BROWSE
Detail
Conference Paper
효율적인 체화 질의 응답을 위한 에피소드 수준 멀티모달 KV Caching
Cited - time in 
Share 
Share
- Issue Date
- 2026-02
- Citation
- 한국로봇학회 종합 학술 대회 2026, pp.129-131
- Publisher
- 한국로봇학회
- Language
- Korean
- Type
- Conference Paper
- Abstract
- Embodied Question Answering (EQA) requires agents to sustain a representation of the world while answering multi-turn queries in real time. A key challenge is how to maintain and update this world model efficiently under resource constraints. Existing approaches repeatedly re-encode visual inputs or apply retrieval-augmented generation, both of which introduce latency that limits interactive use. We propose an episode-level multimodal KV cache that is constructed once from uniformly sampled frames and reused across all queries in the same episode. This cache serves as a lightweight multimodal memory that reduces redundant computation while pre serving relevant context. On the openEQA benchmark, our method achieves up to an 82% reduction in total question-answering time compared to naïve multi-image inference, with only a modest drop in accuracy. These findings demonstrate that reusing an episode-level cache provides an effective mechanism for maintaining and updating world models to achieve efficient reasoning in EQA.
- Keyword
- Embodied Question Answering, KV Cache
- KSP Keywords
- Efficient reasoning, Existing Approaches, Multi-image, Question Answering, Relevant context, World model, real time, resource constraints