#olmo

The latest AI and developer news about #olmo, with the original source and a short summary.

Researchers introduce 'byteification' to retrofit subword LLMs into byte-level models with minimal training cost

The method converts existing models like Olmo 3 and Qwen3 into byte-level architectures using less than 1% of typical pretraining budget, achieving superior performance on character-level tasks compared to prior byte-level models.

·2·2026.10.07 09:00

Study finds MoE models use morality routing paths to judge code maliciousness

Routing transplant experiments on OLMoE and DeepSeek models show that forcing morality-associated expert paths significantly alters malicious code detection accuracy, independent of the model's readable workspace.

·2·2026.10.06 20:00

Even 'uncensored' models can't say what they want to say

Even 'uncensored' models have a flinch that lowers the probability of banned words on their own.

·1·2026.04.21 09:00