Search arXiv⌕ Search
Home/Archive/arXiv

arXiv · 2609.37624

Correct, Don't Delete: Mitigating Emergent Misalignment with Corrective Supervision

Jacob Epifano
View original source ↗Read source document ↗Export metadata ↓

Abstract

Fine-tuning a language model on a narrow set of harmful demonstrations, such as bad medical advice, can make it broadly misaligned on unrelated questions, a phenomenon known as emergent misalignment (EM). The usual defense is to find the offending rows and delete them, but a row locator failed our held-out test and deleting rows helps less than expected. We ask a different question: given a fixed set of poisoned rows, is it better to correct them than to remove them? We fine-tune Qwen2.5-14B-Instruct on a mixture of bad medical advice and benign chat data, select a quarter of the poison rows in advance, and either delete them or replace each with a corrected answer to the same prompt, keeping everything else the same. Replacing the rows cuts the EM rate by about a third and improves answers on held-out medical questions, while deleting the same rows has little measurable effect. The advantage is larger when half the poison rows are corrected, and it holds on a second base model and a second misaligned model organism. The content of the replacement appears to matter: paraphrasing the rows while keeping their bad advice shows no clear benefit, and the correct answers distributed with the dataset appear to do about as well as our rewriter's. Realigning an already-poisoned model with further fine-tuning is known to work, but which data does the work has not been compared directly. We find that a short round of training on corrections beats the same amount of training on generic chat data, that corrections on other medical prompts do roughly as well as corrections of the poisoned prompts themselves, and that instructing the correction writer to model a careful, harm-avoiding assistant adds no measurable benefit over plain corrections. In the settings we tested, correcting harmful training data reduces EM more than deleting it.

Explore related subjects

cs.CLcs.AIcs.LG

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Jacob Epifano. 2026-09-29. Correct, Don't Delete: Mitigating Emergent Misalignment with Corrective Supervision. https://arxiv.org/abs/2609.37624

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Are We Really Making Much Progress in Text Classification? A Comparative Review

We survey the literature on single-label, multi-label, and hierarchical text classification and provide a quantitative comparison of methods categorized into bag-of-words, sequence-based, and graph- or hierarchy-based approaches. Despite a recent surge in graph-based methods, they do not provide an improvement over fine-tuned transformer models on most evaluated datasets. Decoder-only generative language models show promise in few-shot in-context learning, but appear to lag behind fine-tuned language models when sufficient training data is available. The amount of training data needed for a fine-tuned language model to exceed the performance of a generative model is task-dependent. We further highlight the variance in reported numbers across the literature when applying the same model to the same dataset, which can be traced to the use of different hyperparameter values, such as the fine-tuning learning rate. For practitioners, we recommend using a fine-tuned language model when sufficient training data is available. Otherwise, a frozen generative model, enhanced by few-shot in-context learning or reasoning, is preferable. The source code and further information are available at: https://github.com/ascherp/text-classification-survey

cs.CL↗
arXiv2026

Dynamic Optimizations of LLM Ensembles with Two-Stage Reinforcement Learning Agents

The advancement of LLMs and their accessibility have triggered renewed interest in multi-agent reinforcement learning as robust and adaptive frameworks for dynamically changing environments. This paper introduces \texttt{RL-Focal}, a two-stage RL agent framework that routes and ensembles LLMs. \textit{First}, we develop the Decider RL-agent, which learns to dynamically select an ensemble of small size ($m_i$) among $N$ LLMs ($m_i \ll N$) for incoming queries from a user-defined downstream task $i$, by maximizing both error-diversity and reasoning-performance of the selected ensemble through iterative updates of task-adaptive rewards and policy. \textit{Second}, to enable effective fusion of dynamically selected LLMs, we develop the stage-2 Fusion RL-agent, which learns to resolve reasoning conflicts from different LLMs and dynamically adapt to different ensemble teams composed by the Decider Agent for different downstream tasks. {\em Third}, we introduce the focal diversity metric to better model the error correlations among multiple LLMs further improving the generalization performance of the Decider Agent, which actively prunes the ensemble combinations. By focal diversity, we enhance performance across tasks by effectively promoting reward-aware and policy-adaptive ensemble selection and inference fusion. Extensive evaluations on five benchmarks show that RL-Focal achieves the performance improvement of 8.48\% with an ensemble of small size compared to the best individual LLM in a pool and offers stronger robustness. Code is available \href{https://github.com/git-disl/RL-Focal}{here}.

cs.CL↗
arXiv2026

TagPR: Tag-Guided Process Supervision for Personalization Reasoning in Large Language Models

Recent advancements have endowed Large Language Models with impressive general reasoning capabilities. However, these reasoning models often perform worse than non-reasoning models on personalization tasks. While some methods use outcome-based RL to improve personalization reasoning, they fail to supervise the reasoning process. As a result, models may reach correct answers through flawed reasoning chains, limiting further improvement. To address this, we propose TagPR, a novel framework that adds semantic tags to the reasoning process for step-by-step guidance. TagPR first automatically generates a structured, tagged dataset for Supervised Fine-Tuning. It then employs a multi-stage RL process guided by a composite reward signal, which integrates tag-based process supervision with a novel Personalization Reward Model with User Embeddings to achieve fine-grained alignment with user-specific logic. Extensive experiments on public LaMP, LongLaMP, PGraphRAG, and a self-constructed dataset demonstrate that our approach achieves state-of-the-art results, delivering an average improvement of 32.65% over the base model across all LaMP benchmark tasks. Our work demonstrates that tag-guided process supervision is an effective approach for personalization reasoning.

cs.CL↗