Weicong Chen

Current approaches to LLM safety rely on a brittle pattern of identifying and blocking known threats via guardrails. This paper introduces Trust The Typical (T3), a framework that reframes safety as an out-of-distribution detection problem, learning the distribution of acceptable prompts in a semantic space and flagging significant deviations as potential threats. Unlike prior methods, T3 requires no training on harmful examples yet achieves state-of-the-art performance across 18 benchmarks spanning toxicity, jailbreaking, multilingual harms, and over-refusal—reducing false positive rates by up to 40× relative to specialized safety models. A single model trained on safe English text transfers effectively to over 14 languages without retraining.

Curating high-quality, domain-specific datasets is a major bottleneck for deploying robust vision systems. This paper introduces Labeling Copilot, the first data curation deep research agent for computer vision, powered by a large multimodal language model that uses multi-step reasoning to execute specialized tools across three core capabilities: Calibrated Discovery for sourcing in-distribution data from large repositories, Controllable Synthesis for generating rare-scenario data with robust filtering, and Consensus Annotation for producing accurate labels via a novel multi-model consensus mechanism. On the dense COCO dataset, the Consensus Annotation module achieves an annotation mAP of 37.1%, and on Open Images it discovers 903 new bounding box categories.

Existing log anomaly detection methods are often slow, dependent on error-prone parsing, and use unrealistic evaluation protocols. This paper introduces K4 (Knowing the Unknown by Knowing only the Known), a fully unsupervised, parser-independent framework that transforms arbitrary log embeddings into compact four-dimensional descriptors—Precision, Recall, Density, Coverage—using efficient k-nearest neighbor statistics. Under a realistic online chunk-based evaluation protocol, K4 achieves state-of-the-art AUROC of 0.995–0.999 across HDFS, BGL, and Thunderbird datasets, with training under 4 seconds and inference as low as 4 μs.