Don't Judge an LLM Only by Its Activations: Discovering Suppressed Safety Features via Counterfactual Activation Potential

cs.LG, cs.AI, cs.CL, cs.CR

Submitted: 2026-10-04

Updated: 2026-10-04

Comments: 28 pages, 3 figures, 15 tables. Submitted to ICLR 2027

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Sources

• Understanding intermediate layers using linear classifier probes
• Sparse Autoencoders are Capable LLM Jailbreak Mitigators
• Understanding Jailbreak Success: A Study of Latent Space Dynamics in Large Language Models
• Gemma 2: Improving Open Language Models at a Practical Size
• Attention Slipping: A Mechanistic Understanding of Jailbreak Attacks and Defenses in LLMs
• What Features in Prompts Jailbreak LLMs? Investigating the Mechanisms Behind Attacks
• Efficient Dictionary Learning with Switch Sparse Autoencoders
• Beyond I'm Sorry, I Can't: Dissecting Large Language Model Refusal
• Using Mechanistic Interpretability to Craft Adversarial Attacks against Large Language Models
• Universal and Transferable Adversarial Attacks on Aligned Language Models

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