Wrong Before Right: Late Rescue and Interface Failure in Aligned Language Models
2026-07-06 • Computation and Language
Computation and Language
AI summaryⓘ
The authors investigate how aligned language models internally process correct answers, discovering a 'wrong-dip' where middle layers briefly favor incorrect answers before later layers fix them. They show this effect varies with the model size and architecture, affects how models respond to compression methods, and can be reduced by targeted fine-tuning. This internal correction process influences model robustness and how well final answers reflect true understanding. Their work highlights that correct outputs can hide complex internal struggles within the model layers.
aligned language modelslayer-wise difference-in-differenceswrong-dipactivation transplantationmodel compressionLoRA fine-tuningstructured pruningquantizationlate-layer correction
Authors
Jiaqi Deng
Abstract
We study how correctness is assembled inside aligned language models, not only whether the final answer is right. Using layer-wise difference-in-differences (DiD) trajectories over polarity-controlled minimal pairs, we identify the wrong-dip: in mid layers (25-90% depth), internal preference transiently commits to the incorrect answer and is rescued only by late-layer correction. We verify this causally with patchscope-style activation transplantation across 17 models, three families, and 64x scale (0.5B-32B). Four findings follow. (1) Alignment amplification of the causal wrong-dip is recipe-specific and emergent: it emerges at 3B in Qwen2.5, remains high, and peaks at 32B (paired t up to 9.7), reverses in Llama-3-8B (t=-2.31), and sits between for Mistral-7B. (2) The dip predicts real compression failures: high-dip items are 3-7x more likely to flip under late-layer low-rank compression, block dropping, or structured pruning, while quantization flips are dip-blind, a double dissociation confirmed by late-layer ablation. (3) The dip is trainable: a LoRA fine-tune with a mid-layer wrong-margin penalty matches output-only SFT accuracy while cutting the causal dip by 67-70% and improving compression robustness; output-only SFT worsens the causal dip by up to 2.8x at perfect surface accuracy. (4) With controlled readouts, the phenomenon survives natural-language I/O: dip stratification of structural-damage failures is significant on naturalistic vignettes, and free-form fragility separates into a dip-auditable late-rescue layer and a dip-blind interface layer. Together, output-level correctness can hide a late-rescue process that governs compression risk, post-training quality, and evaluation distortion.