探索全球前沿学术脉络

A blinded, counterbalanced rater design for evaluating AI-assisted summarisation of tertiary clinical genomics reports: methodology of the QNOMX-VHIR-CPSP-001 Phase 1 study

Background. Tertiary clinical genomics reports condense layered molecular findings into documents that treating oncologists must read, translate, and act upon; manual summarisation of these reports is time-consuming and variable. Tools that assist summarisation and translation into local languages are emerging, yet the field lacks an agreed methodology for evaluating such tools before any downstream clinical use. The appropriate first endpoint is fidelity of the generated summary to its source report, assessed by qualified human raters under blinded scoring, not downstream variant classification. Methods. QNOMX-VHIR-CPSP-001 Phase 1 is a single-site, non-interventional clinical performance study conducted at Vall d'Hebron Institut de Recerca (VHIR) under ISO 20916:2019 as a Clinical Performance Study Protocol. De-identified tertiary cancer genomics reports from pediatric oncology cases are summarised by the AI-assisted summarisation system under evaluation and, in parallel, by the standard manual workflow. Qualified raters score both summary types against the source genomics report using the Quality Summary Index (QSI), a six-dimension, five-point rubric adapted from the Provider Documentation Summarization Quality Instrument, under a blinded, counterbalanced, two-period crossover with a minimum fourteen-day washout. Two co-primary composite endpoints, content and presentation, are analysed for non-inferiority under a Bayesian hierarchical model, with a frequentist linear mixed model as the convergence check. Inter-rater reliability is reported as Krippendorff's ; a Monte-Carlo power analysis of the fixed clustered design is pre-specified. Discussion. The design isolates summarisation quality from clinical decision-making by scoring both summary types against the same source report under blinding, counterbalancing, and a fourteen-day washout. Conclusion. The QSI rubric, the counterbalanced crossover, and the pre-specified Bayesian primary with frequentist convergence check define a replicable protocol for early-stage evaluation of AI-assisted summarisation in tertiary genomics reporting; observed variance components will inform sample-size determination for Phase 2.

The Reward Was in Your Data All Along: Correcting Flow Matching with Discriminator-Guided RL

Score- and flow-matching models often rely on preference-based reinforcement learning for two purposes: aligning with subjective preferences and, surprisingly, recovering properties such as visual realism and coherent object structure that matching-based training is intended to learn from the data itself. We argue that this reflects a structural mismatch. Matching losses measure $\ell_2$ regression error on the velocity or score field under training-time marginals, a proxy poorly aligned with the visual and semantic properties that determine sample quality at inference. Given a reward aligned with these properties, RL sidesteps the mismatch by evaluating the model on its own samples and following the reward landscape directly. The challenge is to obtain such a reward without relying on human preferences, which are expensive and conflate data realism with annotator inclinations. We propose Discriminator-Guided RL (DRL). DRL trains a discriminator to separate data from base-model samples in a pretrained representation space and uses its logit as the reward in KL-regularized RL. The pretrained space restricts the discriminator to perceptually meaningful directions, and the logit estimates the log-likelihood ratio between data and model, which is the optimal reward for targeting the data distribution. Across SiT, JiT, REPA, and RAE, DRL reduces guidance-free FID (e.g., $9.38 \to 2.62$ on SiT) and semantic-space FD (e.g., $88.2 \to 19.3$ on DINOv3 for SiT), with consistent gains across all backbones, and improves human-preference rewards without training on them. It also yields a better Pareto frontier between preference reward and image fidelity under subsequent preference-based post-training, increasing alignment while reducing low-level artifacts such as oversaturation and excessive brightness.

Grounding Computer Use Agents on Human Demonstrations

arXiv:2511.07332v2 Announce Type: replace-cross Abstract: Building reliable computer-use agents requires grounding: accurately connecting natural language instructions to the correct on-screen elements. While large datasets exist for web and mobile interactions, high-quality resources for desktop environments are limited. To address this gap, we introduce GroundCUA, a large-scale desktop grounding dataset built from expert human demonstrations. It covers 87 applications across 12 categories and includes 56K screenshots, with every on-screen element carefully annotated for a total of over 3.56M human-verified annotations. From these demonstrations, we generate diverse instructions that capture a wide range of real-world tasks, providing high-quality data for model training. Using GroundCUA, we develop the GroundNext family of models that map instructions to their target UI elements. At both 3B and 7B scales, GroundNext achieves state-of-the-art results across five benchmarks using supervised fine-tuning, while requiring less than one-tenth the training data of prior work. Reinforcement learning post-training further improves performance, and when evaluated in an agentic setting on the OSWorld benchmark using o3 as planner, GroundNext attains comparable or superior results to models trained with substantially more data,. These results demonstrate the critical role of high-quality, expert-driven datasets in advancing general-purpose computer-use agents.