Challenge: Despite substantial progress in safety alignment techniques, aligned large language models can still produce unsafe responses under minor internal perturbations.
Approach: They introduce Activation Steering Attack (ASA) and leverage the Negative Log-Likelihood (NLL) as a diagnostic signal to probe the local sensitivity of safety behaviors in latent space.
Outcome: The proposed method is model-agnostic and supervision-free, enabling a general and reproducible diagnostic metric for analyzing safety robustness.

Similar Papers

Preventing Safety Drift in Large Language Models via Coupled Weight and Activation Constraints (2026.findings-acl)

Copied to clipboard

Challenge: Existing defenses constrain either weights or activations in isolation, without considering their coupled effects on safety.
Approach: They propose a weight-activation constraint that enforces a precomputed safety subspace on weight updates and applies regularization to safety-critical features identified by sparse autoencoders.
Outcome: The proposed model outperforms baselines even under high harmful data ratios.
Why Safeguarded Ships Run Aground? Aligned Large Language Models’ Safety Mechanisms Tend to Be Anchored in The Template Region (2025.acl-long)

Copied to clipboard

Challenge: Infilling a fixed template between the input instruction and initial model output is a common practice for existing LLMs, but it is vulnerable to inference-time jailbreak attacks.
Approach: They propose to fill a fixed template between the input instruction and initial model output and to detach safety mechanisms from the template region to mitigate the risk of inference-time jailbreak attacks.
Outcome: The proposed method is widespread across aligned LLMs and shows that it mitigates inference-time jailbreak vulnerabilities.
How Alignment and Jailbreak Work: Explain LLM Safety through Intermediate Hidden States (2024.findings-emnlp)

Copied to clipboard

Challenge: Large language models (LLMs) rely on safety alignment to avoid malicious user inputs.
Approach: They employ weak classifiers to explain LLM safety through the intermediate hidden states.
Outcome: The proposed model can identify malicious and normal inputs and detect malicious ones without jailbreak.
Unraveling LLM Jailbreaks Through Safety Knowledge Neurons (2026.eacl-long)

Copied to clipboard

Challenge: Large Language Models (LLMs) have achieved significant progress in alignment, ensuring safer and more reliable outputs.
Approach: They propose a neuron-level interpretability method that focuses on the role of safety-related knowledge neurons to improve model robustness against jailbreak attacks.
Outcome: The proposed method reduces attack success rates across multiple LLMs and outperforms all baseline defenses.
Analysing the Safety Pitfalls of Steering Vectors (2026.findings-acl)

Copied to clipboard

Challenge: Activation steering has emerged as a powerful tool to shape LLM behaviour without the need for weight updates.
Approach: They propose to audit steering vectors obtained with Contrastive Activation Addition (CAA) and propose a mechanistic explanation for this finding.
Outcome: The proposed approach significantly improves the success rate of jailbreak attacks, with stronger amplification under simple template-based attacks.
Jailbreaks as Inference-Time Alignment: A Framework for Understanding Safety Failures in LLMs (2026.eacl-long)

Copied to clipboard

Challenge: Large language models are safety-aligned to prevent harmful response generation . prior work on jailbreak effectiveness has focused on analyzing success rate of jailbreaks .
Approach: They propose to frame jailbreaks as inference-time alignment and draw suboptimal bounds . they also propose a Safety-Net to measure how vulnerable an LLM is to jailbreak attacks .
Outcome: a new framework allows researchers to show how vulnerable an LLM is to jailbreaks . a Safety-Net measures how vulnerable the model is to attacks, the authors say .
Steering Safely or Off a Cliff? Rethinking Specificity and Robustness in Inference-Time Interventions (2026.eacl-long)

Copied to clipboard

Challenge: Existing studies have shown that model steering can preserve fluency and unrelated abilities, but it fails to preserve robustness specificity.
Approach: They propose a framework that distinguishes three dimensions of specificity: general, control, and robustness.
Outcome: The proposed framework distinguishes three dimensions of specificity: general (preserving fluency and unrelated abilities), control (preserving related control properties), and robustness (preserving control properties under distribution shifts).
Safe-Unsafe Concept Separation Emerges from a Single Direction in Language Models Activation Space (2026.eacl-long)

Copied to clipboard

Challenge: Existing approaches to ensuring the safety of Large Language Models (LLMs) rely on invasive fine- tuning or external generation-based checks, which can be opaque and resource-inefficient.
Approach: They propose a mechanistic method that identifies the layer where safe and unsafe concepts are maximally separable within a pretrained representation space.
Outcome: The proposed method can be used across multiple domains, diverse tasks, and 16 non-English languages on encoder and decoder architectures.
On the Vulnerability of Safety Alignment in Open-Access LLMs (2024.findings-acl)

Copied to clipboard

Challenge: Large language models (LLMs) are susceptible to malicious exploitation, but are often rejected and limited harmfulness is limited.
Approach: They propose two types of reverse alignment techniques: reverse supervised fine-tuning (RSFT) and reverse preference optimization (RPO).
Outcome: The proposed methods can significantly enhance the success rate and harmfulness of jailbreak attacks, but they face high rejection rates and limited harmfulness.
Beyond Surface Alignment: Rebuilding LLMs Safety Mechanism via Probabilistically Ablating Refusal Direction (2025.findings-emnlp)

Copied to clipboard

Challenge: Jailbreak attacks pose persistent threats to large language models . current safety alignment methods have insufficient safety alignment depth and unrobust internal defense mechanisms.
Approach: a new safety alignment framework is developed to overcome jailbreak attacks . the framework forces the model to dynamically rebuild its refusal mechanisms from jailbreak states .
Outcome: a new safety alignment framework reduces attack success rates by approximately 95% on four open-source LLM families and six representative attacks.

What is GenGO?

GenGO is an NLP powered publication search system. It currenctly indexes 30k+ papers from ACL Anthology, and implements multi-aspect summarization, semantic search, and more!

Information

About
Limitations