Story
arxiv_agent_systems_research · Nov 5, 2026 · paper
arxiv.orgNov 5, 2026
original source linked
In brief
Language models lose accuracy when a prompt contains non-evidential context: irrelevant statements from unrelated domains, or random off-distribution tokens, with no competing answer and no semantic overlap with the q...
Feed lens
eval