Layer of Truth: Probing Belief Shifts under Continual Pre-Training Poisoning

• URL: http://arxiv.org/abs/2510.26829v1
• Date: Wed, 29 Oct 2025 14:35:03 GMT
• Title: Layer of Truth: Probing Belief Shifts under Continual Pre-Training Poisoning
• Authors: Svetlana Churina, Niranjan Chebrolu, Kokil Jaidka,
• Abstract summary: Large language models continually evolve through pre-training on ever-expanding web data.<n>This adaptive process also exposes them to subtle forms of misinformation.<n>We investigate whether repeated exposure to false but confidently stated facts can shift a model's internal representation away from the truth.
• Score: 11.28752240109815
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Large language models (LLMs) continually evolve through pre-training on ever-expanding web data, but this adaptive process also exposes them to subtle forms of misinformation. While prior work has explored data poisoning during static pre-training, the effects of such manipulations under continual pre-training remain largely unexplored. Drawing inspiration from the illusory truth effect in human cognition - where repeated exposure to falsehoods increases belief in their accuracy - we ask whether LLMs exhibit a similar vulnerability. We investigate whether repeated exposure to false but confidently stated facts can shift a model's internal representation away from the truth. We introduce Layer of Truth, a framework and dataset for probing belief dynamics in continually trained LLMs. By injecting controlled amounts of poisoned data and probing intermediate representations across checkpoints, model scales, and question types, we quantify when and how factual beliefs shift. Our findings reveal that even minimal exposure can induce persistent representational drift in well-established facts, with susceptibility varying across layers and model sizes. These results highlight an overlooked vulnerability of continually updated LLMs: their capacity to internalize misinformation analogously to humans, underscoring the need for robust monitoring of factual integrity during model updates.

Related papers

• LLM Knowledge is Brittle: Truthfulness Representations Rely on Superficial Resemblance [19.466678464397216]
  We show that internal representations of statement truthfulness collapse as the samples' presentations become less similar to those seen during pre-training.<n>These findings offer a possible explanation for brittle benchmark performance.
  arXiv  Detail & Related papers  (2025-10-13T20:13:56Z)
• Unsupervised Hallucination Detection by Inspecting Reasoning Processes [53.15199932086543]
  Unsupervised hallucination detection aims to identify hallucinated content generated by large language models (LLMs) without relying on labeled data.<n>We propose IRIS, an unsupervised hallucination detection framework, leveraging internal representations intrinsic to factual correctness.<n>Our approach is fully unsupervised, computationally low cost, and works well even with few training data, making it suitable for real-time detection.
  arXiv  Detail & Related papers  (2025-09-12T06:58:17Z)
• Analyzing and Mitigating Object Hallucination: A Training Bias Perspective [108.09666587800781]
  We introduce a new benchmark, POPEv2, which consists of counterfactual images collected from the training data of LVLMs with certain objects masked.<n>We find that current LVLMs suffer from training bias: they fail to fully leverage their training data and hallucinate more frequently on images seen during training.<n>We propose Obliviate, an efficient and lightweight unlearning method designed to mitigate object hallucination via training bias unlearning.
  arXiv  Detail & Related papers  (2025-08-06T15:51:02Z)
• Factual Self-Awareness in Language Models: Representation, Robustness, and Scaling [56.26834106704781]
  Factual incorrectness in generated content is one of the primary concerns in ubiquitous deployment of large language models (LLMs)<n>We provide evidence supporting the presence of LLMs' internal compass that dictate the correctness of factual recall at the time of generation.<n>Scaling experiments across model sizes and training dynamics highlight that self-awareness emerges rapidly during training and peaks in intermediate layers.
  arXiv  Detail & Related papers  (2025-05-27T16:24:02Z)
• Are the Hidden States Hiding Something? Testing the Limits of Factuality-Encoding Capabilities in LLMs [48.202202256201815]
  Factual hallucinations are a major challenge for Large Language Models (LLMs)<n>They undermine reliability and user trust by generating inaccurate or fabricated content.<n>Recent studies suggest that when generating false statements, the internal states of LLMs encode information about truthfulness.
  arXiv  Detail & Related papers  (2025-05-22T11:00:53Z)
• Understanding Knowledge Drift in LLMs through Misinformation [11.605377799885238]
  Large Language Models (LLMs) have revolutionized numerous applications, making them an integral part of our digital ecosystem. We analyze the susceptibility of state-of-the-art LLMs to factual inaccuracies when they encounter false information in a QnA scenario. Our experiments reveal that an LLM's uncertainty can increase up to 56.6% when the question is answered incorrectly.
  arXiv  Detail & Related papers  (2024-09-11T08:11:16Z)
• Enhanced Language Model Truthfulness with Learnable Intervention and Uncertainty Expression [19.69104070561701]
  Large language models (LLMs) can generate long-form and coherent text, yet they often hallucinate facts. We propose LITO, a Learnable Intervention method for Truthfulness Optimization. Experiments on multiple LLMs and question-answering datasets demonstrate that LITO improves truthfulness while preserving task accuracy.
  arXiv  Detail & Related papers  (2024-05-01T03:50:09Z)
• Distinguishing the Knowable from the Unknowable with Language Models [15.471748481627143]
  In the absence of ground-truth probabilities, we explore a setting where, in order to disentangle a given uncertainty, a significantly larger model stands in as a proxy for the ground truth. We show that small linear probes trained on the embeddings of frozen, pretrained models accurately predict when larger models will be more confident at the token level. We propose a fully unsupervised method that achieves non-trivial accuracy on the same task.
  arXiv  Detail & Related papers  (2024-02-05T22:22:49Z)
• Exploring Model Dynamics for Accumulative Poisoning Discovery [62.08553134316483]
  We propose a novel information measure, namely, Memorization Discrepancy, to explore the defense via the model-level information. By implicitly transferring the changes in the data manipulation to that in the model outputs, Memorization Discrepancy can discover the imperceptible poison samples. We thoroughly explore its properties and propose Discrepancy-aware Sample Correction (DSC) to defend against accumulative poisoning attacks.
  arXiv  Detail & Related papers  (2023-06-06T14:45:24Z)
• Mitigating Temporal Misalignment by Discarding Outdated Facts [58.620269228776294]
  Large language models are often used under temporal misalignment, tasked with answering questions about the present. We propose fact duration prediction: the task of predicting how long a given fact will remain true. Our data and code are released publicly at https://github.com/mikejqzhang/mitigating_misalignment.
  arXiv  Detail & Related papers  (2023-05-24T07:30:08Z)

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.