Sahaj.blogs

Day page

4 Mar 2026

Paper summary

Paper summary: Utonia: Toward One Encoder for All Point Clouds

Utonia trains one self-supervised point-transformer encoder across very different point-cloud domains (remote sensing, LiDAR, indoor RGB-D, CAD, video-lifted points), and its features also help robot policies and spatial reasoning in VLMs.

  • Despite different sensors, densities and priors, one encoder learns a shared representation across point-cloud domains.
  • Training domains jointly produced behaviours that didn't appear when they were trained separately.

Read the post →

Study

LLMs · QLoRA

QLoRA keeps the frozen base model in 4-bit NormalFloat (NF4) and trains LoRA adapters on top; building blockwise 4-bit quantisation by hand and comparing NF4 with plain 4-bit on normal-shaped weights.

  • QLoRA stores the frozen base weights in 4 bits and trains only small LoRA adapters in higher precision.
  • NF4 places its 16 levels at quantiles of a normal distribution, which suits neural-network weights.

Read the post →