Brief

A new method called SkillOpt, developed by Microsoft and three Chinese universities, has demonstrated that a single Markdown file can significantly improve AI agent performance. When used as context during inference, this file boosted GPT-5.5's scores by an average of 23 points across six procedural benchmarks. This approach outperformed handwritten instructions, LLM-generated instructions, and four specialized training methods, suggesting a notable trend for AI agent optimization. AI

Researchers have developed GEPA, a new framework designed to enhance the problem-solving capabilities of language models, particularly for arithmetic word problems. This system begins with basic prompts and iteratively refines them by creating deterministic benchmarks, establishing structured evaluation methods, and simultaneously evolving both the instructions and the format of the model's output. The improvements demonstrated by GEPA have shown to generalize effectively to new, unseen datasets. AI

Researchers have developed a new method called P^2O (Joint Policy and Prompt Optimization) to address the issue of advantage collapse in Reinforcement Learning with Verifiable Rewards (RLVR) for large language models. This technique alternates between continuous policy updates and discrete prompt evolution, using the GEPA algorithm to discover effective prompts for challenging samples. By distilling these prompts into the model's parameters, P^2O improves out-of-distribution generalization and achieves up to a 9.5% performance increase over existing methods. AI

A new research paper explores the effectiveness of automated prompt optimization compared to expert-crafted prompts for large language models. The study systematically compared hand-crafted prompts, base DSPy signatures, and GEPA-optimized DSPy signatures across translation, terminology insertion, and language quality assessment tasks. Results indicated that automated and manual prompts often yield similar quality, with performance varying by task and model configuration. AI

Apple researchers have developed a "Reinforced Agent" that proactively verifies tool calls before execution, aiming to prevent errors rather than correcting them post-hoc. This approach demonstrated significant improvements on benchmarks like BFCL irrelevance and τ²-Bench, with reasoning-model reviewers achieving a 3:1 helpful-to-harmful ratio. The system also saw a modest gain with the GEPA prompt optimization without requiring model retraining. AI

Researchers have developed GEPA, a new method for optimizing prompts in complex AI systems. GEPA analyzes failed execution paths and automatically refines the prompts of the specific modules responsible for the errors. In tests across six tasks, GEPA outperformed the GRPO method by an average of 6%, achieving this with significantly fewer rollouts. AI