3,996
cited sources scored across 12 AI systems
Proposed and built the evaluation pipeline for SourceBench, a NeurIPS 2026 benchmark of how well AI search retrieves sources.
← All work 
WukLab
Applied AI Researcher · UCSD · 1+ year · Sep 2025 – Nov 2026
Evaluating whether AI answers cite quality web sources. The work behind SourceBench, accepted at NeurIPS 2026.
My work spanned
- Research Proposal
- AI Evaluation
- LLM-as-Judge Evaluation
- Experiment Design
- Prompt Design
- Prompt Bias Detection
- Co-Authoring the Paper
- Google Search Study Lead
Key Highlights
12 of 12
systems' results saved from a bias bug
Caught a prompt-bias bug that would have invalidated every system's results, forcing a full re-run that made SourceBench publishable.
16% vs 55%
GPT-5 vs Tavily overlap with Google's top results
Led the Google vs. AI-search retrieval study and built the Google collection pipeline, showing GPT-5 cites sources Google rarely ranks at the top.
DeepSeek
controlled retrieval vs. reasoning experiment
Designed an experiment separating the effect of retrieval quality from model reasoning, showing better sources can narrow the gap between reasoning and non-reasoning models.
What We Concluded
- The best AI search checks who is behind a source. GPT-5 scored highest overall because it picked sources with clear owners, credible authors and accurate facts, not just pages that matched the question’s keywords.
- Good sources can matter more than reasoning. In our DeepSeek experiment, a simpler non-reasoning model given high-quality search results beat a reasoning model given weaker ones.
- AI search finds sources Google buries. Only 16% of GPT-5’s sources appeared in Google’s first five pages of results, compared with 55% for Tavily, the lowest-scoring system.
- Answer-only benchmarks miss what people care about. AI systems often give correct answers while citing outdated, biased or untrustworthy pages. Signals like author credibility and page clarity shape whether people trust an answer, yet most benchmarks ignore them.