Users appreciate the study showing how SFT and RL enable compositional generalization in reasoning models, thanking authors and engaging enthusiastically with related questions and suggestions.
Based on 2 visible X reactions from 5 accounts; directional sample.
Ask a question below.
Published answers will appear here.
@_Suresh2 @IFM_MBZUAI Love this question, thank you! By “reasoning creativity,” we mean compositional generalization: solving a problem by building a reasoning path that no training example supplied. The exciting part is that the model must discover reusable pieces and learn how to connect them.
@henryhndev @rsalakhu Thanks for the great idea and suggestion! We kept prompts fixed in this study, but exploring how prompt structure and tool-use workflows shape compositional flexibility is something we’d love to try next -- happy to discuss more:)
@rsalakhu Why can't modules be discovered and recombined by supervised training?
The approach trains models to decompose novel tasks into reusable sub-skills
Users appreciate the study showing how SFT and RL enable compositional generalization in reasoning models, thanking authors and engaging enthusiastically with related questions and suggestions.
Based on 2 visible X reactions from 5 accounts; directional sample.
Ask a question below.
Published answers will appear here.