AI Task vector

Creator
Creator
Seonglae Cho
Created
Created
2025 Jan 26 18:30
Editor
Edited
Edited
2025 Mar 6 12:10
Refs
Refs

Function Vector, Task feature

sparse SAE task vector fine-tuning (gradient-based cleanup)

https://www.lesswrong.com/posts/5FGXmJ3wqgGRcbyH7/extracting-sae-task-features-for-in-context-learning
Obtain a more accurate steering vector through gradient-based cleanup of the steering vector obtained from the SAE decoder since it has reconstruction error with linear combination of SAE features.

Gradient-based cleanup

Fine-tuning is applied to the target vector to efficiently reconstruct neuron activation patterns present in the residual using the SAE basis. Through gradient-based cleanup, features with small gradients were removed to create a compact SAE. This shows improved performance compared to the existing task vector and provides interpretability.
 
 
 
 

In-context learning

TVP-loss to emerge task vector of ICL into specific layer (2025)
Neuron SAE
features mimic task vector steering based on Task detector features and task feature from separator token’s residual mean as task vector with Gradient based Cleanup (2024)
Task vectors play an important role in in-context learning and appear in latter layers (2023)
icl_task_vectors
roeehendelUpdated 2025 Feb 13 5:38

Finding visual task vector using
Policy Gradient Learning

function vector

TaskVec

 
 

 

Recommendations