Teaching Models to Understand (but not Generate) High-risk Data

• URL: http://arxiv.org/abs/2505.03052v2
• Date: Wed, 15 Oct 2025 17:30:34 GMT
• Title: Teaching Models to Understand (but not Generate) High-risk Data
• Abstract summary: We introduce Selective Loss to Understand but Not Generate (SLUNG)<n>SLUNG is a pre-training paradigm through which models learn to understand high-risk data without learning to generate it.<n>We show that SLUNG consistently improves models' understanding of high-risk data without increasing its generation.
• Score: 47.384144793060955
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Language model developers typically filter out high-risk content -- such as toxic or copyrighted text -- from their pre-training data to prevent models from generating similar outputs. However, removing such data altogether limits models' ability to recognize and appropriately respond to harmful or sensitive content. In this paper, we introduce Selective Loss to Understand but Not Generate (SLUNG), a pre-training paradigm through which models learn to understand high-risk data without learning to generate it. Instead of uniformly applying the next-token prediction loss, SLUNG selectively avoids incentivizing the generation of high-risk tokens while ensuring they remain within the model's context window. As the model learns to predict low-risk tokens that follow high-risk ones, it is forced to understand the high-risk content. Through our experiments, we show that SLUNG consistently improves models' understanding of high-risk data (e.g., ability to recognize toxic content) without increasing its generation (e.g., toxicity of model responses). Overall, our SLUNG paradigm enables models to benefit from high-risk text that would otherwise be filtered out.

Related papers

• From Hazard Functions to Language Space: Cox-Supervised Distillation of Survival Risk into a Large Language Model [0.7537475180985093]
  We investigate whether information about time-to-event risk can be transferred into a generative large language model.<n>We propose a text-based survival modelling pipeline and a Qwen-based large language model is fine-tuned to generate patient-specific survival risk.
  arXiv  Detail & Related papers  (2026-06-08T02:47:05Z)
• The Unseen Threat: Residual Knowledge in Machine Unlearning under Perturbed Samples [16.030881842099998]
  We show that slight perturbeds of forget samples may still be correctly recognized by the unlearned model.<n>We propose a fine-tuning strategy, named RURK, that penalizes the model's ability to re-recognize forget samples.
  arXiv  Detail & Related papers  (2026-01-29T22:10:13Z)
• Breaking the Gold Standard: Extracting Forgotten Data under Exact Unlearning in Large Language Models [26.5039481643457]
  We introduce a novel data extraction attack that compromises even exact unlearning.<n>We demonstrate our attack's effectiveness on a simulated medical diagnosis dataset.
  arXiv  Detail & Related papers  (2025-05-30T09:09:33Z)
• Risk-aware Direct Preference Optimization under Nested Risk Measure [23.336246526648374]
  Risk-aware Direct Preference Optimization (Ra-DPO) is a novel approach that incorporates risk-awareness by employing a class of nested risk measures.<n> Experimental results across three open-source datasets demonstrate the proposed method's superior performance in balancing alignment performance and model drift.
  arXiv  Detail & Related papers  (2025-05-26T08:01:37Z)
• Uncertainty-Aware Decoding with Minimum Bayes Risk [70.6645260214115]
  We show how Minimum Bayes Risk decoding, which selects model generations according to an expected risk, can be generalized into a principled uncertainty-aware decoding method.<n>We show that this modified expected risk is useful for both choosing outputs and deciding when to abstain from generation and can provide improvements without incurring overhead.
  arXiv  Detail & Related papers  (2025-03-07T10:55:12Z)
• On the Privacy Risk of In-context Learning [36.633860818454984]
  We show that deploying prompted models presents a significant privacy risk for the data used within the prompt. We also observe that the privacy risk of prompted models exceeds fine-tuned models at the same utility levels.
  arXiv  Detail & Related papers  (2024-11-15T17:11:42Z)
• DPP-Based Adversarial Prompt Searching for Lanugage Models [56.73828162194457]
  Auto-regressive Selective Replacement Ascent (ASRA) is a discrete optimization algorithm that selects prompts based on both quality and similarity with determinantal point process (DPP) Experimental results on six different pre-trained language models demonstrate the efficacy of ASRA for eliciting toxic content.
  arXiv  Detail & Related papers  (2024-03-01T05:28:06Z)
• C-RAG: Certified Generation Risks for Retrieval-Augmented Language Models [57.10361282229501]
  We propose C-RAG, the first framework to certify generation risks for RAG models. Specifically, we provide conformal risk analysis for RAG models and certify an upper confidence bound of generation risks. We prove that RAG achieves a lower conformal generation risk than that of a single LLM when the quality of the retrieval model and transformer is non-trivial.
  arXiv  Detail & Related papers  (2024-02-05T16:46:16Z)
• Mutual Information Alleviates Hallucinations in Abstractive Summarization [73.48162198041884]
  We find a simple criterion under which models are significantly more likely to assign more probability to hallucinated content during generation: high model uncertainty. This finding offers a potential explanation for hallucinations: models default to favoring text with high marginal probability, when uncertain about a continuation. We propose a decoding strategy that switches to optimizing for pointwise mutual information of the source and target token--rather than purely the probability of the target token--when the model exhibits uncertainty.
  arXiv  Detail & Related papers  (2022-10-24T13:30:54Z)
• Optimizing the Performative Risk under Weak Convexity Assumptions [0.0]
  In performative prediction, a predictive model impacts the distribution that generates future data. Prior work has identified a pair of general conditions on the loss and the mapping from model parameters to distributions that implies convexity the performative risk. In this paper, we relax these assumptions, without sacrificing the amenability of the performative minimization risk problem for iterative optimization methods.
  arXiv  Detail & Related papers  (2022-09-02T01:07:09Z)
• Principled Knowledge Extrapolation with GANs [92.62635018136476]
  We study counterfactual synthesis from a new perspective of knowledge extrapolation. We show that an adversarial game with a closed-form discriminator can be used to address the knowledge extrapolation problem. Our method enjoys both elegant theoretical guarantees and superior performance in many scenarios.
  arXiv  Detail & Related papers  (2022-05-21T08:39:42Z)

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.