Paper summary
Paper summary: ChronoReplay: One Answer Interface Is Not Enough for Streaming Spatial Intelligence
Training a spatial VLM on one kind of answer (e.g. left/right) doesn't teach related ones (e.g. distance), which the authors call "interface lock-in"; ChronoReplay replays mixed spatial supervision as time-ordered episodes and improves three streaming benchmarks together.
- Supervision for "which side?" doesn't force a model to keep the metric distance behind the answer, and vice versa.
- Replaying heterogeneous supervision in chronological order, balanced across answer types, improved three benchmarks at once.