Steering the Language Axis: From Linear Decodability to Causal Control
The episode discusses a paper titled "Steering the Language Axis: From Linear Decodability to Causal Control." Hosts detail how researchers found a steerable 'language axis' in AI models, allowing them to causally control output language. They conclude this tool offers a way to fix language biases without retraining, though it is highly specific to model architecture and language pairs.