I think a few things will happen soon:
🚀 Scale beyond 8B
🎯 Multi-modal capabilities
⚡️Faster inference
🔄 Reinforcement learning integration

Exciting to see alternatives to autoregressive models succeeding at scale!

Paper: ml-gsai.github.io/LLaDA-demo/

(8/8)

Results vs LLaMA3 8B:

- Matches/exceeds on most tasks
- Better at math & Chinese tasks
- Strong in-context learning
- Improved dialogue capabilities

(7/8) 🧵

A major result: LLaDA breaks the "reversal curse" that plagues autoregressive models. 🔄

On tasks requiring bidirectional reasoning, it outperforms GPT-4 and maintains consistent performance in both forward/reverse directions.

(6/8) 🧵

For generation, they introduce clever remasking strategies:

- Low-confidence remasking: Remask tokens the model is least sure about

- Semi-autoregressive: Generate in blocks left-to-right while maintaining bidirectional context

(5/8) 🧵

Training uses random masking ratio t ∈ [0,1] for each sequence.

The model learns to predict original tokens given partially masked sequences. No causal masking used.

Also enables instruction-conditioned generation with the same technique. No modifications.

(4/8) 🧵

💡Core insight: Generative modeling principles, not autoregression, give LLMs their power.

LLaDA's forward process gradually masks tokens while reverse process predicts them simultaneously. This enables bidirectional modeling.

(3/8) 🧵

Key highlights:
- Successful scaling of masked diffusion to LLM scale (8B params)
- Masking with variable ratios for forward/reverse process
- Smart remasking strategies for generation, incl. semi-autoregressive
- SOTA on reversal tasks, matching Llama 3 on others

(2/8) 🧵

Project page: bujiazi.github.io/light-a-vide...
Code: github.com/bcmi/Light-A...

Could be a game-changer for quick video mood/lighting adjustments without complicated VFX pipelines! 🎬

The results are pretty good ✨
They can transform regular videos into moody noir scenes, add sunlight streaming through windows, or create cyberpunk neon vibes -- works on everything from portrait videos to car commercials! 🚗

Technical highlights 🔍:
- Consistent Light Attention (CLA) module for stable lighting across frames
- Progressive Light Fusion for smooth temporal transitions
- Works with ANY video diffusion model (AnimateDiff, CogVideoX)
- Zero-shot - no fine-tuning needed!

Project page: liuisabella.com/RigAnything/
Code: not available yet

Really excited to try this out once the code is available!

Authors claim that the model generalizes well across diverse shapes - from humanoids to marine creatures! And works with real-world images & arbitrary poses. 🤩

Technical highlights:
- BFS-ordered skeleton sequence representation
- Autoregressive joint prediction with diffusion sampling
- Hybrid attention masking: full self-attention for shape tokens, causal attention for skeleton
- e2e trainable pipeline without clustering/MST ops

@sebastianraschka.com this is such an interesting discussion! I haven't tried this myself, but I think this can be analyzed theoretically by looking at the rank of the attention matrix in both cases.

I have posted my thoughts on the discussion here: github.com/rasbt/LLMs-f...

Interesting how they handle the domain gap between 2D latent space and 3D representations through their three-stage pipeline. The correspondence-aware encoding significantly reduces high-frequency noise while preserving geometry.

Project: latent-radiance-field.github.io/LRF/

Technical approach:
- Correspondence-aware autoencoding to enhance 3D consistency in VAE latent space
- Builds 3D representations from 3D-aware 2D features
- VAE-Radiance Field alignment to bridge domain gap between latent and image space

#nerf #ai #research

Project: research.nvidia.com/labs/dir/edg...
Training and inference code available here: github.com/NVlabs/EdgeR...

The architecture uses a lightweight encoder and auto-regressive decoder to compress variable-length meshes into fixed-length codes, enabling point cloud and single-image conditioning.

Their ArAE model controls face count for varying detail while preserving mesh topology.