Fusion Check: a zero-data law that predicts whether merging two fine-tunes preserves behavior

Описание и применение

Fusion Check – инструмент, позволяющий за секунду определить, совместно ли два дообученных LLM можно объединить, не ухудшая их производительность.

Задача

Blind weight averaging (soup) can dilute model performance when fine‑tuned LLMs are trained on different tasks.

Результат

Fusion Check accurately predicts fusion outcomes with 0.81 LOOCV correlation and 0.072 MAE across 13 task pairs, enabling safe model merging and avoiding performance dilution.

Рабочий процесс

• 1. Install Fusion Check via pip.
• 2. Run fusion-check verificar <base> <m1> <m2> --acc1 <acc1> --acc2 <acc2> to compute cosine of weight deltas and effective dimensionality.
• 3. Read the decision output (MERGE/NO_MERGE/PRIORITIZED/REFUSE/VERIFY) and the optimal mixing coefficient α*.
• 4. If MERGE, apply the recommended α* to fuse the models; otherwise, follow the suggested action.

Доказательства

Fusion Check — a law that tells you, before merging, whether two fine‑tuned LLMs are compatible.

Источник

Proven three independent ways: Analytic derivation, Formal verification with z3 (28 theorems), Empirical validation (147 combinations, max deviation 1.13%, zero violations).

Источник

Real result: Two weak, diverse estimators of the same task (0.762/0.756) → merged soup 0.822 (+6–7 points).

Источник