Paper summary: ChronoReplay: One Answer Interface Is Not Enough for Streaming Spatial Intelligence

Problem

The authors observe interface lock-in. Training a spatial VLM to answer which side an object is on doesn’t require keeping the metric displacement behind that answer, and distance supervision may leave categorical relations unconstrained. Gains concentrate on the trained answer format and transfer poorly to others.

Method

ChronoReplay replays heterogeneous spatial supervision as chronologically ordered episodes, keeps each source’s native targets, and balances supervision across answer families. At inference, adaptive frame selection changes which frames the model sees, without retraining.

Key results

From the abstract:

• one checkpoint improves OVO-S-Bench, OST-Bench and UCS-Bench together, including +13.41 points on OST-Bench
• same-data controls separate the ingredients: family-balanced learning beats ordinary joint SFT, and chronological replay beats a shifted-order control by about 2.41 OST points across two paired seeds
• adaptive frame selection adds further gains at lower visual cost

Limitations

There’s no dedicated limitations section in the extracted text. Some ablations use only two paired seeds.