Linear probing is a common technique in mechanistic interpretability despite the fact that the decodability of a given feature does not necessarily imply the network uses that feature. Intervention faithfulness, not merely decodability, is the appropriate criterion for whether a feature serves a defined active role in the model. Meanwhile, probe regularization has been used for feature discovery, but the impact of regularization on feature causality remains unclear. We study probe regularization systematically, starting with unregularized ordinary least squares (OLS) regression and considering increasing amounts of regularization via ridge regression. We show theoretically that OLS probes amplify low-variance directions in activation space that have spurious correlations with the target feature, whereas regularization reduces influence from these directions. Empirically, we find in a toy transformer and Llama-2-7B that intervention faithfulness increases up to a point, and then decreases again, and that the optimal regularization strength leads to features bearing striking similarity (in both direction and variance) to those learned by optimizing for intervention outcomes directly. Together, these results provide mechanistic justification for regularizing regression probes.
Figures & tables
Testbed
Direction
Probe loss (MSE)
Int. loss (MSE)
Variance ( v⊤Σv )
∠(⋅,DAS)
Mess3 HMM
OLS probe
9.46×10−5
2.14×10−1
1.16×10−8
89.8∘
Ridge probe ( λ=0.1 )
3.99×10−3
1.95×10−3
5.10×100
28.9∘
DAS
3.21×10−3
1.88×10−3
9.72×100
0∘
Llama-2-7B
OLS probe
6.71×10−2
1.37×101
2.10×10−2
84.7∘
(birth year)
Ridge probe ( λ=104 )
2.51×10−1
7.07×100
3.83×100
26.3∘
DAS
6.89×10−1
6.42×100
8.28×100
0∘
Table 1: The two testbeds tell similar stories. The OLS probes achieve the lowest probe loss (sampled, not exhaustive, on Llama-2-7B), yet their directions carry orders of magnitude less variance, are near-orthogonal to DAS, and interventions along them barely move network outputs. By contrast, intervention-optimal ridge regression probes are similar to the DAS directions in probe loss, intervention loss, variance, and direction. DAS is technically not a probe as it does not optimize for decodability, but its reported “probe loss” is nevertheless that of a linear readout fit on its subspace. Also, the intervention loss of 13.7 for the OLS probe on Llama-2-7B is the same as what one gets from not intervening at all, whereas the ridge regression probe (7.1) and DAS (6.4) roughly halve it.
Figure 1: Network output change under intervention ( y ) against the change the ground truth predicts ( x ). On the HMM (top) we steer 20 random start beliefs toward each simplex corner at 20 strengths, and each steering step gives three dots, one per emission probability. On Llama-2-7B (bottom) we run 400 base-source interchanges, and each gives one dot, the shift in expected decade. A faithful direction puts dots on the diagonal, an inert one on a flat line at zero. The dashed lines are the best fits. OLS probes are causally inert whereas intervention-optimal ridge regression probes are mostly intervention-faithful, and similar in intervention effect to DAS.
Figure 2: Probe loss (blue, log scale) and intervention loss (red) as a function of the ridge penalty λ . Probe loss rises monotonically with λ while intervention loss has an interior minimum.
Figure 3: Belief simplex reconstructions for ground truth, the OLS probe, intervention-optimal ridge regression probe and DAS. The OLS probe recovers the simplex near-perfectly yet is causally inert. Although the specific distortion shapes differ, the two intervention-faithful methods both produce visibly distorted reconstructions with similar probe loss (more than a magnitude higher than the OLS probe).
Figure 4: Intervention loss across the belief simplex for the OLS probe (top) vs. the intervention-optimal ridge regression probe (bottom). From each of the three start states (stars), OLS probe interventions have substantially higher loss than intervention-optimal ridge probe interventions.
Figure 5: Probe mass by causal operativity. Each bubble is a principal component, area proportional to its probe mass mi , position its operativity under in-range swaps. OLS mass (red) concentrates at low operativity on both testbeds; intervention-selected ridge mass (green) concentrates at high operativity. The Llama frame shows the heaviest-mass PCs, covering 95% of the ridge probe’s mass but only ∼31% of the OLS probe’s, whose remainder spreads over a long tail of small bubbles piling onto the same low-operativity side.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 6: Intervention loss (red, MSE) and probe loss (blue, MSE) of the birth-year interchange by layer, based on 200 base–source pairs. Each layer has its own intervention-optimized ridge regression probe. Interventions work best at layer 8, the site used for our Llama-2-7B experiments.
Figure 7: Angle from the ridge probe to DAS (green) and to OLS (red) as λ grows. On the HMM study, we report the larger of the two principal angles between the steering subspaces. On the Llama-2-7B study, the angle is between single directions. The intervention-selected λ is dashed. The sweep’s positive case for ridge regression is clearest on the Llama-2-7B experiment, where the principal angle to DAS is minimized at the intervention-optimal λ . The case against OLS is clear on both testbeds.
Figure 8: Standardized birth-year covariance zi=ncovi/(σiσf) for all 4096 PCs of Llama-2-7B layer 8, computed split-half (one half of the prompts fixes the PC directions, the other supplies σ and cov ). 3976 of 4096 PCs lie within 3 standard errors of zero (shaded), the sampling-noise scale of Section 3.2 .
Figure 9: Per-sample intervention effect for DAS confined to the heaviest-OLS-weight principal components: 100 on Llama-2-7B and 7 on the HMM transformer (over 99.9% of OLS probe mass). No causally operative direction is found in either.