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

ChronoReplay. One Answer Interface Is Not Enough for Streaming Spatial Intelligence. By Yihao Hu and ten others, from Westlake University, Zhejiang University, and five other institutions. Published on the twenty-fifth of October.

A spatial vision-language model can be trained to say which side an object lies on without ever having to preserve the metric displacement that decides that answer. Train it instead to estimate distances, and the categorical relations may be left largely unconstrained. Each way of reading out an answer pins down some distinctions and leaves others free. So when post-training leans on one readout, the improvement can pool in that one interface, while transfer to the others stays weak or even gets worse. The authors call this observed specialization interface lock-in.

ChronoReplay is their response. Rather than favouring one kind of supervision, it replays heterogeneous spatial supervision as episodes ordered in time. Each source keeps its own native targets instead of being converted into a single answer format, and supervision is balanced across the families it comes from. The aim is that no single readout dominates what the model learns. Categorical questions, metric questions and the rest each keep their own form of answer, arrive in the order a stream would present them, and carry comparable weight.

Three ingredients, then: chronological replay, source-native targets, and family balance. Same-data controls are used to pull them apart. Answer serialization redistributes the gains between OST and UCS. Family-balanced learning improves on ordinary joint SFT, and on natural turn weighting. Chronological replay beats a shifted-order control by about two point four one OST points, across two paired seeds. A separate step works at inference rather than in training. Adaptive frame selection changes which evidence the model gets to see, while the trained model stays fixed, and it adds a further gain at lower visual cost.

Under the same observation access, one ChronoReplay checkpoint improves three benchmarks together: OVO-S-Bench, OST-Bench and UCS-Bench. The gain on OST-Bench is thirteen point four one points. Among the supervision settings compared, it reaches the highest mean across the three benchmarks. The point is that a single model improves on all three at once, rather than a separate model specializing in each. That is the opposite of interface lock-in, where gains on one benchmark would come at the expense of the others.

The authors draw the lesson in three parts. Streaming spatial performance depends on which distinctions supervision preserves. It depends on how heterogeneous histories are organized during learning. And it depends on what causal evidence inference makes available at test time. The title's claim follows from the first part: train through one answer interface, and what the model learns can stay locked inside it.