In briefSPADE: Soft-Prompt Method for Detecting AI-Generated Text
SummaryThis paper introduces SPADE, a framework for detecting AI-generated text that uses soft prompts. Unlike existing detectors that look only at the generated text, SPADE measures how well a prompt explains a text using pointwise mutual information. It trains one soft prompt per class and classifies a text by the prompt that best explains it. The authors report state-of-the-art performance on out-of-distribution datasets.
In briefReport on the 8th LSVOS Challenge's Mobile Unified Multimodal Understanding Track
SummaryThis report summarizes the Mobile Unified Multimodal Understanding track of the 8th LSVOS Challenge, held with ECCV 2026. The track requires a single model under 0.5B parameters and 8 GB peak memory to perform multi-concept image tagging, open-vocabulary object detection, and concise captioning. It describes the tasks, rules, evaluation protocol, final results, and the methods of the top three teams, all of which build on Florence-2-base. The findings suggest a compact prompt-based vision-language model can handle all three outputs when each task is designed carefully, while detecting novel categories remains the main bottleneck.
In briefBeyond Word Error Rate: A Failure Taxonomy for Live Dictation Speech Recognition
SummaryThe paper argues that speech recognition is usually compared by word error rate, which ignores the mistakes users must actually fix when dictating live text. The authors streamed multilingual test recordings to eleven systems and sorted each error into four levels: recognition, formatting, language handling, and audio dropped before the model. They report that systems with similar word error rates fail in very different ways, such as with whispered speech or mixed-language recordings.
Harras Mansoor, Muhammad Wisal, Hassan Bhatti et al.
In briefCADET: Generating Synthetic PCB Routing Data with Executable Demonstrations
SummaryThe paper introduces CADET, a method that jointly generates synthetic printed circuit boards and demonstrations built from electronic design automation (EDA) engine actions. It releases RouteDemos, a dataset of 100,000 synthetic boards routed with zero design rule violations, along with their demonstrations. A routing policy trained with behavior cloning and then reinforcement learning raised the rate of DRV-free complete routing on real test boards from 38.0% to 67.2% over an RL-only baseline. The dataset and generation code are released as open source.
In briefAutomated PD-L1 Tumor Scoring with Uncertainty-Based Triage in Lung Cancer
SummaryThis paper presents a three-stage framework for automatically calculating the PD-L1 tumor proportion score (TPS) in non-small cell lung cancer. It identifies tumor regions, counts PD-L1-positive and negative tumor cells with a dual-branch network, and uses an uncertainty-guided triage mechanism called TPS-Guard to flag slides for review. In a retrospective single-center study of 154 whole-slide images, the model's mean absolute error on 114 unseen slides was 6.96 percentage points, within the range of three pathologists. Model uncertainty correlated with TPS error and inter-pathologist disagreement.
