Seonglae Cho- Event Segmentation
- Calculate how "surprised" the LLM is at the next token (Bayesian Surprise)
- Set points of high surprise as event boundaries → divide context into natural event units
- Boundary Refinement
- Ensure segmented events are internally similar and distinct from each other
- Adjust boundaries using graph-based metrics (modularity/conductance)
- Memory Retrieval
- (1) Similarity-based k-NN event retrieval
- (2) Temporal contiguity-based retrieval
→ Mimics human memory recall patterns (sequential replay of adjacent items)
- LLM Input Construction
- Local context + initial tokens + retrieved events (similarity/contiguity)