ChronoReplay reduces interface lock-in in spatial VLMs
Original title: ChronoReplay: One Answer Interface Is Not Enough for Streaming Spatial Intelligence
One checkpoint improves OVO-S-Bench, OST-Bench and UCS-Bench together, including a 13.41-point gain on OST-Bench.
The paper describes interface lock-in, where post-training a spatial vision-language model on one answer interface improves that readout while transfer to others stays weak or degrades. ChronoReplay replays heterogeneous spatial supervision as chronologically ordered episodes, keeps source-native targets, and balances supervision across families. Under the same observation access, a single checkpoint improves OVO-S-Bench, OST-Bench and UCS-Bench and records the highest three-benchmark mean among the compared settings. Same-data controls attribute transfer to answer serialization, family-balanced learning and chronological order, and adaptive frame selection adds an inference-time gain at lower visual cost.
- Chronological replay beats a shifted-order control by about 2.41 OST points.
- That comparison uses two paired seeds.
- Family-balanced learning improves over ordinary joint SFT and natural turn weighting.
- Adaptive frame selection keeps the trained model fixed while changing evidence access.
From the original article
1 min read at huggingface.co
Training a spatial VLM to answer which side an object lies on does not require it to preserve the metric displacement that determines that answer, while supervision for a distance estimate may leave categorical relations largely unconstrained.
By Yihao Hu, Yitong Lin, Pan Wang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.