Scaling cross-situational word learning: 2,441 hours of child-perspective video from Stanford team
RichmanRonald · x · 2026-10-29
Michael C. Frank's group released an arXiv preprint testing cross-situational word learning at scale using the BabyView corpus — 2,441 hours of longitudinal egocentric video from 48 children, yielding 1.69M training pairs. Contrastive vision-language models trained on it did learn object–label mappings, but performance was far higher with internet-pretrained visual representations than with representations learned solely from egocentric video. Using Gemini 2.5 Flash to score pair alignment, only 10% of pairs scored ≥50 and 7% had a spoken referent. Conclusion: cross-situational learning scales, but its ceiling depends on visual representation quality from naturalistic child-view data.
More from Models
- ARC nearing ceiling effect: Greg Kamradt proposes difficulty-ranked Elo list for public evals — GregKamradt · 2026-10-29
- Leaker: Google shipped Carbon ahead of OpenAI and Anthropic's next-gen models, still ~2 months behind — scaling01 · 2026-10-29
- ARC team: v1 hits ceiling effect as v2 extends headroom for frontier models — MindsAI_Jack · 2026-10-29
- Gemini 4 Argon Posts Lowest Hallucination Rate at 4% on AA-Omniscience — haider1 · 2026-10-29
- Suhail launches Particle: particle-1 matches performance with ~45% fewer tokens — chrisalbon · 2026-10-29
- Exploring a whole repo with a 1M-context open model cost just $0.63 — HowDevelop · 2026-10-29