Gen-Verse/LatentMAS

Embeds agent reasoning in the model's latent space rather than text, where agents exchange compressed representations through a shared working memory buffer. Achieves 50–80% token reduction and 3–7× wall-clock speedups while maintaining performance, with training-free alignment and compatibility across HuggingFace models and optional vLLM backends. Supports both sequential and hierarchical multi-agent architectures across math, science, commonsense reasoning, and code generation tasks.

800 stars. Actively maintained with 2 commits in the last 30 days.