Sparse autoencoders (SAEs) compress model activations into sparse codes, but equal reconstruction error and sparsity can preserve different linearly decodable signals. We formalize this ambiguity as a matrix-valued distortion between optimal ridge-prediction operators and train decoder-preserving SAEs by combining this distortion with reconstruction loss. In a rank relaxation, an isotropic task prior saturates per-mode omission costs without changing PCA's ordering, whereas a structured prior can change which modes are retained. A controlled sparse experiment shows that a declared prior protects held-out combinations from its task subspace. On GPT-2 small block 8, DPSAE reduces held-out decoder distortion by 10.6--11.4% across three paired runs while matching reconstruction NMSE. The same checkpoints pass an average natural-text output-KL noninferiority test, but one matched Pythia pair shows no improvement in probes restricted to a few sparse features. These results show that reconstruction quality does not determine which refitted linear readouts survive sparse compression, and that readout preservation is distinct from learning cleaner benchmark concepts or preserving every frozen-model behavior.
Sankaran Vaidyanathan, David Arbour, Aaron Mueller +2cs.LG cs.CL
Activation patching is the primary tool in mechanistic interpretability. It attributes causal responsibility for a model behavior to each of its individual components by estimating its natural indirect effect (NIE). Re-deriving the activation patching estimand from causal mediation analysis, we find that the NIE does not solely capture the causal effect through the specific component. It also contains interaction effects (INT) that measure how much the component's causal effect itself depends on the state of other components in the model. A natural response may be to try to eliminate INT by adjusting the estimator or unit of analysis, but each of these potential remedies has predictable failure modes. We demonstrate these failure modes in the GPT-2 IOI circuit; components whose causal importance is conditional on the state of other components are either invisible or artificially inflated, and INT variance explains the previously documented instability of faithfulness scores. We prove that INT scales with the distance between clean and patched component activations, is negligible when the model is locally affine, and decomposes combinatorially into pairwise and higher-order group interactions. Despite its inevitability, INT is not a nuisance to be eliminated, but rather a diagnostic for interpretability studies. Its individual and group-level magnitude and sign signal when causal conclusions are prompt-dependent, and when greedy NIE-based component ranking will miss mechanisms only discoverable through combinatorial search.
Weight-space geometry plays a central role in neural network optimization, yet manifold constraints are often applied uniformly across all weight matrices. In this work, we ask whether different transformer modules prefer different manifold geometries. We study Manifold Muon for GPT-2 pretraining and compare layer-wise assignments of Stiefel and DGram constraints across attention and MLP blocks. Our results show a clear asymmetry: constraining attention layers with Stiefel geometry while assigning DGram geometry to MLP layers gives the best performance among the tested configurations, whereas the inverted assignment and all-DGram configuration become unstable under the shared hyperparameter setting. We trace this failure to singular value growth in DGram-constrained attention weights, which can amplify attention logits and induce softmax saturation. These findings suggest that symmetry-aware and geometry-aware optimization for transformers should be module-specific rather than uniform.
We report a small, reproducible audit of which sparse-autoencoder (SAE) features of GPT-2 small fire differently on failed versus successful trials of the Indirect Object Identification (IOI) task. On 300 prompts, GPT-2 small reaches 79.7% accuracy; 146 of the 24,576 features in the layer-8 residual-stream SAE release of Bloom (2024) clear a Holm-corrected significance threshold and 105 reach a large effect size (|Cohen's d| > 0.8). The strongest single correlate of failure -- feature 17,491, d=+2.93, Neuronpedia label 'cryptographic keys' -- is essentially silent except when the prompt's transferred object is 'the keys,' on which GPT-2 small fails 93.3% of the time vs. 7.5% on the other seven objects (Fisher exact p = 8.79 x 10^-33). We put this correlate through three controls that a mechanistic claim should pass. (i) A causal ablation: zeroing feature 17,491 in the residual stream across all token positions of the 45 keys prompts does not restore accuracy (6.7% -> 4.4%); the feature is a correlate, not a sufficient cause at this layer. (ii) A representation baseline: a logistic regression on the raw 768-dimensional residual stream reaches 5-fold ROC AUC = 0.929, matching the top-100 SAE features (0.927); the SAE basis adds interpretability, not predictive power. (iii) A seed-robustness check: across five random seeds the keys-subset failure rate stays in 75.0--93.3% (the behavioural effect is real), but feature 17,491 is the top-|d| feature in only 1 of 5 runs. The methodological contribution is therefore the audit pipeline (cheap, model-agnostic, surfaces named correlates) rather than any single feature found through it. We release the code, the 300-prompt corpus, the 300x24,576 activation matrix, the ablation and baseline scripts, and the figures. The full pipeline runs on a laptop (Apple M3 Max, no discrete GPU).