The Nonlinear Library LW - Transformer Circuit Faithfulness Metrics Are Not Robust by Joseph Miller
Jul 12, 2024
11:08
Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Transformer Circuit Faithfulness Metrics Are Not Robust, published by Joseph Miller on July 12, 2024 on LessWrong.
When you think you've found a circuit in a language model, how do you know if it does what you think it does? Typically, you ablate / resample the activations of the model in order to isolate the circuit. Then you measure if the model can still perform the task you're investigating.
We identify six ways in which ablation experiments often vary.[1][2]
How do these variations change the results of experiments that measure circuit faithfulness?
TL;DR
We study three different circuits from the literature and find that measurements of their faithfulness are highly dependent on details of the experimental methodology. The IOI and Docstring circuits in particular are much less faithful than reported when tested with a more precise methodology.
The correct circuit for a set of prompts is undefined. The type of ablation you use to isolate the circuit determines the task that you are asking the circuit to perform - and therefore also the optimal circuit.
This is especially important because previous work in automatic circuit discovery has tested algorithms by their ability to recover these "ground-truth" circuits from the literature - without considering these potential pitfalls and nuances.
Case Studies
We look at three circuits from the mech interp literature to demonstrate that faithfulness metrics are highly sensitive to the details of experimental setup.
Indirect Object Identification Circuit
The IOI circuit is the most well known circuit in a language model. It computes completions to prompts of the form:
"When Mary and John went to the store, John gave a bottle of milk to ____"
The circuit is specified as a graph of important attention heads (nodes) and the interactions between them (edges) as applied to a specific sequence of tokens. The authors report that the circuit explains 87% of the logit difference between the two name tokens. They find this number by passing some inputs to the model and ablating all activations outside of the circuit. Then they measure how much of the logit difference between the correct and incorrect name logits remains.
However, an important detail is that they arrived at this number by ablating the nodes (heads) outside of the circuit, not by ablating the edges (interactions between heads) outside of the circuit. So they don't ablate, for example, the edges from the previous token heads to the name mover heads, even though these are not part of the circuit (effectively including more edges in the circuit).
We calculate the logit difference recovered (defined below) when we ablate the edges outside of the circuit instead.
They ablate the heads by replacing their activations with the mean value calculated over the "ABC distribution", in which the names in the prompts are replaced by random names.[3] In our experiments, we also try resampling the activations from different prompts (taking individual prompt activations instead of averaging).
The first thing that jumps out from the box plots above is the very large range of results from different prompts. The charts here are cut off and some points are over 10,000%. This means that although the average logit difference recovered is reasonable, few prompts actually have a logit difference recovered close to 100%.
And we see that ablating the edges instead of the nodes gives a much higher average logit difference recovered - close to 150% (which means that the isolated circuit has a greater logit difference between the correct and incorrect names than the un-ablated model). So the edge-based circuit they specified it is much less faithful than the node-based circuit they tested.
The authors calculate the 87% result as the ratio of the expected difference (over a set of prompts) in the ...
When you think you've found a circuit in a language model, how do you know if it does what you think it does? Typically, you ablate / resample the activations of the model in order to isolate the circuit. Then you measure if the model can still perform the task you're investigating.
We identify six ways in which ablation experiments often vary.[1][2]
How do these variations change the results of experiments that measure circuit faithfulness?
TL;DR
We study three different circuits from the literature and find that measurements of their faithfulness are highly dependent on details of the experimental methodology. The IOI and Docstring circuits in particular are much less faithful than reported when tested with a more precise methodology.
The correct circuit for a set of prompts is undefined. The type of ablation you use to isolate the circuit determines the task that you are asking the circuit to perform - and therefore also the optimal circuit.
This is especially important because previous work in automatic circuit discovery has tested algorithms by their ability to recover these "ground-truth" circuits from the literature - without considering these potential pitfalls and nuances.
Case Studies
We look at three circuits from the mech interp literature to demonstrate that faithfulness metrics are highly sensitive to the details of experimental setup.
Indirect Object Identification Circuit
The IOI circuit is the most well known circuit in a language model. It computes completions to prompts of the form:
"When Mary and John went to the store, John gave a bottle of milk to ____"
The circuit is specified as a graph of important attention heads (nodes) and the interactions between them (edges) as applied to a specific sequence of tokens. The authors report that the circuit explains 87% of the logit difference between the two name tokens. They find this number by passing some inputs to the model and ablating all activations outside of the circuit. Then they measure how much of the logit difference between the correct and incorrect name logits remains.
However, an important detail is that they arrived at this number by ablating the nodes (heads) outside of the circuit, not by ablating the edges (interactions between heads) outside of the circuit. So they don't ablate, for example, the edges from the previous token heads to the name mover heads, even though these are not part of the circuit (effectively including more edges in the circuit).
We calculate the logit difference recovered (defined below) when we ablate the edges outside of the circuit instead.
They ablate the heads by replacing their activations with the mean value calculated over the "ABC distribution", in which the names in the prompts are replaced by random names.[3] In our experiments, we also try resampling the activations from different prompts (taking individual prompt activations instead of averaging).
The first thing that jumps out from the box plots above is the very large range of results from different prompts. The charts here are cut off and some points are over 10,000%. This means that although the average logit difference recovered is reasonable, few prompts actually have a logit difference recovered close to 100%.
And we see that ablating the edges instead of the nodes gives a much higher average logit difference recovered - close to 150% (which means that the isolated circuit has a greater logit difference between the correct and incorrect names than the un-ablated model). So the edge-based circuit they specified it is much less faithful than the node-based circuit they tested.
The authors calculate the 87% result as the ratio of the expected difference (over a set of prompts) in the ...