§ BEAT
Research
New DRPO Method Fixes Long-Tail Vocabulary Collapse in LLM RL
Router Matching 50 Retries with 10 Samples Cuts LLM Test-Time Compute
SafeSteer cuts alignment tax by targeting sparse safety tokens
Claude Code Spent 58% of Sessions Optimizing a Broken Architecture
RLHF Training Amplifies Model Bias to 100 Percent
MemAudit Cuts Memory-Poisoning Attacks to 0%
Rensselaer and IBM Expose KV Cache Leakage in Multi-Agent LLMs
Matching Principle Unifies Seven Robustness Families
Self-Modifying Agents Boost Benchmark Score to 0.61
LCGuard Patches KV-Cache Leakage in Multi-Agent Systems
Fine-tuning erases reasoning chains while accuracy stays high
Medical LLMs Underweight Patient Autonomy
Microsoft Finds GPT-5 Fails Against Implausible Attacks
LLM Formalization Catches 18.8% Ambiguous Requirements in Safety Specs
Negation Neglect Drives False Belief Rate to 88.6% in Fine-Tuned LLMs
Reward Hacking Undetected in Single-Verifier Training
Google's RubricEM trains research agents without ground truth
Every Guardrail Classifier Tested Fails Formal Safety Verification
AI Agents Bypass Software Engineering, Risk Production Failure
CIVeX Logs Zero False Executions in Confounded Workflows
Paper Dismantles Causal Discovery Claim in Prediction Models
Flow-OPD Raises Stable Diffusion Accuracy to 92 From 63
Conformal Path Reasoning cuts knowledge graph answer sets by 40 percent
Longer Context Degrades LLM Cooperation, Study Finds
Math AI Training Solver Accuracy Rises 21.4% With Verifier-Backed Generation
Q2RL Reaches 100% Success on Peg Insertion, Outpacing BC and IBRL
Dreadnode Framework Cuts AI Red Teaming from Weeks to Hours
Staging malicious requests bypasses safety in 9 coding agents
LLM hallucination detector beats eight baselines without retraining
