Legible Transformers Offer Clearer AI Understanding and Editing

Understanding and editing the internal mechanisms of large transformer models remains a significant challenge due to their opaque "black box" nature. Current research aims to make these models more interpretable by constructing them from "legible" operators, which function like fuzzy set operations rather than dense, uninterpretable activations. A new training methodology has been developed to enhance this legibility. Previous attempts to sharpen these operators often led to them collapsing into inert constants. The key innovation is a per-channel variance floor, which acts as a target legibility metric during training, preventing collapse and recovering both interpretability and model quality. Additionally, a learned per-unit fraction dynamically determines the optimal mix of legible and conventional operations. The resulting "most legible transformer" routes a significant portion of its computation through crisp, contextual detectors, making its internal logic more transparent. This increased legibility allows for far more localized and precise edits to the model's behavior, targeting explicit logical conjunctions that single neurons cannot express. The approach maintains quality parity with conventional baselines while offering a "legibility dial" to trade circuit reuse for unit independence, turning complex concepts into surgically editable components.