ChronoReplay Fixes Interface Lock-in in Spatial VLMs, +13.41 on OST-Bench
LINs-lab · hf · 2026-10-28
Westlake University with ZJU, USTC and others released ChronoReplay, addressing a newly identified failure mode in spatial vision-language models: interface lock-in.
- Problem: post-training that emphasizes one readout interface (e.g. categorical side queries) concentrates gains there while transfer to others (e.g. distance estimation) stays weak or degrades.
- Method: replay heterogeneous spatial supervision chronologically as ordered episodes, retaining source-native targets and balancing supervision across families.
- Results: a single checkpoint improves OVO-S-Bench, OST-Bench and UCS-Bench jointly, with a 13.41-point gain on OST-Bench and the best three-benchmark mean among compared settings; ablations separate the contributions of answer serialization, family balancing and chronological ordering.
More from Research
- Norn Group raises $9.1M of $15M to build the datasets AI for aging will need — juanbenet · 2026-10-28
- Researcher admits SETH/CSETH mixup, revises odds in P vs NP debate to >80% on CSETH — geoffreyirving · 2026-10-28
- Softmax's standard implementations hide two numerical bugs, Google paper shows — srush_nlp · 2026-10-28
- Percy Liang: experts rate 1/3 of AI-proposed research questions genuinely interesting — StanfordAILab · 2026-10-28
- Protenix team spins out Molence Bio, releases open biomolecular model Mol-1 — LeoTZ03 · 2026-10-28
- Just Action Model: decoupling what separates VLA from world-action models — Yonghao Yu · 2026-10-28