✦
Summary
ALoDLM-8B is an adaptively looped diffusion language model for parallel text and code generation. Shared transformer layers repeatedly refine unresolved latent states, with confidence-based token commitment and learned halting allocating computation where needed.
Key Features
Eight-billion-parameter checkpoint.
Diffusion text generation.
Code generation.
Shared recurrent transformer layers.
Persistent latent refinement.
Confidence-guided token commitment.
Learned loop halting.
Noncommercial model weights.
About ALoDLM-8B
Researchers can examine adaptive inference rather than using a fixed number of identical denoising passes at every token position. Core capabilities include Eight-billion-parameter checkpoint, Diffusion text generation, Code generation, and Shared recurrent transformer layers.
Public code or model weights are linked from the project, with setup instructions for running or adapting the system. Additional characteristics include Persistent latent refinement, Confidence-guided token commitment, Learned loop halting, and Noncommercial model weights.