large-language-models/ model-optimization · gemma

Targeted weight edits curb repetitive loops in Gemma 4 LLMs

Researchers show that silencing a handful of neurons stops most enumeration failures in Gemma 4 models, though deeper knowledge gaps remain.

  • Gemma 4 instruction‑tuned models often get stuck when asked to list long series, repeating the same answer or collapsing to a single entry.
  • Experiments identified a small cluster of MLP neurons (or routed experts in the 26B‑A4B MoE) that trigger the loops. By flipping the sign of a single weight in the E2B variant, or applying a few more edits in larger models, the authors eliminated the repetition while keeping benchmark scores intact.
  • The fix matters because it proves a concrete pathology can be isolated to a few parameters, offering a cheaper alternative to full‑model retraining. However, the edits do not address “doom loops” that arise when the model circles around a missing fact under longer generation budgets; those stem from knowledge precision, not a removable circuit.
  • In short, neuron‑level surgery is a useful tool for polishing specific failure modes, but it won’t replace better data or architecture when the model simply lacks the needed information.
TR

The Revision

Written by an AI system from the public sources credited above. How we write →