Research map / Probabilistic and scientific modeling / Causal discovery
Causal Abstractions of Neural Networks
Atticus Geiger, Hanson Lu, Thomas F Icard, Christopher Potts
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add labeled-subphrase data augmentation to training (ablation) | BERT-base · MQNLI hard generalization test | accuracy: 54.51 -> 88.50 | Improved |
| add dev-set comparison for augmentation (ablation) | BERT-base · MQNLI hard generalization dev | accuracy: 55.42 -> 88.25 | Improved |
| use BERT pretraining instead of no pretraining (ablation) | BERT-base · MQNLI hard generalization | accuracy: 49.1 -> 88.50 | Improved |
| use BERT instead of BiLSTM | MQNLI classifier · MQNLI NP_Obj causal abstraction | clique size (%): 0.7 -> 38.3 | Improved |
| remove N_Obj^P from the NP_Obj composition (ablation) | NP_Obj causal model · MQNLI on BERT | clique size (%): 38.3 -> 33.8 | Hurt |
| remove A_Obj^H from the NP_Obj composition (ablation) | NP_Obj causal model · MQNLI on BERT | clique size (%): 38.3 -> 15.7 | Hurt |
| remove N_Obj^H and N_Obj^P from the NP_Obj composition (ablation) | NP_Obj causal model · MQNLI on BERT | clique size (%): 38.3 -> 14.1 | Hurt |
| add N_Subj^P to the NP_Obj composition (ablation) | NP_Obj causal model · MQNLI on BERT | clique size (%): 38.3 -> 37.2 | Hurt |
| add Adj_Subj^P to the NP_Obj composition (ablation) | NP_Obj causal model · MQNLI on BERT | clique size (%): 38.3 -> 30.5 | Hurt |
| use interchange interventions instead of probing to locate representations | BERT analysis · MQNLI NP_Obj | discriminativeness of location | Improved |
| use selectivity instead of probe accuracy | probing · MQNLI NP_Obj | probe result | No clear effect |
Closest papers by meaning
- The Non-Linear Representation Dilemma: Is Causal Abstraction Enough for Mechanistic Interpretability? (NeurIPS 2025 · similarity 0.91)
- Neural Causal Abstractions (AAAI 2024 · similarity 0.89)
- Internal Causal Mechanisms Robustly Predict Language Model Out-of-Distribution Behaviors (ICML 2025 · similarity 0.88)
- Causal Concept Graph Models: Beyond Causal Opacity in Deep Learning (ICLR 2025 · similarity 0.87)
- Causal Abstraction Inference under Lossy Representations (ICML 2025 · similarity 0.87)
- Can Large Language Models Infer Causation from Correlation? (ICLR 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
