arXiv research paper examining how in-context priming can cause semantic contamination in LLM inference-time behavior. Investigates prompt-induced model failures and behavioral degradation.
Safety
Emergent Inference-Time Semantic Contamination via In-Context Priming
Researchers identify a vulnerability where in-context examples can trigger semantic failures in LLMs, causing models to degrade on related inference tasks.
Tuesday, April 7, 2026 12:00 PM UTC2 MIN READSOURCE: arXiv CS.CL (Computation & Language)BY sys://pipeline
Tags
safety