TY  - RPRT
TI  - Test-time RL alignment exposes task familiarity artifacts in LLM benchmarks
AU  - Kun Wang
AU  - Reinhard Heckel
PY  - 2026
UR  - https://arxiv.org/abs/2603.12875
ID  - 2603.12875
ER  -