 84.20% to 88.09% across external OOD datasets were achieved in whole pancreas segmentation. Dice scores of 80.53% on MRI and 83.05% on CT were achieved for downstream subregion segmentation, without using CT subregion annotations. These results demonstrate that a unified anatomical representation can support both cross-modality pancreas segmentation and label-efficient downstream transfer.

arXiv:2609.13071v1 Announce Type: cross Abstract: Enterprise Digital Twins (EDTs) promise data-driven decision support at organizational scale, but realizing them requires navigating siloed departments, tacit knowledge, and high-stakes decisions with long-horizon consequences. Existing approaches involve domain experts during model development but focus less on early organizational buy-in in EDTs. We present a vision for trustworthy EDT engineering grounded in an `involving before evolving' paradigm: rapidly involving stakeholders through a working prototype before evolving toward federation and full interoperability. Our three-stage approach combines foundation models for rapid prototyping, an ontological backbone for federated interoperability, and observability tooling for stakeholder trust. We ground our vision in an ongoing collaboration with Michelin, a multinational manufacturer, where an initial prototype has helped support stakeholder buy-in.

arXiv:2609.13072v1 Announce Type: cross Abstract: Fine-grained control of language model behaviors (e.g., steering) is among the more actionable outcomes of interpretability research. For binary concepts such as refusal, a single direction in activation space often suffices for steering. However, many concepts are not binary: Animals and Countries contain many subcategories, each with multiple instances. For these concepts, the search space over possible representation geometries is far larger than for binary concepts; it is thus not clear what geometries are most appropriate, nor what methods are most effective at recovering them. In this work, we introduce MAxBench, a geometry-agnostic evaluation framework for multinomial concept representations based on sampling from the recovered concept representation. We use MAxBench to compare 10 localization methods (covering 5 geometry types) across 6 concepts and 4 models. Using this framework, we find that (i) affine subspaces steer more reliably and have greater recall than rank-one or linear subspaces; (ii) much of this advantage is due to better non-zero offsets rather than the choice of bases; (iii) manifold steering is competitive with the best methods when applicable; and (iv) no method consistently outperforms prompting, in alignment with prior findings on binary concepts. These findings underscore the importance of expanding the scope of interpretability research and meta-evaluation to concepts with more varied structure.

arXiv:2506.19960v2 Announce Type: replace-cross Abstract: Reliable description of bond breaking remains a major challenge for quantum chemistry due to the multireference character of the electronic structure in dissociating species. Multireference methods in particular suffer from large computational cost, which under the normal paradigm has to be paid anew for each system at a full price, ignoring commonalities in electronic structure across molecules. Quantum Monte Carlo with deep neural networks uniquely offers to exploit such commonalities by pretraining transferable wavefunction models, but all such attempts were so far limited in scope. Here, we bring this paradigm to fruition with Orbformer, a transferable wavefunction model pretrained on 22,000 equilibrium and dissociating structures that can be fine-tuned on unseen molecules reaching an accuracy-cost ratio rivalling classical multireference methods. On established benchmarks as well as more challenging bond dissociations and Diels-Alder reactions, Orbformer is the only method that consistently converges to chemical accuracy (1 kcal/mol). This work turns the idea of amortizing the cost of solving the Schr\"odinger equation over many molecules into a practical approach in quantum chemistry.

arXiv:2602.21223v2 Announce Type: replace-cross Abstract: It is not only what we ask large language models (LLMs) to do that matters, but also how we ask them. Phrases like ``This is urgent'' or ``As your supervisor'' can shift model behavior without altering task content. We study this effect as pragmatic framing, contextual cues that shape directive interpretation rather than task specification. While prior work exploits such cues for prompt optimization or probes them as security vulnerabilities, pragmatic framing itself has received comparatively little attention as a target of controlled measurement in instruction following. To support its systematic study as a measurable property, we introduce a framework that combines three components: directive-framing decomposition separating framing context from task specification; a taxonomy organizing 400 instantiations of framing into 13 strategies across 4 mechanism clusters; and priority-based measurement that quantifies influence through observable shifts in directive prioritization. Evaluating five open-weight LLMs across different families and scales, we find that pragmatic framing produces systematic shifts in directive prioritization, and the effectiveness ranking of different strategies proves highly consistent across models. This reveals that susceptibility to pragmatic framing is a structured behavioral property of instruction-tuned systems. Measuring this susceptibility is a prerequisite for any deliberate response to it, and this work provides the framework to do so.

arXiv:2603.02638v2 Announce Type: replace-cross Abstract: Machine learning for mobile network analysis, planning, and optimization is often limited by the lack of large, comprehensive real-world datasets. This paper introduces the Vienna 4G/5G Drive-Test Dataset, a city-scale open dataset of georeferenced Long Term Evolution (LTE) and 5G New Radio (NR) measurements collected across Vienna, Austria. The dataset combines passive wideband scanner observations with active handset logs, providing complementary network-side and user-side views of deployed radio access networks. The measurements cover diverse urban and suburban settings and are aligned with time and location information to support consistent evaluation. For a representative subset of base stations (BSs), we provide inferred deployment descriptors, including estimated BS locations, sector azimuths, and antenna heights. The release further includes high-resolution building and terrain models, enabling geometry-conditioned learning and calibration of deterministic approaches such as ray tracing. To facilitate practical reuse, the data are organized into scanner, handset, estimated cell information, and city-model components, and the accompanying documentation describes the available fields and intended joins between them. The dataset enables reproducible benchmarking across environment-aware learning, propagation modeling, coverage analysis, and ray-tracing calibration workflows.

arXiv:2603.07357v3 Announce Type: replace-cross Abstract: Latent generative models have emerged as powerful priors for solving inverse problems. These models typically represent a class of natural signals at a single, fixed complexity, governed by the latent dimensionality. This can be limiting: depending on the problem, a latent dimensionality that is too small may result in high representation error, while one that is too large may overfit to noise. We develop tunable latent priors for diffusion models, normalizing flows, and variational autoencoders, leveraging nested dropout. Across tasks including compressed sensing, inpainting, denoising, and phase retrieval, we show empirically that tunable priors consistently achieve lower reconstruction errors than fixed-complexity baselines. In the linear denoising setting, we derive the optimal complexity in closed form, showing how it depends on the noise level and the signal spectrum. This work demonstrates the potential of tunable latent generative priors and motivates both the development of supporting theory and their application across a wide range of inverse problems.

arXiv:2606.07167v2 Announce Type: replace-cross Abstract: Meaningful multilingual evaluation must test models in the target language and educational context. Urdu, spoken by more than 230 million people, lacks a broad MMLU-style benchmark built from native educational sources. We introduce UrduMMLU, a benchmark of 26,389 Urdu MCQs across 26 subjects and five domains, collected from native Urdu MCQ banks and public examination PDFs. Unlike translation-based benchmarks, UrduMMLU combines academic subjects with content specific to Urdu and regional education. We label the exam-derived portion through dual human annotation with strict consensus filtering. We evaluate 30 LLMs under English and Urdu prompts, yielding 60 zero-shot evaluations, and further evaluate four open-source LLMs under multiple few-shot settings across both prompt languages. Gemini-3.5-Flash performs best, reaching 90.23% and 90.45% accuracy, while no other model exceeds 85%. The strongest open-source model trails by 7.78 and 9.12 points, and many models lose 25 to 40 points on Urdu-centered Humanities subjects compared with STEM. Few-shot prompting yields only modest gains. Results on UrduMMLU show that current LLMs have uneven Urdu knowledge, particularly for content grounded in the regional context.

arXiv:2608.15424v2 Announce Type: replace-cross Abstract: The rapid adoption of large language models has enabled the development of clinical multi-agent systems (MAS) capable of integrating multimodal patient data and supporting increasingly complex clinical decision-making. However, the deployment of these systems in real-world healthcare settings raises critical ethical concerns related to safety, fairness, accountability, transparency, and patient trust. While numerous organizations, including the World Health Organization, the National Academy of Medicine, and the FUTURE-AI consortium, have proposed ethical frameworks and governance principles for healthcare AI, these efforts remain largely conceptual. To address this challenge, we present ETHOS (Ethics and Trust through Hierarchical Oversight System), a modular ethics framework designed as a governance meta-agent that can be integrated with any existing multi-agent system without requiring changes to its underlying architecture. ETHOS translates stakeholder-informed ethical requirements into executable runtime oversight through a layered governance approach consisting of deterministic checks, contextual reviews, and a final ethics critic. These components continuously evaluate intermediate reasoning steps and final outputs, enabling the system to identify ethical risks, request revisions, or suppress responses that fail predefined safety and trustworthiness criteria. We demonstrate ETHOS within a hepatology clinical decision-support MAS. Results show that ETHOS improves decision reliability by detecting incomplete, inconsistent, or out-of-scope evidence and appropriately increasing abstention when safe recommendations cannot be supported. By embedding ethical governance directly into system operation, ETHOS provides a practical and auditable mechanism for transforming high-level AI ethics principles into deployable safeguards.

arXiv:2608.18138v2 Announce Type: replace-cross Abstract: In recent years, the rapid development of language models has transformed the field of Natural Language Processing through a wide range of applications. However, the development of language models has not progressed uniformly across all languages. In the case of the Portuguese language, there has recently been a growing effort by academia and companies to develop language models and create data resources for Portuguese. These efforts have resulted in the rise of an increasingly diverse ecosystem of language models for Portuguese. However, information on these models remains dispersed in scientific publications, technical reports, model repositories, and project documentation. This survey presents a systematic mapping study of language models developed for Portuguese, providing a comprehensive overview of the current state of the field. We map a total of 46 models, characterizing them by various aspects, including base model, architecture, computational resources, training datasets, licensing, code availability, data, and model weights. Furthermore, we analyzed the evolution and relationships among these models through a phylogenetic perspective, identified current research gaps and opportunities, and discussed future directions for the development of language models for Portuguese.