acceptodds
Desk rejected at ICLR 2027

Renewal–Persistence Replay for Returning Concepts in Timestamped Multimodal Streams

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

An image–text pair may become useful again after its concept disappears and returns, but a finite replay buffer can evict the pair before its renewed value is visible. Renewal–Persistence Replay (RPR) addresses both decisions: a frozen harmful-staleness ranker updates priority from teacher disagreement, embedding migration, and stable-anchor support, while cluster sampling mass and eviction floors keep low-frequency history reachable. Recurring, low-displacement concepts account for 28.5–35.8% of future queries in our three timestamped streams, making this return regime substantial rather than exceptional. With the model, contrastive objective, buffer, update budget, and checkpoint rule held fixed, five-seed SciCapStream comparisons reduce forgetting by 0.70.5 points versus MIR-style replay and 0.90.5 versus GSS; RPR also lowers forgetting from 14.10.6 to 9.40.4 relative to strict timestamp ordering without reducing future retrieval. The ordering repeats on PMC-Quarterly and RedCaps-Monthly and transfers to ViT-L/14. Its mechanism is localized: the SciCap RPR–Recency effect expands to -6.60.9 points for high recurrence but contracts to -0.20.5 for weak recurrence; corrupting visible time erases RPR's margin over a shared timestamp-agnostic MIR trajectory; and oscillating concepts in ConceptDrift-5K gain +14.71.8 R@1. These results identify returning-concept streams as a consequential replay regime and show that coupling current utility to continued reachability improves replay there.

Primary area:
foundation or frontier models, including LLMs
Keywords:
multimodal and embodied learning, Renewal--Persistence Replay for Returning Concepts in Timestamped Multimodal Streams

Related papers

Discussion (0)

Sign in to comment.