Researchers claim a new safety technique called Tripwire reduces jailbreak attack success rates to under 2% while keeping utility loss between 0.5% and 5.3% on MT-Bench, using per-neuron hypothesis te…
A developer building a production AI agent in Spring Boot found that a system prompt change based on intuition lost in a pairwise A/B test against the old prompt, despite appearing better in manual ch…
Researchers introduced JudgeArena, an open-source framework unifying major LLM-judge benchmarks (AlpacaEval, Arena-Hard, MT-Bench, and m-Arena-Hard) under a single interface with swappable judges and …
Researchers introduce MeRLa (Meta-Learned Reward Shaping), a framework that meta-learns a task-aware shaping function for Reinforcement Learning from Human Feedback (RLHF) to improve alignment of larg…
Three prompting techniques that worked in 2023 now actively cost accuracy in 2026, according to recent research. Expert personas reduced factual accuracy to 68.0% versus 71.6% without on the MMLU know…
Researchers introduced Masked Language Flow Models (MLFMs), combining masked diffusion and flow-based methods for efficient language generation. MLFMs enable conditional generation via continuous flow…
A systematic evaluation of 21 LLM-as-a-Judge models across 118 runs and 541,000 judgments reveals that exact-match agreement overstates discriminative ability, with kappa deflation of 33–41 percentage…
A new study reveals that LLM judges, widely used to evaluate AI outputs, can be manipulated after making an initial decision through targeted conversation, overturning stable judgments and shifting be…