When an AI answers a complex question, it often needs to piece together facts from multiple sources. Current systems retrieve those facts (like passages or database entries) but leave the logic of how to connect them hidden inside the model's own processing. This means every new question that follows the same logical pattern must be solved from scratch.
PatternBloom changes what gets retrieved. Instead of facts alone, it retrieves a 'reasoning skeleton', a blueprint for how to connect information, like 'find the director, then find their birthplace'. These skeletons are learned by an AI that explores different reasoning paths and is rewarded for finding efficient ones.
The result is a 7-billion-parameter model that beats 27 other systems by 15.0 points on average. The improvement grows on questions requiring deeper reasoning, because the system reuses learned thinking patterns instead of rebuilding them each time.