arxiv:2610.34283

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

Published on Oct 25

· Submitted by

Yihao Hu on Oct 28

Authors:

Pan Wang ,

Abstract

ChronoReplay mitigates interface lock-in in spatial vision-language models by replaying heterogeneous spatial supervision chronologically, improving cross-task transfer and enabling efficient inference-time evidence selection.

Generated by thinkingmachines/Inkling-Small

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. When post-training emphasizes one readout, improvements can therefore concentrate on that interface while transfer to others remains weak or even degrades; we call this observed specialization interface lock-in. ChronoReplay addresses it by replaying heterogeneous spatial supervision as chronologically ordered episodes, retaining source-native targets, and balancing supervision across families. Under the same observation access, a single ChronoReplay checkpoint improves OVO-S-Bench, OST-Bench, and UCS-Bench together, including a 13.41-point gain on OST-Bench, and achieves the highest three-benchmark mean among the supervision settings compared here. Same-data controls further separate the ingredients of this transfer: answer serialization redistributes gains across OST and UCS, family-balanced learning improves over ordinary joint SFT and natural turn weighting, and chronological replay exceeds a shifted-order control by about 2.41 OST points across two paired seeds. Adaptive frame selection then changes evidence access while keeping the trained model fixed, providing an additional inference-time gain at lower visual cost. Streaming spatial performance depends on what distinctions supervision preserves, how heterogeneous histories are organized during learning, and what causal evidence inference makes available at test time.

View arXiv page View PDF Project page GitHub 7 Add to collection

Community

HU22333

Paper submitter about 14 hours ago

ChronoReplay is a training framework designed to improve the spatial reasoning capabilities of streaming vision-language models (VLMs).

Existing spatial VLMs often specialize in specific answer formats, such as categorical spatial relations, metric distances, or motion trajectories. This specialization, which we call interface lock-in, limits generalization across different spatial reasoning tasks.

To address this challenge, ChronoReplay organizes heterogeneous spatial supervision into chronologically ordered episodes, preserves task-specific answer formats, and balances learning across supervision families. This enables a single model to learn complementary spatial capabilities while maintaining temporal consistency.

Key Highlights:

Unified Spatial Learning: Supports categorical, metric, and trajectory-based spatial reasoning within a single model.

Chronological Replay: Preserves temporal dependencies by organizing training examples into ordered interaction episodes.

Family-Balanced Training: Improves cross-task generalization by balancing heterogeneous supervision signals.

Strong Benchmark Performance: Achieves consistent improvements across OVO-S-Bench, OST-Bench, and UCS-Bench, including a 13.41-point improvement on OST-Bench.

Efficient Streaming Inference: Adaptive frame selection further improves performance while reducing visual processing costs.

ChronoReplay demonstrates that effective streaming spatial intelligence requires preserving complementary spatial information across tasks, learning from temporally structured experiences, and selecting relevant visual evidence at inference time.