LLM Security
Special Token Attacks: The 96% LLM Jailbreak Exploit
Security researchers uncover how special tokens in LLM architectures create hidden attack surfaces, enabling jailbreak success rates as high as 96% across major models.
Machine Learning
LLM Research
Soft Recursive Least-Squares: A New Approach to Lifelong LLM Edit
Researchers introduce S-RLS, a novel method for continuous LLM knowledge updates that avoids catastrophic forgetting through soft memory preservation instead of rigid constraints.
neuro-symbolic AI
Tensor Networks Bridge Neural and Symbolic AI Reasoning
New research proposes tensor network mathematics to unify neural networks with symbolic AI, potentially enabling more interpretable and reasoning-capable AI systems.
synthetic data
New Research Exposes Key Limitations of Learning from Synthetic D
Researchers analyze why Empirical Risk Minimization fails when models train on synthetic data, revealing fundamental barriers that affect AI video generation and deepfake systems.
Generative AI
Ambient Dataloops: Using AI to Refine Its Own Training Data
New research explores how generative models can iteratively improve their own training datasets, potentially enhancing quality across AI video, image synthesis, and synthetic media generation.
LLM Agents
New Testing Framework Ensures LLM Agents Behave Predictably
Researchers introduce a determinism-faithfulness assurance harness for tool-using LLM agents, enabling reliable replay testing to catch unpredictable AI behavior in critical applications.
LLM Agents
Aeon: Neuro-Symbolic Memory Boosts Long-Horizon LLM Agents
New research introduces Aeon, a memory management system combining neural and symbolic approaches to help LLM agents maintain coherent reasoning across extended task sequences.
LLM alignment
GRADE: New Backpropagation Method Replaces Policy Gradients for L
Researchers introduce GRADE, a technique that replaces traditional policy gradient methods with direct backpropagation for aligning large language models, potentially offering more efficient training.
LLM Research
New Research Maps LLM Embeddings Using Hamiltonian Physics
Researchers propose a physics-inspired framework treating LLM token embeddings as discrete semantic states governed by Hamiltonian dynamics, offering new insights into transformer interpretability.
LLM Agents
New LLM Agent Framework Tackles ML Feature Engineering Reliabilit
Researchers propose a constrained-topology planning approach for LLM agents that improves reliability in automated feature engineering, addressing key challenges in ML pipeline automation.
LLM Research
How Similarity Retrieval Creates Reasoning Biases in LLMs
New research reveals how LLMs develop 'directional attractors' during reasoning tasks, showing that similarity-based retrieval mechanisms systematically steer iterative summarization toward predictable patterns.
AI research
Research Challenges AI Certainty-Scope Trade-Off Assumptions
New research formally disproves the assumed universal trade-off between certainty and scope in AI systems, with implications for how we understand LLM reliability and knowledge boundaries.