The field of large language models (LLMs) is rapidly evolving, with a growing focus on safety and interpretability. Recent research has highlighted the importance of understanding how LLMs learn and represent knowledge, as well as the need to develop effective methods for detecting and preventing harmful behavior. One key area of research is the development of techniques for analyzing and interpreting LLM internals, such as concept-driven neuron attribution and activation transport operators. These methods have the potential to provide valuable insights into how LLMs work and how they can be improved. Another important area of research is the development of safety guardrails and defense mechanisms, such as speculative safety-aware decoding and prompt injection detection. These mechanisms can help to prevent LLMs from being used for malicious purposes and ensure that they are used in a responsible and beneficial way. Notable papers in this area include 'A Review of Developmental Interpretability in Large Language Models', which provides a comprehensive overview of the field of developmental interpretability, and 'NEAT: Concept driven Neuron Attribution in LLMs', which proposes a new method for locating significant neurons in LLMs. Overall, the field of LLM safety and interpretability is rapidly advancing, with new techniques and methods being developed to address the challenges and risks associated with these powerful models.
Advances in Large Language Model Safety and Interpretability
Sources
A Review of Developmental Interpretability in Large Language Models
NEAT: Concept driven Neuron Attribution in LLMs
PickleBall: Secure Deserialization of Pickle-based Machine Learning Models
Time Series Based Network Intrusion Detection using MTF-Aided Transformer
LLMSymGuard: A Symbolic Safety Guardrail Framework Leveraging Interpretable Jailbreak Concepts
Retrieval-Augmented Defense: Adaptive and Controllable Jailbreak Prevention for Large Language Models
Confusion is the Final Barrier: Rethinking Jailbreak Evaluation and Investigating the Real Misuse Threat of LLMs
HAMSA: Hijacking Aligned Compact Models via Stealthy Automation
QueryBandits for Hallucination Mitigation: Exploiting Semantic Features for No-Regret Rewriting
ObjexMT: Objective Extraction and Metacognitive Calibration for LLM-as-a-Judge under Multi-Turn Jailbreaks
Towards Safeguarding LLM Fine-tuning APIs against Cipher Attacks
Activation Transport Operators
School of Reward Hacks: Hacking harmless tasks generalizes to misaligned behavior in LLMs
MalLoc: Toward Fine-grained Android Malicious Payload Localization via LLMs
Speculative Safety-Aware Decoding
Learning from Few Samples: A Novel Approach for High-Quality Malcode Generation
Neither Valid nor Reliable? Investigating the Use of LLMs as Judges
Tricking LLM-Based NPCs into Spilling Secrets
Stand on The Shoulders of Giants: Building JailExpert from Previous Attack Experience
Language Models Identify Ambiguities and Exploit Loopholes
The Art of Hide and Seek: Making Pickle-Based Model Supply Chain Poisoning Stealthy Again
FlowletFormer: Network Behavioral Semantic Aware Pre-training Model for Traffic Classification
Evaluating Language Model Reasoning about Confidential Information
Forewarned is Forearmed: Pre-Synthesizing Jailbreak-like Instructions to Enhance LLM Safety Guardrail to Potential Attacks
Smart Contract Intent Detection with Pre-trained Programming Language Model
IntentionReasoner: Facilitating Adaptive LLM Safeguards through Intent Reasoning and Selective Query Refinement
FlowMalTrans: Unsupervised Binary Code Translation for Malware Detection Using Flow-Adapter Architecture
Ransomware 3.0: Self-Composing and LLM-Orchestrated
Token Buncher: Shielding LLMs from Harmful Reinforcement Learning Fine-Tuning
JADES: A Universal Framework for Jailbreak Assessment via Decompositional Scoring
PromptSleuth: Detecting Prompt Injection via Semantic Intent Invariance