Learning from Mistakes: Can LLM Self-Recover after Misalignment?

• URL: http://arxiv.org/abs/2606.00003v1
• Date: Wed, 25 Mar 2026 15:36:15 GMT
• Title: Learning from Mistakes: Can LLM Self-Recover after Misalignment?
• Abstract summary: This paper introduces a new perspective on advancing alignment techniques.<n>Rather than developing stronger alignment techniques, we investigate the model's intrinsic ability to recover its alignment after corruption.<n>We propose a methodology for modeling the safety trajectories of user-assistant interactions.
• Score: 1.8374319565577155
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Responsible AI initiatives place great emphasis on the safety of Large Language Model (LLM)-based systems. In particular, it has become standard practice to subject these models to an alignment procedure aimed at preventing harmful outputs. However, once aligned, a model is not guaranteed to maintain this alignment throughout its lifecycle. Moreover, the likelihood of misalignment increases as malicious actors may deliberately employ jailbreaking techniques to compromise LLM safety. To counter this, much research has focused on improving alignment methods and post-processing filters. In this paper, we introduce a new perspective on advancing LLM alignment: rather than developing stronger alignment techniques, we investigate the model's intrinsic ability to recover its alignment after corruption. We propose a methodology for modeling the safety trajectories of user-assistant interactions and for detecting recovery trends within them. We apply this approach to a jailbreaking scenario, presenting a preliminary recovery analysis based on a dataset of adversarial multi-turn dialogues and examining the influence of the content moderation model chosen for safety evaluation. Project page with an interactive data visualizer is available at https://lab-rococo-sapienza.github.io/LearningfromMistakes.

Related papers

• Rendering Data Unlearnable by Exploiting LLM Alignment Mechanisms [3.648393062009244]
  Large language models (LLMs) are increasingly trained on massive, heterogeneous text corpora.<n>This raises serious concerns about the unauthorised use of proprietary or personal data during model training.<n>We propose Disclaimer Injection, a novel data-level defence that renders text unlearnable to LLMs.
  arXiv  Detail & Related papers  (2026-01-06T20:34:15Z)
• Do We Really Need Curated Malicious Data for Safety Alignment in Multi-modal Large Language Models? [83.53005932513155]
  Multi-modal large language models (MLLMs) have made significant progress, yet their safety alignment remains limited.<n>We propose finetuning MLLMs on a small set of benign instruct-following data with responses replaced by simple, clear rejection sentences.
  arXiv  Detail & Related papers  (2025-04-14T09:03:51Z)
• Improving LLM Safety Alignment with Dual-Objective Optimization [81.98466438000086]
  Existing training-time safety alignment techniques for large language models (LLMs) remain vulnerable to jailbreak attacks.<n>We propose an improved safety alignment that disentangles DPO objectives into two components: (1) robust refusal training, which encourages refusal even when partial unsafe generations are produced, and (2) targeted unlearning of harmful knowledge.
  arXiv  Detail & Related papers  (2025-03-05T18:01:05Z)
• Fooling LLM graders into giving better grades through neural activity guided adversarial prompting [26.164839501935973]
  We propose a systematic method to reveal such biases in AI evaluation systems.<n>Our approach first identifies hidden neural activity patterns that predict distorted decision outcomes.<n>We demonstrate that this combination can effectively fool large language model graders into assigning much higher grades than humans would.
  arXiv  Detail & Related papers  (2024-12-17T19:08:22Z)
• Course-Correction: Safety Alignment Using Synthetic Preferences [17.897817682322053]
  We introduce the textscC$2$-Eval benchmark for quantitative assessment and analyze 10 popular language models. Using an automated pipeline, we create textscC$2$-Syn, a synthetic dataset with 750K pairwise preferences. Experiments on 2 LLMs, textscLlama2-Chat 7B and textscQwen2 7B, show that our method effectively enhances course-correction skills without affecting general performance.
  arXiv  Detail & Related papers  (2024-07-23T16:54:28Z)
• Model Merging and Safety Alignment: One Bad Model Spoils the Bunch [70.614652904151]
  Merging Large Language Models (LLMs) is a cost-effective technique for combining multiple expert LLMs into a single versatile model. Current approaches often overlook the importance of safety alignment during merging, leading to highly misaligned models. We evaluate several popular model merging techniques, demonstrating that existing methods do not only transfer domain expertise but also propagate misalignment.
  arXiv  Detail & Related papers  (2024-06-20T17:59:58Z)
• DeAL: Decoding-time Alignment for Large Language Models [58.368979253590794]
  Large Language Models (LLMs) are nowadays expected to generate content aligned with human preferences.<n>We propose DeAL, a framework that allows the user to customize reward functions and enables Decoding-time alignment of LLMs.
  arXiv  Detail & Related papers  (2024-02-05T06:12:29Z)
• InferAligner: Inference-Time Alignment for Harmlessness through Cross-Model Guidance [56.184255657175335]
  We develop textbfInferAligner, a novel inference-time alignment method that utilizes cross-model guidance for harmlessness alignment. Experimental results show that our method can be very effectively applied to domain-specific models in finance, medicine, and mathematics. It significantly diminishes the Attack Success Rate (ASR) of both harmful instructions and jailbreak attacks, while maintaining almost unchanged performance in downstream tasks.
  arXiv  Detail & Related papers  (2024-01-20T10:41:03Z)
• Fake Alignment: Are LLMs Really Aligned Well? [91.26543768665778]
  This study investigates the substantial discrepancy in performance between multiple-choice questions and open-ended questions. Inspired by research on jailbreak attack patterns, we argue this is caused by mismatched generalization.
  arXiv  Detail & Related papers  (2023-11-10T08:01:23Z)
• Open Sesame! Universal Black Box Jailbreaking of Large Language Models [0.0]
  Large language models (LLMs) are designed to provide helpful and safe responses. LLMs often rely on alignment techniques to align with user intent and social guidelines. We introduce a novel approach that employs a genetic algorithm (GA) to manipulate LLMs when model architecture and parameters are inaccessible.
  arXiv  Detail & Related papers  (2023-09-04T08:54:20Z)

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.