[KongchangAI]
Concepthidden state steering / 表示操控

Hidden State Manipulation

A causal intervention method in mechanistic interpretability research. It adjusts states along specific directions in the hidden representation space to amplify or suppress the encoding of a given factor, and observes changes in model output to provide causal evidence.

Timeline (last 90 days)

Oct 21

The researchers conducted hidden state steering experiments on six models of 7-9B parameters, and shifting hidden states along the direction corresponding to each factor changed the models' judgments as expected.

Unverified50%
Oct 21

The hidden state steering experiments provide causal evidence that psycholinguistic factors such as commonness and distinctiveness are encoded inside the models, rather than being merely correlational observations.

Unverified50%
Oct 21

Hidden state steering is a causal intervention method in mechanistic interpretability research that locates concept directions in hidden representation space and adds or subtracts vectors along that direction to amplify or suppress the strength of a factor's encoding.

Unverified50%

All Facts (3)

Source Articles