Researchers propose Amplitude Gating (AG), a non-destructive inference-time method that modulates feed-forward network activation magnitudes to improve structured outputs in large language models without retraining weights. This approach addresses the limitations of previous direction-changing repairs by preserving pretrained weight directions while correcting small format or argument errors.
- AG defines a fine-grained intervention system spanning P1/P2/P3 and branch-specific sites, evaluated via a protocol separating combination-oracle headroom from learned gates.
- On Qwen3.5-9B, category-level learned gates improve tool/structured/agentic performance from 38.66% to 42.92%, with Hermes function-call tasks gaining approximately +7.6 points.
- On Qwen3-8B, Hermes JSON mode improves by +11.36 points.
- Qwen2.5-7B retains oracle headroom but current learned gates fail to capture it, indicating deployment requires model-specific routing.
- Comparisons show neither entropy AG nor Newton-Schulz-windowed AG is uniformly dominant across models.
The results identify tool-structured inference as the most credible target for safe FFN-level optimization, though broader cross-model evaluation remains necessary.