Objects vanish mid-sequence in video models, not because they moved, but because the tokenizer dropped them. 3D-DLP (ICML 2026, from Lambda) learns the world as

AI News • 2026-08-03 12:35 UTC • By Billy Odell Tucker-Robinson

Objects vanish mid-sequence in video models, not because they moved, but because the tokenizer dropped them. 3D-DLP (ICML 2026, from Lambda) learns the world as objects, not tokens. Paper: [📷](https://pbs.twimg.com/media/HOzETDgawAAXm_k.jpg) [↧](https://t.co/xv2rRgpc7j)

Share this article

𝕏 X / Twitter Facebook LinkedIn WhatsApp

📚 More from Billy’s World

📰 Banking With Billy News All breaking financial news & market analysis 📚 Banking With Billy Books Deep-dive books on every financial story 💬 Join the Community discord.gg/VHxwmR5j4Y 🎥 Banking With Billy on YouTube Live streams & market commentary

Banking With Billy News Network — bankingwithbilly.com
Discord: discord.gg/VHxwmR5j4Y • YouTube: @BankingWithBilly