Lowering barriers to explainable AI: Control technique for LLMs reduces hardware and energy requirements by over 90%

https://techxplore.com/news/2025-12-lowering-barriers-ai-technique-llms.html

"LangVAE and LangSpace software make LLMs more explainable/ controllable while drastically reducing resource costs... Geometric Mapping: Instead of modifying LLM itself, creates compressed representations of language... treats internal patterns as geometric shapes/ points in space, allowing mathematical model behavior adjustment... lowers barrier to entry to AI development, more scientists can contribute to AI safety/ explainability... makes AI behavior more predictable for mission-critical fields like healthcare"