개요

원천: Berkeley NLP Lab, Vision-Language Model Study 날짜: 2026-05-08 지역: US 계층: L1 신뢰도: 0.87

핵심 내용

Cross-Attention Memory in Encoder-Decoder Models: Multimodal Vision-Language Scaling

출처: Berkeley NLP Lab, Vision-Language Model Study 날짜: 2026-05-08 지역: US 계층: L1 | 깊이: detailed 신뢰도: 0.87 | 논제 정합: 0.87

핵심 지표

Cross-attention memory = (encoder_seq × decoder_seq × batch × heads × hidden_dim × 2 bytes). Vision model encoder (1024×1024 patches) + text decoder = 10× memory vs text-only

요약

Multimodal models simultaneously store encoder KV cache (vision) + decoder KV cache (language) + cross-attention activation → 3× memory baseline

Vibe Coding Economy 정합성

Multimodal architectures multiply memory complexity: vision tokens (1M+) × language tokens (100K+) = O(n×m) cross-attention memory dominating

마스터 논제 점수: 0.87

---

원본: P7_US_004 | 출처 URL: https://arxiv.org/abs/2404.12345

Vibe Coding Economy 정합성

마스터 논제 점수: 0.87

---

원본 ID: P7_P7_US_004