Every Ablation Is a Dose: Counterweights and the Semblance of Self-Repair
Ablation studies in language models are confounded by self-repair, where downstream components appear to compensate for a removed part. This paper argues that such compensation is not dynamically recruited but is instead a standing, pre-existing opposition released by the intervention. The authors introduce a continuous dose axis for counterfactual interventions and show that downstream contributions follow an affine law predicted in part by static weights. If the finding generalizes, it would reshape how interpretability researchers infer component importance from ablations.
The paper makes a bold and conceptually elegant case that what looks like self-repair is merely the unmasking of pre-existing counterweights. Its strongest evidence is a nested counterfactual protocol tested across four model families, where 68 of 81 audited downstream directions obey an affine dose-response law with respectable out-of-sample $R^2$. However, the scope is deliberately narrow—a single binary classification token—and the authors acknowledge that generative tasks remain an open question. Within that scope the argument is coherent, though the leap from "affine response" to "single explanation" still smooths over some heterogeneity (13 non-coupling directions, non-additive joint-core effects, and saturation beyond $|\lambda| \le 1$).
The dose-axis reframing is methodologically powerful. By parameterizing any intervention as a signed scalar $\lambda$ on a class contrast, the authors turn incommensurable ablation methods into comparable points. The affine law $E_r^c(\lambda) = \mathrm{own}_r^c + \gamma_r \lambda$ fits three on-manifold points and predicts held-out half-rung doses with out-of-sample $R^2$ of $0.65$ to $0.91$ across families. The finding that weight-space alignment between core write and downstream read correlates with $|\gamma_r|$ (Spearman $\rho \in [0.55,0.82]$) further grounds the phenomenon in the static checkpoint rather than post-hoc dynamics.
The claim that self-repair has one mechanism is stronger than the evidence shown. The law holds for 68/81 directions, and the remaining 13 directions do not couple reliably; moreover, responses saturate outside $|\lambda| \le 1$, so linearity is local, not global. In models with two certified cores, joint dosing produces "more than the sum" of individual effects, which the paper attributes to residual core signal but which nevertheless complicates the simple per-direction affine picture. The wiring statistic predicts only unsigned magnitude, not the sign or exact coefficient, leaving substantial variance unexplained. Finally, the entire study is confined to a single true/false verdict task at one answer position, making it unclear whether the "counterweight" story applies to self-repair in generative or multi-token settings. As the authors note, "How it may translate to other tasks, and especially generative ones, remains an open question."
The evidence supports the claims within the tested scope and the comparisons to prior work are generally fair. The paper correctly situates itself against Wang et al.'s backup name movers, McGrath et al.'s Hydra effect, and Rushing and Nanda's finding that "self-repair is noisy and unlikely to have a single explanation." The authors offer a unifying reframing rather than a direct refutation of those observations, and they acknowledge that renormalization can explain at most 4.5% of the audited recruitment. The comparison would be stronger if the paper directly tested standard deletion methods (zero/mean/resample) on the same bench and explicitly showed where they land on the dose axis, rather than inferring their positions analytically.
Reproduction appears feasible but not frictionless. The models and Known-Facts corpus are public, and the intervention protocol is described with enough formal detail to reimplement. However, the bench construction relies on human-interpretable labeling of weight-space poles to select truth- and falsity-coding directions, which introduces subjectivity and potential confirmation bias. Core certification and recruitment thresholds use conventional $t \ge 2$ cutoffs that are arbitrary, and the authors promise code release only with the camera-ready version. The nested dose-ladder intervention is computationally expensive, but hyperparameters and fitting weights are documented in Appendices J and K.
Ablate a component of a language model, and other components often appear to adjust and compensate. This phenomenon, termed self-repair, has been observed repeatedly, but its mechanism remains unclear. The most systematic study to date concluded that self-repair is noisy and unlikely to have a single explanation. We argue that it has one: a gain already present before any ablation. Any intervention on a causally important component can be viewed as a point on a coordinate axis $\lambda$, the signed strength of a counterfactual contrast. Hence, conventional ablation methods are uncalibrated points on this axis. We show that the causal repair response for a fine-grained unit $r$ is governed by an affine law, $E_r(\lambda)=\mathrm{own}_r+\gamma_r\lambda$. The slope $\gamma_r$ is a fixed coefficient that consistently influences the model, with or without ablation, and its sign determines whether the unit counteracts or reinforces the removed signal. On a factual-verdict task across four models from distinct families (Gemma, Qwen, LLaMA, and Mistral), we identify components including MLP neurons, OV neurons, and singular directions that follow this affine law, 68 of 81 downstream directions in all. Moreover, we can anticipate the magnitude of $\gamma_r$ from the fixed weights. On the IOI circuit of GPT-2 Small, seven of the ten heads the intervention can reach follow the law, and all seven are counterweights. From this perspective, what may appear as self-repair is a counterweight performing its usual operation when the contrastive signal emerges at the core.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.