Tag
Presents TEXAS, a method for downstream adaptation of Mixture-of-Experts LLMs that discovers task-relevant experts via correctness-conditioned activations and applies token-level supervision allocation, improving performance across multiple benchmarks.