Rune
A mechanistic-interpretability research project testing whether a language model's internal activations can reveal the operation and operands of an arithmetic prompt well enough to route it to a calculator without parsing the prompt text.
Rune studies whether a model's internal activations expose the operation and operands behind an arithmetic prompt, and whether those internal readouts can route computation to a calculator without falling back to parsing the prompt text at inference time. On published Llama-3.1-8B slices, activation-derived operation and operand readouts can supply calculator arguments for several arithmetic tasks; the prompt is treated as opaque, with the route getting token IDs and captured activations rather than a regex over the text.
The project explicitly does not claim the model's native arithmetic has been repaired, and is not a residual JIT replacement that resumes generation from a patched hidden state. It builds on and cites prior work on embodied mathematical cognition, trigonometric representations of addition inside language models, and causal mediation analysis of arithmetic circuits.
The repository includes an article abstract, an ACM Queue-style pitch abstract, an interactive HTML article with an exported PDF, a documented claim boundary, and commands to rebuild the main checks. Written in Python.