
The directions obtained through sparse coding showed higher interpretability compared to random directions, PCA, and ICA when used as neuron basis
2021 MIT
Natural language descriptions of deep visual features
2023 OpenAI
All of activations are quantized to be between 0-9 inclusive.
2024
intervention scoring
delphi
EleutherAI • Updated 2025 Jun 29 17:16
sparsify
EleutherAI • Updated 2025 Jul 1 6:5
Analyze the impact of a feature on model outputs by comparing the baseline output with the output after intervention
Gradual improvement with hypothesis Best-of-k sampling and small model by knowledge distillation
In order to get feature explanations Claude 2 is provided with a total of 49 examples: ten examples from the top activations interval; two from the other 12 intervals; five completely random examples; and ten examples where the top activating tokens appear in different contexts. Finally, we ask the model to be succinct in its answer and not provide specific examples of tokens it activates for.
Using the explanation generated, in a new interaction Claude is asked to predict activations for sixty examples: six from the top activations; two from the other 12 intervals; ten completely random; and twenty top activating tokens out of context. For the sake of computational efficiency, Claude scores all sixty examples in a single shot, repeating each token followed by its predicted activation. In an ideal setting, each example would be given independently as its own prompt.