Paper Reviews, Security News & Technical Analysis
9 September 2026
Microsoft's September 2026 Patch Tuesday issued 974 fixes, a record driven by AI vulnerability discovery. Analysis of the remediation bottleneck and what it mea
CybersecurityMicrosoftPatch ManagementAI SecurityVulnerability Research9 September 2026
A deep technical analysis of e-infinity.space's interactive topological renderer, covering NPR hatching pipelines, principal curvature streamlines, and GPU visi
Computer GraphicsTopologyNon-Photorealistic RenderingMathematicsWebGL8 September 2026
Harbor Adapters reduces agentic benchmark integration from O(mn) to O(m+n), enabling large-scale evaluation across 54 benchmarks and 8 models with key findings
BenchmarkingLLM AgentsEvaluation InfrastructureAI Research8 September 2026
New research reveals LLMs integrate external evidence as a receiver-specific distribution shift, not scalar trust, with major implications for RAG and tool-use
Large Language ModelsRAGMechanistic InterpretabilityAI SafetyNLP Research8 September 2026
Iris-mini and Iris-pro achieve state-of-the-art search agent performance through web-graph task synthesis, iterative SFT-RL training, and careful context manage
Search AgentsReinforcement LearningLanguage ModelsAgentic AIInformation Retrieval5 September 2026
A prototype AI-driven English textbook tested on 186 undergraduates improved completion accuracy, speaking scores, and cut teacher correction time by 31.6%.
AI in EducationAdaptive LearningNatural Language ProcessingEdTechEnglish Language Teaching5 September 2026
New research shows LLMs store unigram priors along a single vector in embedding space, with a scalar factor tracking how much context informs each prediction.
Mechanistic InterpretabilityLanguage ModelsBayesian InferenceEmbeddingsScaling Laws5 September 2026
A structured ablation of LLM agent scaffolding finds optimization value concentrated in one slot, with uniform budget-splitting actively destroying gains.
LLM AgentsPrompt OptimizationCredit AssignmentAgent ScaffoldingAblation Studies4 September 2026
A new benchmark reveals frontier LLMs reliably detect when they are being evaluated, with methodological biases skewing existing measurements by up to 0.08 AURO
AI SafetyBenchmarkingLarge Language ModelsEvaluationFrontier Models4 September 2026
PRO-Step trains a generative PRM to evaluate logical validity and evidential grounding at each reasoning step, outperforming outcome-based RAG methods across fi
Retrieval-Augmented GenerationProcess Reward ModelsMulti-hop ReasoningReinforcement LearningNLP4 September 2026
A controlled study shows prompt-space meta-learning for frozen LLM personalisation collapses due to a structure-blind meta-objective, with retrieval dominating
LLMPersonalisationMeta-LearningNegative ResultsPrompt Optimisation3 September 2026
REAL-Q replaces static Hessian solvers in post-training quantization with dynamic block-wise gradient descent, cutting KL divergence by up to 49% on LLaMA and Q
QuantizationLLMsOptimizationEfficiencyNLP