ResearchRandom Cloud Finds Minimal Neural Networks Without Training
Training-free method discovers feedforward architectures through stochastic exploration, challenging conventional NAS approaches.
Latest
67 articles total
ResearchLLM Judges Show Systematic Bias Toward Own Outputs
New research quantifies how language models favor their own responses in evaluation tasks, threatening benchmark validity.
ResearchHow LLMs Detect and Correct Their Own Errors Without External Feedback
New research reveals internal confidence signals enable models to identify reasoning failures autonomously, reshaping debugging approaches.
ResearchPExA Agent Balances Speed and Accuracy in Text-to-SQL
New method parallelizes SQL generation to resolve the latency-performance tradeoff in LLM-based database queries.
ResearchLinear-Time B-splines KAN Architecture Reduces Computational Cost
New approach accelerates Kolmogorov-Arnold Networks while maintaining expressibility and interpretability.
ResearchLLMs Fail to Generate Random Numbers From Statistical Distributions
New research reveals large language models cannot faithfully sample from probability distributions, a critical gap for stochastic systems.
ResearchContent Moderation Systems Fail When Multiple Answers Are Right
New research exposes how agreement metrics penalize valid AI decisions in rule-governed environments.
ResearchAI System Automates Military Course of Action Planning
Researchers develop automated CoA generation to match acceleration of modern warfare speeds.
ResearchEnterprise AI Agent Systems Fail at Rates Up to 86.7 Percent
New research identifies coordination failures and specification mismatches as primary causes of multi-agent LLM system breakdowns.
ResearchNew Method Slashes Memory Cost of Training Deep Networks
BASIS technique reduces activation memory scaling, freeing up GPU resources for larger models.