Startup targets LLM groupthink with diversity-focused training

MIT Technology Review identified a pattern in large language models described as a “groupthink groove,” where systems repeatedly produce safe, statistically common answers at the expense of originality. In one benchmark across 100 trials, over 60% of popular LLMs chose 7 when asked for a random number between 1 and 10, and business-idea prompts often converged on familiar categories such as AI-powered analytics or subscription boxes.

The startup highlighted in the findings says current training methods, including RLHF, supervised fine-tuning, and reinforcement learning, push models toward responses most human raters consider acceptable. Its proposed fix uses “contrarian training signals” that reward less common outputs when they remain accurate or logically sound.

The method includes broader human preference data, minority viewpoints grounded in truth, a novelty term in the loss function, and adversarial sampling to retain unexpected but valid responses during fine-tuning. Early results show a 40% increase in output diversity while maintaining factual accuracy on benchmarks such as MMLU and HellaSwag.

Developers and businesses using LLMs for writing, brainstorming, code generation, product design, or strategic planning may need to audit models for repetitive outputs. Diversity-aware fine-tuning could become a way to distinguish AI assistants as more products adopt similar generative features.