How Transformers Solve Propositional Logic Problems: A Mechanistic Analysis

arXiv:2411.04105v1 Announce Kind: cross
Summary: Giant language fashions (LLMs) have proven wonderful efficiency on duties that require planning and reasoning. Motivated by this, we examine the inner mechanisms that underpin a community’s skill to carry out advanced logical reasoning. We first assemble an artificial propositional logic downside that serves as a concrete test-bed for community coaching and analysis. Crucially, this downside calls for nontrivial planning to resolve, however we will prepare a small transformer to attain good accuracy. Constructing on our set-up, we then pursue an understanding of exactly how a three-layer transformer, skilled from scratch, solves this downside. We’re capable of determine sure “planning” and “reasoning” circuits within the community that necessitate cooperation between the eye blocks to implement the specified logic. To develop our findings, we then research a bigger mannequin, Mistral 7B. Utilizing activation patching, we characterize inside elements which are essential in fixing our logic downside. General, our work systemically uncovers novel elements of small and enormous transformers, and continues the research of how they plan and motive.

Source link

#Transformers #Remedy #Propositional #Logic #Issues #Mechanistic #Evaluation