A study of 17 models across three families reveals that aligned language models often transiently commit to incorrect answers in mid-layers (25-90% depth) before being corrected by late-layer mechanisms. This "wrong-dip" is verified causally through activation transplantation and varies significantly based on alignment recipes and model scale.
- The causal wrong-dip amplification is recipe-specific, peaking at 32B in Qwen2.5 and reversing in Llama-3-8B.
- High-dip items are 3-7x more likely to fail under late-layer low-rank compression or pruning, while remaining robust to quantization.
- A LoRA fine-tune with a mid-layer wrong-margin penalty cuts the causal dip by 67-70% without sacrificing accuracy.
- Output-only SFT worsens the causal dip by up to 2.8x even at perfect surface accuracy.
- The phenomenon persists in natural-language I/O, separating fragility into a dip-auditable late-rescue layer and a dip-blind interface layer.
The authors conclude that output-level correctness can hide a late-rescue process that governs compression risk, post-training quality, and evaluation distortion.