Path patching provides a method to express and quantitatively test hypotheses that neural network behaviors are localized to sets of paths.
Title resolution pending
17 Pith papers cite this work. Polarity classification is still indexing.
17
Pith papers citing it
citation-role summary
background 2
method 1
citation-polarity summary
representative citing papers
Grokking arises from gradual amplification of a Fourier-based circuit in the weights followed by removal of memorizing components.
GPT-2 small solves indirect object identification via a circuit of 26 attention heads organized into seven functional classes discovered through causal interventions.
Introduces Causal Functional Signatures grounded in causal evidence and ILP-learned architectural signatures to enable explicit, comparable, and portable mechanistic claims across model scales.
Pretraining and alignment induce asymmetric geometric traces in transformer weights because alignment updates concentrate in read pathways due to activation covariance while write pathways inherit less structure from alignment losses.
Language model circuits show high within-task consistency and necessity but substantial overlap across tasks, making them less specific than assumed.
Refusal in language models is mediated by a single direction in residual stream activations that can be erased to disable safety or added to elicit refusal.
Gated SAEs decouple which features to use from how large their activations should be, applying the L1 penalty only to selection and thereby eliminating shrinkage while halving the number of firing features needed for good fidelity.
Training per-layer affine probes on frozen transformers yields more reliable latent predictions than the logit lens and enables detection of malicious inputs from prediction trajectories.
SafeDiffusion-R1 uses online GRPO with CLIP embedding steering to cut inappropriate content from 48.9% to 18.07% and nudity detections from 646 to 15 in diffusion models while raising GenEval scores from 42.08% to 47.83% and generalizing across seven harm categories without supervised pairs or extra
Interpretability research should be judged by actionability—the degree to which its insights support concrete decisions and interventions—rather than explanatory power alone.
Sentiment is represented as a single linear direction in LLM activation space that is causally relevant across tasks and is summarized at punctuation and names in addition to charged words.
TIDE augments standard transformers with per-layer token embedding injection via an ensemble of memory blocks and a depth-conditioned router to mitigate rare-token undertraining and contextual collapse.
KL divergence of attention heads from uniform distribution predicts LLM answer correctness across datasets and model families.
Varying evaluation metrics and corruption methods in activation patching produces different localization and circuit discovery outcomes in language models, leading to recommendations for preferred practices.
A mechanics of the learning process is emerging in deep learning theory, characterized by dynamics, coarse statistics, and falsifiable predictions across idealized settings, limits, laws, hyperparameters, and universal behaviors.
citing papers explorer
-
Progress measures for grokking via mechanistic interpretability
Grokking arises from gradual amplification of a Fourier-based circuit in the weights followed by removal of memorizing components.
-
Refusal in Language Models Is Mediated by a Single Direction
Refusal in language models is mediated by a single direction in residual stream activations that can be erased to disable safety or added to elicit refusal.
-
Eliciting Latent Predictions from Transformers with the Tuned Lens
Training per-layer affine probes on frozen transformers yields more reliable latent predictions than the logit lens and enables detection of malicious inputs from prediction trajectories.