Route, Don't Fix: Regime-Dependent Decoding Correction and a Trajectory-Gated Router for Reliable Clinical LLM Answer Selection
ALTAS is introduced, which reads terminal entropy and late-layer linearity from one forward pass to choose per question between greedy decoding and late-layer trajectory correction and improves TruthfulQA over greedy at 3B and 8B.
Ze-Yu Dong, Benjamin Wang, Joyee W. Jin
· 0 citations