BAAI releases LLaCO, an open-weight 8B continuous diffusion language model
Article timeline
The articles in this story. Times are ET.
BAAIOfficialLLaCO adds spherical noising and log-rank retention to diffusion language modelsAILLaCO is a diffusion language model method that uses spherical noising, a geodesic interpolation toward a [MASK]-biased endpoint that avoids norm decay while keeping pretrained semantic biases. It also introduces Log-Rank Retention (Q_log), which measures rank decay on a logarithmic scale to focus on top candidates, and trajectory reallocation, which concentrates compute where token identities shift.
- BAAIOfficialBAAI open-sources LLaCO, a continuous diffusion language model for reasoning
AIBAAI, with BUPT and PKU, releases the code and weights for LLaCO, a continuous diffusion model that it says scales to serious reasoning. The post says continuous diffusion gives bidirectional contextual planning and smooth semantic trajectories without autoregressive bottlenecks. Paper, code, weights and project page are published on arXiv, GitHub, Hugging Face and a project site.
- BAAIOfficialLLaCO-8B reports top scores among diffusion-based LLMs on nine benchmarks
AIBAAI says its LLaCO-8B, a diffusion-based LLM, scores 62.12 on MATH, 47.02 points above DSL-LLaDA-8B's 15.10. The post reports 48.17 on HumanEval, 80.74 on GSM8K, 69.89 on MMLU and 31.25 on GPQA, and says these results beat discrete models including Dream-7B (51.20 on MATH) and LLaDA-8B (43.82).
- BAAIOfficialLLaCO, an 8B continuous diffusion language model, beats LLaDA-8B and Dream-7B on math and code
AIBAAI, BUPT, and PKU introduce LLaCO, an 8B continuous diffusion language model. The post says it outperforms previous continuous baselines and discrete diffusion models LLaDA-8B and Dream-7B on math and code tasks.
