Publications

Reinformed Dreamer: An Asymmetric World Model Efficiently Trained through Latent Guidance
Adrien Bolland
Daniel Ebi
Damien Ernst
Much like humans benefit from guidance while learning, reinforcement learning algorithms may benefit from additional supervision beyond rewa… (see more)rds. Leveraging additional information during training to learn better representations and behaviors has been the focus of asymmetric reinforcement learning. This learning paradigm has proven effective under partial observability when additional state information is available, but also under full observability when more refined state information is available. Focusing on model-based reinforcement learning, we study the effect of asymmetric learning on observation representations and on privileged information representations. First, we identify a limitation in the privileged information representations learned by an asymmetric model-based algorithm known as the Informed Dreamer. Then, we propose a novel asymmetric representation learning objective using latent guidance, resulting in a new algorithm called the Reinformed Dreamer. Experiments across several benchmarks show a more consistent improvement over Dreamer than previous asymmetric approaches.
Evaluating Communicative Belief Updates in Large Language Models via Implicature Recognition and Cancellation
Human language is driven by unspoken beliefs and belief updates, making these critical to model for successful communication between large l… (see more)anguage models (LLMs) and their users. In this paper, we evaluate the ability of LLMs to recognize unspoken beliefs made through implicatures and to understand their updates through implicature cancellation: the pragmatic phenomenon whereby an utterance's implied meaning is weakened or negated. We create the first expert-annotated implicature cancellation dataset, ImplicatureX, crowdsourced for human judgements of implicatures and their corresponding cancellations. We find that LLM belief update understanding lags behind that of humans, especially in more naturally-occurring scenarios. Additional control experiments suggest that successes in LLM belief updates may stem in part from a reliance on prior beliefs, and that failures in belief updates may depend on their type and on their form. Overall, our study suggests that current LLMs have not yet reached human-level understanding of unspoken beliefs and belief updates. Code and data are available at https://github.com/cesare-spinoso/ImplicatureX.
Evaluative Stance Toward Artificial Intelligence in High-Quartile Medical Journals (2021– 2026): Large-Scale LLM-Assisted Computational Content Analysis
Lixing Wang
Abstract Background Medical-AI publications do more than report technical performance; they also frame AI as beneficial, uncertain, or risky… (see more). How this evaluative stance has changed across the medical literature is not well characterized. Objective To characterize evaluative stance in published medical-AI discourse abstracts from January 2021 through April 2026 and examine variation over time, concern themes, failure mechanisms, specialties, first-author geography, and publication format. Methods We conducted an LLM-assisted computational content analysis of medical-AI abstracts from first- and second-quartile medical journals. Of 97,492 post-cutoff Q1/Q2 records entering the prefilter, 16,759 were retained as discourse or evaluative. Claude Sonnet 4.6 assigned 16,749 valid stance classifications using Alarm, Caution, Neutral, Cautious Optimism, and Advocacy. Annual analyses used 16,747 records dated 2021–2026. Critical stance was Alarm plus Caution and indexed evaluative scrutiny, rather than opposition or author psychology. Each LLM step was validated against blinded human coding by one author: prefilter Cohen κ=0.51, stance quadratic-weighted κ=0.79 (95% CI 0.72–0.84) for codable, in-scope records, specialty κ=0.75, and mechanism axes κ=0.84 (model type) and κ=0.57 (failure mode). Results Advocacy declined from 2.9% in 2021 to 0.6% in partial 2026, while Cautious Optimism remained the majority stance. Among 16,749 valid classifications, 30.8% were critical. Critical share increased from 25.4% to 32.6%, a 7.25-percentage-point increase based on unrounded estimates. Among critical records, patient safety remained the most prevalent concern. Hallucination/errors increased by 30.9 percentage points. Regulation declined by 22.0 percentage points and ethics/bias by 8.1 percentage points in prevalence share; these declines do not necessarily indicate lower publication counts. Within the hallucination/error theme, factual error was more common than fabrication. Fabrication estimates should be treated as an upper bound, because failure-mode agreement was moderate. Specialty patterns were heterogeneous. Critical rate was inversely associated with FDA cleared-device availability (Spearman ρ=−0.65, two-sided p=0.004), which does not measure adoption, deployment, maturity, or clinical use. First-author geography described publication metadata and discourse, not national attitudes or research quality. Reviews were the least critical and most favourable format. In exploratory forward-validation, 2 of 78 early Advocacy predictions were fully borne out, although the analysis was single-rater and retrieval-dependent. Conclusions Published medical-AI abstracts became modestly less promotional and more focused on specific errors and safety concerns. Unqualified promotion declined, but qualified favourable framing remained dominant, and the rise in critical stance was modest. Concern moved toward errors and patient safety, with factual error discussed more often than fabrication. These findings describe published discourse, not AI capability or whether the evaluations were correct.
Intrinsic and Triangulation-Agnostic Attention: A Simple and Powerful Approach for Learning on Meshes
This work proposes an adaptation of the attention mechanism for triangle meshes. The core observation is that endowing the attention mechani… (see more)sm with critical properties for learning over meshes -- intrinsicality and triangulation-agnosticism -- enables it to attain state-of-the-art results over several learning-based tasks in geometry-processing. The above is achieved by modifying the attention mechanism from the bottom up based on simple principles from geometry-processing. Namely, the quantities used within attention -- queries, keys and values -- are created by an intrinsic, triangulation-agnostic network, and treated as discretizations of continuous functions. From that, we devise an appropriate attention mechanism that operates over triangle meshes through standard FEM discretization of the resulting integrals of the above functions. Surprisingly, as far as we know, this straightforward approach has not been utilized for learning over meshes. Experiments show our method exceeds current state of the art, including both mesh-based architectures as well as point cloud transformers. Namely, we show significant improvements on several common benchmarks and tasks -- predicting canonical high-frequency signals; predicting deformations; computing dense correspondences, both between full shapes and partial ones; and predicting feature descriptors.
Moral Hazard in Multi-Agent Language Models
Cooperation can fail when socially valuable effort is costly, hard to observe, and benefits mainly someone else. Building on Holmstrom's mod… (see more)el of moral hazard in teams, the Dialogue Moral Hazard Game instantiates this hidden-action structure as a textual environment for language agents. An agent chooses between keeping an immediate local reward and paying a query cost to reveal a hidden safety fact that helps another agent's downstream decision. We evaluate fourteen open-weight and four frontier models using measures of information acquisition, communication, downstream use, and team success. In matched 3,015-decision-per-model experiments, GPT-5.6 Sol, Claude Opus 4.8, and Nemotron-3 Ultra track the derived private-share boundary across nine query costs, with mean absolute errors of 0.013, 0.030, and 0.024. Muse Spark 1.1 responds directionally, whereas Fable 5 remains query-saturated. SFT, RLOO, SFT+RLOO, and GEPA produce heterogeneous mechanism changes. GEPA raises Muse team success from 22.2% to 100.0% while reducing query use from 51.1% to 0.3%. Frozen-prompt interventions show that this success depends on a learned rank-label mapping rather than direct revelation: changing the mapping reduces team success from 100.0% to 12.5% and then 0.0%. We introduce CREDIT (Counterfactual Replay for Evidence-Driven Information Transfer), a mechanism-aligned multi-agent prompt-optimization algorithm that uses matched hidden-state twins and total-action replay to reward robust causal contribution rather than query frequency. Across five models and multiple seeds, CREDIT preserves query-mediated behavior while revealing model-specific acquisition and downstream-use bottlenecks. Optimization can reach the same aggregate outcome through direct revelation or a learned effective information structure, motivating mechanism-level evaluation and optimization rather than team success alone.
Pan-Cancer Gene Expression Predicts Per-Cancer Radiation Therapy Response
Background and purpose: Although radiotherapy response involves biological processes shared across cancers, most biomarker studies rely on s… (see more)mall, disease-specific sample sizes that limit the reliability of prognostic models. This study evaluated whether pan-cancer data can increase the information available for estimating genomic associations with progression after radiotherapy, based on Molecular Signatures Database hallmark gene sets applied to The Cancer Genome Atlas patients. Methods and materials: Cross-cancer predictive potential was evaluated in cancer types with at least 100 patients who received radiotherapy and had information on gene expression, progression, age, stage, and gender: breast invasive carcinoma (BRCA, n=553), cervical squamous cell carcinoma (CESC, n=183), head-neck squamous cell carcinoma (HNSC, n=320), low-grade glioma (LGG, n=316), lung adenocarcinoma (LUAD, n=104), skin cutaneous melanoma (SKCM, n=125), thyroid cancer (THCA, n=326), and uterine corpus endometrial carcinoma (UCEC, n=257). Gene set expression scores were calculated at the single-sample level. For each evaluation cancer type, scores from all other types were used to fit pan-cancer Cox proportional hazards models. Linear predictions from these models were evaluated in per-cancer prognostication via Kaplan Meier curves, Cox proportional hazards (CoxPH) estimates, including adjustment for age, stage, and gender, and optimism-corrected C-statistic and Nagelkerke pseudo-R 2 . Results: In Kaplan Meier analysis, pan-cancer scores predicted progression in all cancer types, with varying statistical significance. Univariable CoxPH models also demonstrated consistent association with progression, again with varying significance. Associations persisted when adjusted for clinical variables, and model performance metrics indicated the potential to improve per-cancer prognostication. Associations were particularly strong for BRCA, CESC, HNSC, LGG, and LUAD, with statistically insignificant results for SKCM, THCA, and UCEC. Conclusion: Pan-cancer gene expression data can predict per-cancer progression after radiotherapy, with variable strength across cancer types. These findings support the use of cross-cancer genomic information to address sample size limitations that affect radiotherapy biomarker development.
Backend-Aware Graph Learning for Denoising Outcome Distributions in Quantum Program Testing
Ning Ma
Heng Li
Testing quantum programs on NISQ (Noisy Intermediate-Scale Quantum) backends is challenging because the noise disturbs outcome distributions… (see more) and can affect pass/fail decisions. We present Q-BRIDGE, a graph learning-based approach that converts noisy observations into denoised distributions suitable for oracle-based verification. Q-BRIDGE uses a graph transformer architecture to encode a transpiled quantum circuit, capturing the characteristics of its gates and their connectivity; the physical backend information is encoded together with the logical structure of the circuit. An additional conditioning layer, based on FiLM (Feature-Wise Linear Modulation), takes the encoding as input and integrates noisy observations to produce denoised outcomes. We evaluate Q-BRIDGE on 23 IBM noise backends and 6 circuit families representative of practical workloads. In the first setting, we train a separate Q-BRIDGE model for each backend; in the second setting, we train a single general model shared across all backends. Across both settings, Q-BRIDGE outperforms the state-of-the-art baseline in noise mitigation by a large margin. In testing scenarios with noisy executions, Q-BRIDGE achieves 93.97%-94.90% precision and 82.50%-83.51% recall in detecting bug-induced test failures, significantly outperforming the state-of-the-art baseline. These results indicate that considering the graph structure of the transpiled circuits and the physical characteristics of specific quantum backends is a practical route to more reliable noise-aware quantum program testing.
Online Policy Evaluation for MDPs with Dynamic UBSR Measures
Developing efficient function-approximation methods for policy evaluation is a fundamental challenge in risk-aware reinforcement learning. E… (see more)xisting approaches either focus on restrictive classes of risk measures or rely on access to a simulator, limiting their applicability in fully online settings. In this work, we propose computationally efficient online learning algorithms for policy evaluation in Markov decision processes (MDPs) with dynamic utility-based shortfall risk (UBSR) measures under linear function approximation. Specifically, we introduce the UBSR-TD algorithm, establish conditions under which it converges almost surely, and develop several variants designed to accelerate convergence. Our formulation shows that existing policy evaluation algorithms for risk-neutral MDPs can be readily adapted to dynamic UBSR settings by incorporating a loss function into the temporal-difference error. Numerical experiments support our theoretical findings, and an application to a perishable inventory management problem with shelf-life uncertainty demonstrates the practical effectiveness of the proposed methods.
Modern Convolutional Design Improves Uterine MRI Segmentation, while nnU-Net Remains Most Robust Across Heterogeneous Datasets
Daniel A. Di Giovanni
Akiyo Takada
Evan McNabb
Jérémy Dana
Hajime Yokota
Takahiro Tsuboyama
Rita Zakarian
James Man Git Tsui
Caroline Reinhold
Abstract Purpose To evaluate how segmentation architecture and dataset-adaptive configuration influence uterine MRI segmentation across hete… (see more)rogeneous benign and malignant tasks. Methods U-Net, Swin-UNETR, and MedNeXt were compared with nnU-Net as a self-configuring reference across T2-weighted MRI datasets: public multiclass UMD anatomy/fibroid segmentation (n=300), institutional endometrial cancer tumor segmentation (n=206), and institutional uterine mass lesion segmentation (n=234). A relabeled external UMD-style cohort (n=12) assessed domain shift. Models used fixed partitions, fold ensembling, Dice, HD95, ASSD, volume error, and paired bootstrap comparisons with Holm correction. Results MedNeXt was the strongest manually controlled architecture. nnU-Net achieved the highest performance on all internal datasets and external testing. Macro-Dice reached 0.761, 0.746, and 0.814 for nnU-Net on UMD, endometrial cancer, and uterine mass datasets, respectively, versus 0.722, 0.726, and 0.789 for MedNeXt. The nnU-Net-MedNeXt gap was largest for multiclass UMD segmentation and smaller in binary tasks. External testing degraded all models; nnU-Net remained highest (0.542), followed by MedNeXt (0.490), U-Net (0.396), and Swin-UNETR (0.287). Conclusions Uterine MRI segmentation performance depended on task, architecture, and evaluation domain. MedNeXt supported modern convolutional design as a strong manual baseline, but nnU-Net remained the most robust overall, emphasizing the importance of dataset-adaptive configuration and external validation.
1D Pre‐Acquisition Navigator Correcting Respiratory‐Induced Field Fluctuations in Multi‐Echo Gradient‐Echo Imaging of the Thoracic Spinal Cord
Alicia E. Cronin
Alexandre D’Astous
Nathan Williams
Antoine Guénette
Aimee Salakhov
Seth Stubblefield
Colin D. Mcknight
Lipika Narisetti
Subramaniam Sriram
Seth A. Smith
Ryan K. Robison
Guillaume Gilbert
Julien Cohen‐Adad
Kristin P. O’Grady
PURPOSE: In the spinal cord (SC), multi-echo gradient echo (ME-GRE) increases gray (GM) and white matter (WM) contrast and improves sensitiv… (see more)ity to lesions in people with multiple sclerosis (pwMS). However, SC ME-GRE is susceptible to breathing-induced field fluctuations, causing ghosting artifacts and signal loss. Recent work introduced a 1D phase navigator following the last echo to measure field variations; however, susceptibility to phase wrapping increases at longer echo times. We propose a 1D phase navigator preceding the first echo, reducing phase accumulation and eliminating the need for respiratory monitoring. METHODS: ME-GRE data covering the lower (T9-T12 vertebrae) and upper (T4-T8 vertebrae) thoracic SC were acquired in 20 healthy volunteers and 3 pwMS at 3T. Standard and navigator-corrected images were acquired in the same acquisition. To evaluate image quality, WM and GM signal-to-noise ratio (SNR), WM/GM contrast-to-noise ratio (CNR), and background ghosting signals were measured and compared between the two reconstructions. Both were blindly assessed for artifacts, structural delineation, and diagnostic confidence in pwMS. RESULTS: Navigator correction significantly increased GM and WM SNR and CNR, reduced posterior ghosting across both thoracic regions, and significantly reduced artifacts while increasing structural delineation. Preliminary evaluation in three pwMS showed consistent improvements in artifact mitigation, structural delineation, and lesion conspicuity with navigator correction, providing proof-of-concept for potential clinical application. CONCLUSION: A 1D navigator prior to the first echo reduces ghosting and improves thoracic SC image quality without respiratory monitoring. This approach could improve the diagnostic value and enhance the reliability of thoracic SC ME-GRE.
The curriculum effect in visual learning: The role of readout dimensionality
Christopher C. Pack
Generalization of visual perceptual learning (VPL) to unseen conditions varies across tasks. Previous work suggests that training curriculum… (see more) may be integral to generalization, yet a theoretical explanation is lacking. We propose an explanatory theory of visual learning generalization and curriculum effects by leveraging an artificial neural network (ANN) model of VPL in comparison with humans. We found that easy-to-hard sequential training improved generalization in both humans and ANNs. However, when easy and hard conditions were interleaved, humans and ANNs showed different behaviours: while ANNs performed worse than with sequential training, humans maintained good performance but with large inter-individual variability. Investigating ANN models trained with different curricula, we demonstrated that models relying on low-dimensional neural populations showed superior generalization. This readout subspace dimensionality was directly determined by curriculum: learners who learned from easy tasks early formed lower-dimensional subspaces and generalized better. Our theory provides a mechanistic framework linking curriculum design to VPL generalization through neural population dimensionality.
Towards Trustworthy Physical AI: From Theory to Practice Across Life Cycle
Wang Yang
Hongxuan Liu
Xinghui Xu
Arjun Menon
Xiaoran Cai
Yunyu He
Alex Tarvo
Jingzong Zhou
Mengzhong Ma
Xinpeng Wei
Yi Yu
Shaobo Wang
Cheng Peng
Aoran Jiao
Alexei Korolev
Yanyan Zhang
Kai Ye
Xinpeng Li
Chengquan Guo
Jingjing Fu … (see 23 more)
Nicholas Bai
Yongjun He
Junru Ren
Silei Ren
Mohamad Louai Shehab
Keshu Cai
Nathaniel Dennler
Traian Tus
Gaoyue Zhou
Abdullah Garra
Mason Nakamura
George Ortiz
Marius Urbonas
Lars Johannsmeier
Rohit Sharma
Suraj Deshmukh
Michael Spranger
Vinayak Gupta
Xu Chen
Ashis G. Banerjee
Yuxiang Feng
Peng Qi
Physical AI refers to AI systems that understand, reason about, and act in accordance with the physical world and its underlying laws, dynam… (see more)ics, and constraints. Unlike conventional AI systems, physical AI interacts continuously with uncertain physical environments, and its actions produce consequences that are physically irreversible. As existing trustworthy AI frameworks have been developed primarily for digital AI systems, they do not fully capture the distinctive challenges of physical AI, such as physical safety, cyber-physical security, and physical manufacturing process. To address this gap, we present a survey of trustworthy physical AI principles. First, we characterize the core capabilities and challenges of physical AI. Second, we examine the role of physics in AI. Third, we trace the end-to-end physical AI life cycle across five core stages and introduce Trustworthy Physical AI Operationalization (T-PAIO). Fourth, we develop the Trustworthy Physical AI (T-PAI) framework, a theoretical framework that organizes key trustworthiness principles and provides a foundation for governing trustworthy physical AI systems.