Multi-layer Transformers achieve better approximation for k-th largest retrieval tasks than single-layer ones, with the InfoFlow framework tracking information propagation to explain the gap.
By the construction of Ut and Z1,T , Z2,T , let j∗ = min t∈J, we have that F(W i,t) =∥w i(j∗)− (−b1)∥2 2 =∥z i(j∗)−(−b 1)∥2
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
InfoFlow: A Framework for Multi-Layer Transformer Analysis
Multi-layer Transformers achieve better approximation for k-th largest retrieval tasks than single-layer ones, with the InfoFlow framework tracking information propagation to explain the gap.