Booth 21-25 | AI Data Management Zone | Tokyo Big Sight
Most improvements to AI systems happen before deployment: better training data, better fine-tuning, better RLHF. Once the model is out in the world, you generally get the performance you trained
Here’s a finding that genuinely surprised me when I sat with it. Shaoyang Xu and Wenxuan Zhang at the Singapore University of Technology and Design (arXiv:2601.11227, January 2026) showed that the language an LLM uses for its internal reasoning measurably changes the diversity of its English outputs — even when
