- 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.