Comment by jsrozner
4 days ago
A big problem with some of these supervised* interpretability approaches is that they can find spurious structure. (There are lots of ways to make the model do what you want; which is roughly what Hewitt and Liang 2019 showed). This paper draws a contrast to a previous method, DAS (distributed alignment search) on page 20. These and related methods rest on theories of causal abstraction, which are great in theory, but harder in practice. DAS, for example, has faced numerous recent criticisms (Makelov 2024, Meloux 2025, Sutter 2025, Grant 2026, Kumon 2026). My favorite is the quite approachable Meloux et al.; Sutter 2025 is also really good, but relies on a sort of real number argument that allows a lossless encoding of every input.
My forthcoming paper at EMNLP offers an alternative that instead grounds the notion of representation in a very simple notion of the effect it has on model learning/behavior when you adversarially perturb it. For example, if I tell a model that in the context "I saw a duck quacking" it should replace 'duck' with 'glam', how much does it desire to replace 'duck' with 'glam' in "I need to duck out of the meeting" vs. "At the park a duck protected her ducklings." This method turns out to work quite well, and as we use only a single example, avoids the need for supervision.
The linked paper argues that their method, DISCOVER, is not supervised in the same way as DAS, since it does not directly optimize for causal effect. I have only skimmed this, but I am not so sure it might not suffer from a similar issue. They're still supervising to align representations with their underlying hypothesis, even if they don't directly supervise for causal outcomes.
Refs
- Hewitt and Liang 2019. Designing and interpreting probes with control tasks
- Kumon and Yanaka, 2026. Fine-grained analysis of shared syntactic mechanisms
- Meloux et al., 2025. Everything everywhere all at once
- Rozner and Shain 2026. Perturbation: A simple and efficient adversarial tracer for representation learning in LMs. https://arxiv.org/abs/2603.23821
- Sutter et al. 2025. The nonlinear representation dilemma
I'm new to the mechanistic interpretability field, but from what I have read insofar, a lot of papers have relied on ablation/causal interventions to prove the faithfulness of their models.
Do you have a simple explanation of why this level of proof is not sufficient?
So, basically you're saying: I have only skimmed the article, similar approaches had issues in the past, and my own method from my forthcoming paper is better.
I mean, no disrespect, but that's the core of your argument, yes?
It certainly comes across as disrespectful.
To my reading the nature of supervision creating the structure seems to be the core of the argument.
I have issue with the comment because he says he didn't read it, then unfavorably compares it to a previous method, and finally uses that negative review to plug his own article instead. His criticism might be valid, I'm not in a position to judge, but the self-promotion leaves a sour tastes in my mouth and makes me question how much of the criticism is just drummed up to make his own contribution appear more relevant.
2 replies →
His arguments apply generally to a large family of interpretability methods. Skimming suffices to figure out that the linked article is in that family.