Inference-Time Intervention
A controllable generation technique that directly modifies activation values at specific layers during model inference. Target directions are extracted from contrastive sample pairs and applied during forward propagation, enabling guidance of model behavior without updating any model parameters.
Timeline (last 90 days)
Prompting provides instructions and examples at inference time, has the lowest cost and fastest iteration, and is suitable for solving problems such as expression style, format constraints, and task boundary specification, but it does not change model weights, so it cannot introduce private knowledge the model does not know, nor can it guarantee real-time capability.
Certain inference-time intervention methods are sensitive to answer format and language coverage breadth, being effective on specific formats but potentially failing when switched to other formats.
All Facts (2)
Prompting provides instructions and examples at inference time, has the lowest cost and fastest iteration, and is suitable for solving problems such as expression style, format constraints, and task boundary specification, but it does not change model weights, so it cannot introduce private knowledge the model does not know, nor can it guarantee real-time capability.
50%UnverifiedCertain inference-time intervention methods are sensitive to answer format and language coverage breadth, being effective on specific formats but potentially failing when switched to other formats.
50%