Wrong Before Right: Late Rescue and Interface Failure in Aligned Language Models

2026-07-05

AI Summary

This work reveals a “wrong-then-correct” phenomenon in aligned language models, wherein intermediate layers transiently favor incorrect answers—a behavior termed “causal probing”—before later layers rectify the output. Through minimal pairs, layer-wise differential analysis, and Patchscope-style activation transplantation, the study provides the first evidence that this mechanism is contingent on model scale (≥3B parameters) and alignment methodology, and substantially undermines compression robustness. To mitigate this issue, the authors propose a LoRA-based training strategy that penalizes erroneous intermediate representations. This approach reduces causal probing by 67–70% while preserving task accuracy, thereby significantly enhancing the stability of model compression.

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.

Problem

• aligned language models
• wrong-dip
• late rescue
• compression robustness
• internal representation

Innovation

• wrong-dip
• late rescue
• activation transplantation
• compression robustness
• alignment dynamics