Experiment Families

Each experiment family is a registered class that implements a standardized run() interface. Families are declared in src/mech_interp/experiments/families.py and implemented in src/mech_interp/experiments/<family>.py.

Registered families

Family key	Description	Source paper(s)	Implementation
polysemanticity	Baseline polysemanticity probe — measures superposition in MLP neurons	Elhage et al. 2022, "Toy Models of Superposition"	experiments/polysemanticity.py
polysemanticity_sae	Top-K Sparse Autoencoder training on a hook site; reports live/dead features, explained variance, decoder norms	Cunningham et al. 2023; Bricken et al. 2023; Gao et al. 2024	experiments/polysemanticity_sae.py
superposition	Direct superposition measurement via singular value analysis of weight matrices	Elhage et al. 2022	experiments/polysemanticity.py
circuit_patching	Activation patching (resample ablation) at any (layer, site) pair to measure causal contribution	Wang et al. 2022, "Interpretability in the Wild"	experiments/circuit_patching.py
acdc_lite	Lightweight ACDC — edge-level circuit discovery without full path enumeration	Conmy et al. 2023, "Towards Automated Circuit Discovery"	experiments/acdc_lite.py
acdc_edge	Full ACDC with IOI validation; exhaustive edge attribution and pruning	Conmy et al. 2023	experiments/acdc_edge.py
refusal_direction	Extract and characterize the linear refusal direction in residual stream via PCA/SVM on harmful vs. harmless activations	Arditi et al. 2024, "Refusal in LLMs is mediated by a single direction"	experiments/refusal_direction.py
caa_steering	Contrastive Activation Addition — inject a steering vector and measure behavioral shift	Turner et al. 2023, "Activation Addition: Steering Language Models Without Optimization"	experiments/caa_steering.py
logit_lens	Project residual stream at every layer through the unembedding; track per-token rank and probability	nostalgebraist 2020; Belrose et al. 2023, "Eliciting Latent Predictions from Transformers"	experiments/logit_lens.py
direct_logit_attribution	Decompose final logit into per-component (MLP, attention head, embed) contributions	Elhage et al. 2021, "A Mathematical Framework for Transformer Circuits"	experiments/direct_logit_attribution.py
attribution_patching	Gradient x activation attribution patching — efficient approximation of activation patching	Nanda 2023, "Attribution Patching: Activation Patching At Industrial Scale"	experiments/attribution_patching.py
sparse_probing	Fit sparse linear probes on residual stream activations to detect linear representations	Gurnee et al. 2023, "Finding Neurons in a Haystack"	experiments/sparse_probing.py
cross_model_representation_probe	Compare representations across model families by projecting into a shared probe space	—	experiments/cross_model_representation_probe.py
sae_cross_model	Compare SAE feature dictionaries learned from different models on the same corpus	Lindsey et al. 2024, "Sparse Crosscoders for Cross-Layer Features"	experiments/sae_cross_model.py
crosscoder	Train a sparse crosscoder to find shared features across layers or model variants	Lindsey et al. 2024	experiments/crosscoder.py
causal_scrubbing	Causal scrubbing — test hypotheses about computational circuits by re-running forward passes with controlled substitutions	Chan et al. 2022, "Causal Scrubbing"	experiments/causal_scrubbing.py

Backend support

Family	TransformerLens	NNsight	MLX
polysemanticity_sae	Yes	—	—
circuit_patching	Yes	—	—
acdc_lite / acdc_edge	Yes	—	—
refusal_direction	Yes	Yes	—
caa_steering	Yes	Yes	—
logit_lens	Yes	—	—
direct_logit_attribution	Yes	—	—
attribution_patching	Yes	—	—
sparse_probing	Yes	Yes	—
causal_scrubbing	Yes	—	—
crosscoder	Yes	—	—

Adding a new family

1. Add a StrEnum entry to ExperimentFamily in families.py.
2. Implement class MyFamily(BaseExperiment) in experiments/my_family.py.
3. Register it in experiments/registry.py.
4. Add a YAML spec under experiments/.
5. Add unit tests in tests/test_my_family.py.

See experiments/polysemanticity_sae.py for a reference implementation.