지능형 AI 시스템과
사람 중심의 가치를 연구합니다

Abstract: Large Language Models (LLMs) often suffer from severe memory footprint and latency degradation when reasoning over extended contexts (>64k tokens). In this paper, we propose AdaContext, an entropy-guided attention cache compression method that selectively preserves crucial reasoning traces while discarding redundant intermediate tokens. Comprehensive experiments across 8 benchmark suites demonstrate a 2.4x speedup and 46% memory reduction with virtually zero loss in reasoning fidelity.

Abstract: Aligning heterogeneous visual and linguistic representations requires flexible cross-modal grounding. We introduce DynCross, an adaptive sparse attention framework that calculates semantic affinity across modalities on the fly. The proposed framework significantly outperforms prior cross-attention baselines on VQA v2, MS-COCO Retrieval, and RefCOCO while lowering FLOPs by 34%.

Abstract: Parameter-efficient fine-tuning (PEFT) has democratized adaptation of massive foundation models. However, backpropagation on edge hardware remains constrained by peak SRAM limits. We formulate SpLoRA, introducing structured sparse low-rank factors that decouple gradient tracking across layers, achieving a 58% reduction in activation memory without sacrificing downstream task accuracy.

Abstract: Visual prompting offers an intuitive paradigm to steer pre-trained vision models for downstream perception. We present a self-supervised objective enabling prompt tokens to capture domain-invariant structural priors, dramatically improving zero-shot cross-domain segmentation on medical, satellite, and natural datasets.

Abstract: 본 특허는 자연어 처리 트랜스포머 모델의 셀프 어텐션 연산 시 토큰 중요도에 기반하여 연산 그래프를 동적으로 프루닝(Pruning)하고 온디바이스 메모리 대역폭 한계를 극복하는 고성능 추론 장치 및 방법에 관한 것입니다.