Machine Dialectology (MDia)
A research framework where language models create persistent symbolic dialects (LSFs) for exchanging reasoning between speaker and listener models, then routes among them under accuracy, token, and reliability constraints.
Machine Dialectology (MDia) is a research framework for studying how large language models invent and use compact reasoning dialects. Speaker models create persistent Language Symbolism Frameworks (LSFs): dialect cards with symbols, grammar, operators, validity rules, and answer contracts. Listener models execute those dialects.
MDia routes messages between speakers and listeners under explicit accuracy, token, parseability, and reliability constraints, supporting single, aggregate, compose, and guarded abstain-or-fallback inference plans. CLSR, a homogeneous-agent special case, is configured separately from the main MDia implementation.
The project is written in Python and released under the MIT license, accompanying an ICML 2026 paper.