Huang, S. (CSE) – From Hallucination Detection to Self-Improvement: Towards Reliable LLMs through Fine-Grained Feedback

Large language models can produce unsupported claims in short phrases embedded in otherwise accurate responses. This PhD proposal investigates how to pinpoint and reduce these hallucinations using span-level supervision. I introduce SHAM-Bench, a benchmark spanning six generation tasks, and present a detector that uses retrieved evidence to identify hallucinated spans. The detector achieved the highest average rank in the SemEval 2025 multilingual hallucination detection challenge across 14 languages. In summarization experiments, I find that unlikelihood training is the most effective of the mitigation methods tested. In future work, I will investigate how reinforcement learning with verified reasoning steps can improve knowledge recall and whether jointly training an agent and a trajectory verifier enables self-improvement on agentic task performance.
Event Host: Sicong Huang, PhD Student, Computer Science & Engineering
Advisor: Ian Lane
Zoom: https://ucsc.zoom.us/j/99455562030?pwd=ztbMAYTYNwcbLoYuTnp9kxEtTovVOI.1
Passcode: 012478