Shiyuan (Eric) Zhou 周世源

SALUS detects annotation errors in natural-language-to-SQL benchmarks. SQL generated by multiple LLM agents provides complementary weak supervision signals, which are combined to identify high-confidence training examples without human ground truth. A learned decision model then uses query features and agent reliability to audit reference annotations. Evaluations on BIRD and Spider reveal substantial annotation errors and their effects on leaderboard reliability.

MADRAG combines multi-agent debate with retrieval-augmented generation for analytic essay scoring without task-specific training. An Advocate identifies strengths, a Skeptic critiques weaknesses, and a Judge uses rubric-aligned exemplar essays to calibrate the final score. Experiments show that retrieval improves calibration while debate strengthens reasoning on higher-level writing traits, outperforming prompt-based baselines and approaching supervised systems.

WorkLoadWeaver constructs synthetic workload specifications from query templates and optional domain assumptions. Its multi-agent pipeline models user personas, temporal patterns, transitions between queries, within-template refinements, and event-driven workload shifts. The pipeline supports both automated generation and human checkpoints for reviewing assumptions, adding constraints, and refining the resulting workload.

HEPHA is a mixed-initiative image labeling tool that elicits domain experts’ knowledge through inductive logic learning. Starting from a small set of labeled images, it infers rules composed of visual predicates and applies them to unlabeled data. Experts can then correct labels or edit the rules directly. A user study evaluates this collaboration in both specialized and general image labeling tasks.

DOSA aligns Stack Overflow posts with official documentation to help developers navigate complementary software resources. Weak supervision adapts a language model to the structure of a documentation site, while constrained decoding restricts its output to existing documentation categories and sections. Experiments on Python and Flask documentation show the promise of this approach for low-resource software engineering tasks.

Rapid combines pretrained computer vision models with inductive logic learning to infer interpretable image labeling rules. Domain experts supply a small set of labeled examples, and an active learning strategy selects informative samples for further annotation. Across four tasks, including specialized domains, Rapid achieves strong labeling accuracy with only 12 to 39 labeled samples.