Pinned Loading

1. Semantic-Gravity-RP Semantic-Gravity-RP Public

   Mechanistic investigation of why LLMs fail at negative constraints ("don't use word X"). Violation probability follows semantic pressure (p = σ(−2.40 + 2.27·P₀)). Two failure modes: priming (87.5%)…

   Python

2. Long-Number-Addition Long-Number-Addition Public

   Analysis of the accuracy of Qwen 2.5 7B model on performing calculations for numbers with 10-18 digits. We found that for large numbers, the model has a bias to produce an extra digit, even when no…

   Jupyter Notebook