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

arXiv: https://arxiv.org/abs/2607.04640 · HTML: https://mlsanity.com/paper/2607.04640/

## Key facts

- Cited by 1 papers in the map (data through 14 Oct 2026).
- Placed in LLM preference optimization methods, one of 10 directions in LLM alignment / RLHF / preference optimisation.
- Builds on 18 of its 20 references that are in the map.
- First posted on arXiv 6 Jul 2026.

## Placed in

- LLM alignment / RLHF / preference optimisation / [LLM preference optimization methods](https://mlsanity.com/alignment/llm-preference-optimization-methods/)

## Builds on (most cited first, 18 of 18)

- [The Llama 3 Herd of Models](https://mlsanity.com/paper/2407.21783/) (2024)
- [Qwen2.5 Technical Report](https://mlsanity.com/paper/2412.15115/) (2024)
- [Locating and Editing Factual Associations in GPT](https://mlsanity.com/paper/2202.05262/) (2022)
- [Emergent Abilities of Large Language Models](https://mlsanity.com/paper/2206.07682/) (2022)
- [GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformers](https://mlsanity.com/paper/2210.17323/) (2022)
- [Representation Engineering: A Top-Down Approach to AI Transparency](https://mlsanity.com/paper/2310.01405/) (2023)
- [LLM.int8(): 8-bit Matrix Multiplication for Transformers at Scale](https://mlsanity.com/paper/2208.07339/) (2022)
- [Eliciting Latent Predictions from Transformers with the Tuned Lens](https://mlsanity.com/paper/2303.08112/) (2023)
- [ShortGPT: Layers in Large Language Models are More Redundant Than You Expect](https://mlsanity.com/paper/2403.03853/) (2024)
- [Towards Best Practices of Activation Patching in Language Models: Metrics and Methods](https://mlsanity.com/paper/2309.16042/) (2023)
- [Patchscopes: A Unifying Framework for Inspecting Hidden Representations of Language Models](https://mlsanity.com/paper/2401.06102/) (2024)
- [The Unreasonable Ineffectiveness of the Deeper Layers](https://mlsanity.com/paper/2403.17887/) (2024)
- [The Hydra Effect: Emergent Self-repair in Language Model Computations](https://mlsanity.com/paper/2307.15771/) (2023)
- [The Truth is in There: Improving Reasoning in Language Models with Layer-Selective Rank Reduction](https://mlsanity.com/paper/2312.13558/) (2023)
- [Overthinking the Truth: Understanding how Language Models Process False Demonstrations](https://mlsanity.com/paper/2307.09476/) (2023)
- [Compressing LLMs: The Truth is Rarely Pure and Never Simple](https://mlsanity.com/paper/2310.01382/) (2023)
- [Decoding Compressed Trust: Scrutinizing the Trustworthiness of Efficient LLMs Under Compression](https://mlsanity.com/paper/2403.15447/) (2024)
- [An Empirical Study of Qwen3 Quantization](https://mlsanity.com/paper/2505.02214/) (2025)

## Built on by (most cited first, 1 of 1)

- [Reachability Is Not Realization: Tracing the Sources of LLM Benchmark Gains](https://mlsanity.com/paper/2608.03219/) (2026)

## Abstract (arXiv)

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.