Xinpeng Wang

CS PhD @CS_UVA | Advised by Prof. @zhezeng0908

[Xinpeng Wang @xpwang633]

Read the full paper for more details.
📰Paper: https://t.co/TiDimHvub3
Huge thanks to my collaborators and advisor for their guidance and support throught out this project: William X. Cao, @andrewgwils, @zhezeng0908.

[Xinpeng Wang @xpwang633]

🧵Excited to introduce our work at ICML 2026!
Hallucination signals condense on intermediate layers of LLMs. But the best layer varies across datasets and models.
Can we design an efficient & principled criterion to automatically identify the most informative layer?🤔

[Xinpeng Wang @xpwang633]

Bonus: stop probing at the last token — representations there are often degraded by end-of-sequence noise (repetition, semantic drift, inconsistent continuations). We propose a simple first-sentence truncation (FST) that removes such noise. It boosts every baseline we tested 👏