Thread #3 — LEO-VL: Efficient Scene Representation for Scalable 3D Vision-Language Learning

Date: 2026-03-24
Paper: arxiv.org/abs/2506.09935
Template: Performance Breakthrough (adapted)

---

1/ 3D VLMs have been bottlenecked by token-heavy scene reps that kill scalability.

BIGAI drops LEO-VL: Condensed Feature Grid (CFG) slashes tokens while hitting SOTA on SQA3D/Beacon3D.

700k 3D-VL data across indoor domains. Here's the breakthrough: 🧵

2/ The problem: Voxel/point grids explode tokens. Transformers choke on quadratic cost, limiting training scale and spatial reasoning.

3/ Previous: Dense grids or sparse sampling. Tradeoffs kill either efficiency or perception.

4/ CFG approach: Condensed grid fuses multi-res features via learnable aggregation. Key: Reduces tokens 10x+ while preserving fine details.

[Diagram from paper if possible, but text: low-res backbone + high-res refinement]

5/ Results: • SQA3D: +5-10% over prior 3D VLMs • Beacon3D: New SOTA • Scan2Cap: Better captions

Trained on 700k diverse indoor data (4 domains, 5 tasks)

6/ Why builders care: • Scalable to massive data without token hell • SceneDPO post-training boosts robustness • Unlocks real-world 3D VQA/dialogue

7/ Limits: • Indoor focus (outdoor next?) • Still needs quality 3D scans

8/ Takeaway: Efficient reps > raw scale. CFG patterns stealable for your 3D pipelines.

9/ Paper: https://arxiv.org/abs/2506.09935 Project: https://leo-vl.github.io

Follow @soren_cto for ArXiv → builder insights 🧵1/3