Ziwei Zhao

We define categorized information density as a per-class difficulty signal beyond instance frequency. Its Pearson correlation with AP reaches about -0.72 on Pascal VOC, roughly four times stronger than the instance-count baseline, exposing a long-overlooked bias in long-tailed detector training. A local-to-global covariance merging scheme keeps global information density exact during training and reduces per-class AP variance by up to 45.89% over EFL on LVIS v1.0.

A cultural-scope safety benchmark that treats norm scope as a first-class safety dimension. We build a dual-anchored annotation protocol across 49 countries and 9 cultural clusters (3,323 norms), define three scope error types (Inflation, Deflation, Displacement), and evaluate 21 frontier LLMs across 7 model families.

An evidence-grounded reconstruction pipeline for degraded exam-paper images. Contributions include benchmark construction and OCR-robustness ablations.

A coordination layer for agentic software engineering where a skill’s public view is limited to name, typed schemas, and trust tier, while the body (code and system prompt) is closure-captured and never crosses the wire. Released as an open-source Python SDK over XMTP; featured in awesome-agentic-patterns as the Black-Box Skill Invocation pattern.

A multimodal trajectory-prediction framework fusing visual context (ResNet-18) and motion history (LSTM) via cross-attention and gated fusion to output calibrated probabilistic forecasts. Split conformal calibration converts raw Gaussian uncertainty into statistically reliable risk bounds, improving Coverage@95 by 7 percentage points.

GGTPC addresses local training bias in federated prompt learning: the server reconstructs global geometric priors from per-class statistics without exchanging raw features. As a plug-and-play module across 7 mainstream FL algorithms, it yields +9.17 pp over FedAvg on CIFAR-100 at beta=0.01 where baselines collapse.