R-DEIM Net achieves paraphrase detection with 76M parameters and rationale generation
September 25, 2026
R-DEIM Net utilizes a dual-expert architecture combining multi-scale 2D convolutions with a Flan-T5-small decoder to generate human-readable rationales. This 76M-parameter model avoids full re-encoding by pooling decoder hidden states as features for classification.
HOW THIS AFFECTS YOU
●
builderYou can deploy smaller, interpretable models for paraphrase detection tasks without the latency of full LLMs.
●
researcherYou can explore using decoder hidden states as auxiliary features for efficient classification tasks.