Amplitude-Only FFN Intervention for Tool-Structured LLM Inference: Method, Gated Evaluation Protocol, and Cross-Model Empirical Results
Amplitude Gating (AG) is an inference-time technique that modifies feed-forward network activations to improve tool-structured LLM performance while keeping pretrained weights fixed. It shows positive exploratory gains on specific Qwen models and function-call endpoints, suggesting a task-dependent intervention surface. Deployment requires careful selection of candidates and acceptance policies.