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

• URL: http://arxiv.org/abs/2607.04640v1
• Date: Mon, 06 Jul 2026 03:51:24 GMT
• Title: Wrong Before Right: Late Rescue and Interface Failure in Aligned Language Models
• Abstract summary: We study how correctness is assembled inside aligned language models.<n>Using layer-wise difference-in-differences (DiD) trajectories over polarity-controlled minimal pairs, we identify the wrong-dip.<n>We verify this causally with patchscope-style activation transplantation across 17 models, three families, and 64x scale (0.5B-32B)
• Score: 2.1291393112295935
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• 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.

Related papers

• Steering Under Compression: Dose-Response, Capability Cost, and Failure Asymmetry in Quantized LLMs [0.0]
  Inference-time steering enables behavioral control of large language models without parameter modification.<n>Post-training quantization reduces memory and compute costs for deployment.<n>We systematically study activation steering under weight-only quantization.
  arXiv  Detail & Related papers  (2026-09-06T08:41:42Z)
• On the Design of Qwen3.8-Next Architecture: Evaluation, Efficiency, and Training Stability [57.03393882933515]
  We describe the architecture and ablations of Qwen3.8-Flash-Next, a sparse mixture-of-experts model with 125B parameters, 6B activated per token, and 51B parameters of n-gram embedding tables held off the accelerator.<n>On fourteen pre-training benchmarks the model leads the 397B-A17B predecessor on eight and trails it on the rest by at most 2.6 points, at 1/3 the activated parameters, 1/3 the training tokens, and roughly 1/9 the training FLOPs.
  arXiv  Detail & Related papers  (2026-08-31T06:35:07Z)
• Sensitivity, Causality, and Repair Dissociate: A Layer-Wise Analysis of Perturbation Robustness and Its Scaling [0.0]
  We show that when a language model fails on surface-perturbed input, "which layer is responsible" has three natural operationalizations.<n>We identify two propagation regimes - spike-and-suppress and late-accumulation.<n>A fixed-harness layer sweep across four models (3.8-8B) confirms the core prediction on chain-of-thought GSM8K.
  arXiv  Detail & Related papers  (2026-08-04T15:48:55Z)
• A Decomposable Probe for Few-Step Diffusion Models: Prompt, Latent, and Score Selectivity across Backbone Families and Distillation Paradigms [0.0]
  Few-step distilled diffusion students cut text-to-image inference from 50 to 1-8 network evaluations.<n>We replace the scalar with a decomposable probe that injects controlled perturbations along three layers.<n>We report a bootstrap-median Bures W22 selectivity ratio on Inception features.
  arXiv  Detail & Related papers  (2026-07-03T12:18:24Z)
• Decodable Is Not Grounded: A Vision-Ablation Arbiter for VLM Spatial Reasoning [48.438568700670835]
  The standard way to read latent knowledge out of a model can systematically overstate what a vision-language model grounds in the image.<n>We show this on spatial reasoning, where the error is invisible to both probing and steering yet exposed by a one-line causal control.
  arXiv  Detail & Related papers  (2026-06-30T07:33:18Z)
• Layer-Isolated Evaluation: Gating the Deterministic Scaffold of a Production LLM Agent with a No-LLM, Regression-Locked Test Harness [45.148328075418156]
  A deployed ordering agent is decomposed into a fixed taxonomy of layers.<n>The pure suite runs in CI on every change against a locked per-slice baseline.<n>We validate by controlled regression injection, degrading one layer at a time across seven non-safety layers.
  arXiv  Detail & Related papers  (2026-06-10T05:55:13Z)
• Right Makes Might: Aligning Verified Hidden States Empowers RL Reasoning [55.264863369127774]
  Current methods reduce each correct rollout to a single reward bit, ignoring the geometric structure shared among their hidden states.<n>We propose Hidden-Align, an auxiliary loss function that aligns the last-layer hidden states of correct rollouts at the anchor token during RL training, with zero overhead in both training and inference.<n>On eight mathematical reasoning benchmarks, Hidden-Align improves average pass@1 over the DAPO baseline by 3.8, 6.2, and 5.4 percentage points on Qwen3-1.7B, 4B, and 14B respectively, with consistent pass@k gains across all three scales.
  arXiv  Detail & Related papers  (2026-06-02T06:51:15Z)
• Layer Collapse in Diffusion Language Models [54.880703002010144]
  Diffusion language models (DLMs) have emerged as competitive alternatives to autoregressive (AR) language models.<n>We show that layer collapse in DLMs is not driven by undertraining but by overtraining.<n>Our findings have strong practical implications.
  arXiv  Detail & Related papers  (2026-05-07T14:39:40Z)
• Architectural Observability Collapse in Transformers [0.0]
  Training preserved an internal decision-quality signal that output confidence does not expose.<n>Confidence controls absorb on average 60.3% of raw probe signal across 14 models in 6 families.<n>Training preserved observability where Llama 3.1 8B collapses at identical 32-layer, 32-head, 4096-hidden shape.
  arXiv  Detail & Related papers  (2026-04-27T02:39:02Z)
• Disposition Distillation at Small Scale: A Three-Arc Negative Result [0.0]
  Internal draft reported +33.9-point MCAS and +15.3-point HumanEval gains on a Qwen3-0.6B student; a second-pass sanity check falsified both numbers before publication.<n>We find no operator that moves judge-measured disposition without damaging content or collapsing into stylistic mimicry.<n>We contribute a three-arc negative result with mechanism, a two-failure-mode taxonomy for linear h_last probes, and an honest falsification pipeline that converts the class of false positives we ourselves produced into publishable negatives.
  arXiv  Detail & Related papers  (2026-04-13T17:40:31Z)
• Fragile Reasoning: A Mechanistic Analysis of LLM Sensitivity to Meaning-Preserving Perturbations [2.7946918847372277]
  Large language models demonstrate strong performance on mathematical reasoning benchmarks, yet remain surprisingly fragile to meaning-preserving surface perturbations.<n>We systematically evaluate three open-weight LLMs, Mistral-7B, Llama-3-8B, and Qwen2.5-7B, on 677 GSM8K problems paired with semantically equivalent variants.<n>All three models exhibit substantial answer-flip rates (28.8%-45.1%), with number paraphrasing consistently more disruptive than name swaps.
  arXiv  Detail & Related papers  (2026-04-02T05:30:20Z)
• Information Fidelity in Tool-Using LLM Agents: A Martingale Analysis of the Model Context Protocol [69.11739400975445]
  We introduce the first theoretical framework for analyzing error accumulation in Model Context Protocol (MCP) agents.<n>We show that cumulative distortion exhibits linear growth and high-probability deviations bounded by $O(sqrtT)$.<n>Key findings include: semantic weighting reduces distortion by 80%, and periodic re-grounding approximately every 9 steps suffices for error control.
  arXiv  Detail & Related papers  (2026-02-10T21:08:53Z)
• Token-Level Inference-Time Alignment for Vision-Language Models [58.41370989069588]
  Vision-Language Models (VLMs) have become essential backbones of modern multimodal intelligence.<n>We present TITA, a lightweight framework that freezes the base VLM and instead trains a reward model to approximate its distribution.<n>During inference, implicit preference signals are extracted as log-probability ratios between the reward model and the target VLM, yielding dense autoregressive feedback.
  arXiv  Detail & Related papers  (2025-10-20T09:58:03Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.