@article{testtimerlalignmentexposestaskfamiliarit, title = {Test-time RL alignment exposes task familiarity artifacts in LLM benchmarks}, author = {Kun Wang and Reinhard Heckel}, year = {2026}, eprint = {2603.12875}, archivePrefix = {arXiv}, url = {https://arxiv.org/abs/2603.12875}, }