LLM Interpretability
Brain-Grounded Axes: Reading and Steering LLM Internal States
New research maps LLM internal representations to brain-derived axes, enabling interpretable reading and targeted steering of model behavior without fine-tuning.
AI safety
LLM Security
AdvJudge-Zero: Adversarial Tokens Can Flip LLM Evaluator Decision
New research reveals how adversarial control tokens can manipulate LLM-as-a-Judge systems into completely reversing their binary decisions, exposing critical vulnerabilities in AI evaluation pipelines.
AI safety
Bayesian Uncertainty Methods Could Make AI Systems More Honest
New research explores how Bayesian uncertainty quantification in neural QA systems can improve AI reliability by enabling models to recognize and communicate their own limitations.
mechanistic interpretability
SALVE: New Technique Enables Mechanistic Control of Neural Networ
Researchers introduce SALVE, combining sparse autoencoders with latent vector editing for precise mechanistic control over neural network behaviors and outputs.
AI security
API Access Enables AI Model Cloning and Safety Bypass
New research reveals how anyone with API access can clone AI models and strip away safety guardrails, creating unregulated copies capable of generating harmful content.
LLM Research
ReflCtrl: Steering LLM Reasoning via Representation Engineering
New research introduces ReflCtrl, a method for controlling when large language models engage in extended reasoning by manipulating internal representations rather than prompts.
AI safety
AI Models Can Learn to Hide Thoughts From Safety Monitors
New research reveals language models can learn to conceal internal states from activation-based monitoring systems, raising critical questions for AI safety and detection systems.
AI safety
New Research Proposes Behavioral Guidance for Trustworthy AI Agen
Researchers present a framework for making multi-turn LLM agents more trustworthy through behavioral guidance, addressing critical safety concerns as AI systems become more autonomous.
LLM Research
Study Reveals How Users Build Trust When LLMs Hallucinate
New research examines how users develop calibrated trust strategies when interacting with hallucination-prone LLMs, offering frameworks for safer human-AI collaboration.
DeepMind
DeepMind Expands UK AI Security Institute Partnership
Google DeepMind deepens collaboration with UK AI Security Institute on frontier AI safety evaluation, establishing frameworks that could shape how synthetic media and generative models are assessed globally.
LLM Agents
SABER Framework Tackles Error Cascades in LLM Agents
New research introduces SABER, a safeguarding framework that identifies how small errors in LLM agent actions can cascade into significant failures, proposing intervention mechanisms.
AI safety
CCA Framework: Lifecycle Supervision for Aligned AI Agents
New research proposes Cognitive Control Architecture, a supervision framework designed to maintain AI agent alignment throughout their operational lifecycle through structured oversight mechanisms.