Baseline Preference
The default inclination a model naturally exhibits toward a value dimension in the absence of steering or special prompting. It arises from the distribution of pretraining corpora, the tendencies of instruction-tuning data, and annotator consensus in preference alignment, and it self-reinforces across multi-turn conversations.
Timeline (last 90 days)
In multi-turn conversations, when the assistant and the user steer in opposite directions, it is often the side that matches the assistant's baseline bias that ultimately prevails.
Baseline bias refers to the default tilt a model naturally exhibits toward a certain value dimension when there is no steering at all, originating from the distribution of pretraining corpora, the selection tendencies of instruction fine-tuning data, and the group consensus of human annotators during preference alignment.
In multi-turn conversations, baseline bias continuously reinforces itself through the model's ongoing reinterpretation of context, with wording biased toward one side being repeatedly cited and amplified, forming inertia.
For product teams, it is necessary to recognize that when user and assistant values conflict, the model's default tendency may be the actual dominant factor.
All Facts (4)
In multi-turn conversations, when the assistant and the user steer in opposite directions, it is often the side that matches the assistant's baseline bias that ultimately prevails.
50%UnverifiedBaseline bias refers to the default tilt a model naturally exhibits toward a certain value dimension when there is no steering at all, originating from the distribution of pretraining corpora, the selection tendencies of instruction fine-tuning data, and the group consensus of human annotators during preference alignment.
50%UnverifiedIn multi-turn conversations, baseline bias continuously reinforces itself through the model's ongoing reinterpretation of context, with wording biased toward one side being repeatedly cited and amplified, forming inertia.
50%UnverifiedFor product teams, it is necessary to recognize that when user and assistant values conflict, the model's default tendency may be the actual dominant factor.
50%