Right Answer, Wrong Mechanism: Detecting Pernicious Divergence in Causal Interventions
cs.LG
Submitted: 2026-09-30
Updated: 2026-10-04
Code: https://github.com/TransformerLensOrg/TransformerLens
Sources
- Addressing divergent representations from causal interventions on neural networks
- How to use and interpret activation patching
- Learning a Generative Meta-Model of LLM Activations
- Is This the Subspace You Are Looking for? An Interpretability Illusion for Subspace Activation Patching
- The Hydra Effect: Emergent Self-repair in Language Model Computations
- Qwen2.5 Technical Report
- The Curse of Multiple Mediators: Hidden Interaction Effects in Activation Patching
- A Reply to Makelov et al. (2023)'s "Interpretability Illusion" Arguments
Related papers
- Polynomial-Augmented Neural Networks (PANNs) with Weak Orthogonality Constraints for Enhanced Function and PDE Approximation
- AIRL-S: Unifying Reinforcement Learning and Search-Based Test-Time Scaling via Adversarial Inverse Reinforcement Learning
- Transformers as Bayesian In-Context Experimenters: Smoothness-Adaptive Efficient ATE Estimation
- Convergence issues in Relational Concept Analysis based on AOC-posets
- Beliefs Beyond Posteriors: The Predictive Cost of ELBO Binding
- Understanding Diffusion Models via Ratio-Based Function Approximation with SignReLU Networks