Show HN: Gave Claude Code ADHD.. Now it thinks 3x better (adhdstack.github.io)

A new method called ADHD (Attention Divergence and Hierarchical Determination) has been developed to tackle the issue of "premature convergence" in language models, which limits their ability to generate diverse and innovative ideas during problem-solving tasks. Unlike existing methods that focus on optimizing correctness within a closed answer space, ADHD encourages the generation of a wide range of non-obvious viable options. By promoting structural divergence during the ideation phase and using a separate evaluation phase, it prevents early tokens from anchoring later responses. This two-phase mechanism allows multiple language models to think independently and creatively, resulting in ideas that a single, conventional model might miss. The significance of ADHD lies in its potential to enhance ideation processes, particularly in engineering and design domains. In evaluations across six complex engineering problems, ADHD significantly outperformed a strong baseline model on dimensions like breadth, novelty, and trap detection, indicating its effectiveness in surfacing diverse and innovative solutions. For instance, it flagged misleading ideas and generated structurally distinct approaches that the baseline overlooked, showcasing its utility in scenarios where conventional reasoning may fall short. By investing a bit more computational effort for a richer output, ADHD offers a promising tool for decision-makers aiming to navigate complex problem spaces creatively.