A model can answer “which side?” without preserving the distance that supports that judgment. This mismatch helps explain why improving one spatial answer format may not improve another. The paper’s central insight is to treat the distinctions preserved by supervision—and the order in which evidence is presented—as part of the learning problem, not just the final answer format.
Key Findings
- A single ChronoReplay checkpoint improves on OVO-S-Bench, OST-Bench, and UCS-Bench under matched observation access. Reported scores rise from 48.98 to 53.89, 50.52 to 63.93, and 45.33 to 52.77, respectively.
- The largest reported matched-access gain is 13.41 points on OST-Bench. Same-data controls suggest that family-balanced learning and chronological replay contribute beyond ordinary joint fine-tuning; chronological replay exceeds a shifted-order control by about 2.41 OST points across two paired seeds.
- Adaptive frame selection is a separate inference-time change, not a training comparison: with the trained checkpoint fixed, it raises OVO-S from 53.89 to 57.98 while using 80.50 frames on average versus 128 for uniform sampling.
Method and Scope
The training pool combines categorical preservation, numeric geometry or camera-motion targets, and causal trajectory episodes while retaining their original answer formats. Chronological replay ties trajectory turns to their observation boundaries, and family-balanced losses prevent the largest supervision source from dominating shared learning. The visual encoder remains frozen while the language model and visual merger are trained.
The evidence is limited to three benchmarks and reported experimental results; the available sources do not establish independent replication or performance in operational settings. The findings are most useful for researchers studying spatial VLM transfer and causal video understanding, rather than as evidence of deployment-ready spatial reliability.