SRIQ is an open AI research group working on language models that reason more efficiently.

Reasoning models often spend thousands of tokens reaching answers that need far less. We train and study models that use shorter reasoning traces while keeping answer quality high.

Research

• Sparse and compressed reasoning
• Efficient post-training
• Reasoning quality per token
• Inference and token reduction
• Evaluation and benchmarking

We publish our models, datasets, and experiments openly for others to test and build on.