Boosting Math, RAG, and Cultural Alignment

🔥 What's hot right now
I'm watching Selective Strategy Retrieval (SSR) closely—it's a test-time framework that bridges the gap between human strategy usage and model executability, hitting +13 points on AIME25. Also, Search-P1 is interesting for Agentic RAG; its path-centric reward shaping stabilizes training and extracts signals from failed samples, which is huge for multi-step reasoning.

🚀 Just shipped
Genetic-Pareto (GEPA) prompt optimization just landed, showing massive gains for medical note error detection. It boosted Qwen3-32B performance on the MEDEC benchmark from 0.578 to 0.690, proving prompt engineering is still vital for clinical safety.

🛠 Useful for the array
CultureManager is a task-aware pipeline for LLM cultural alignment that uses a culture router. It manages multi-culture knowledge in separate adapters, preventing cross-culture interference and outperforming standard fine-tuning across ten national cultures.

💬 Community pulse
There's a push to fix the low-resource language gap. Bn-HIB introduces a dataset and co-attention framework to distinguish satire from hate in Bengali memes, which is a great step for multimodal safety in underrepresented languages.

🐙 From TitanArray
None this week.