Path-Lock Expert uses two mode-specific MLP experts per decoder layer with a deterministic router to cleanly separate think and no-think modes, improving no-think accuracy and conciseness on math benchmarks while keeping think performance intact.
The dense model incurs excess loss ∆conflict :=L dense(β⋆ dense)−L ⋆ PLE = 1 2 ∑ r∈{0,1} πr(β⋆ dense −β ⋆ r )⊤Hr(β⋆ dense −β ⋆ r )≥0
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Path-Lock Expert: Separating Reasoning Mode in Hybrid Thinking via Architecture-Level Separation
Path-Lock Expert uses two mode-specific MLP experts per decoder layer with a deterministic router to cleanly separate think and no-think modes, improving no-think accuracy and conciseness on math benchmarks while keeping think performance intact.