From Noise to Narrative: Tracing the Origins of Hallucinations in Transformers

• URL: http://arxiv.org/abs/2509.06938v1
• Date: Mon, 08 Sep 2025 17:50:45 GMT
• Title: From Noise to Narrative: Tracing the Origins of Hallucinations in Transformers
• Authors: Praneet Suresh, Jack Stanley, Sonia Joseph, Luca Scimeca, Danilo Bzdok,
• Abstract summary: hallucinations in the output of a transformer model can be reliably predicted from the concept patterns embedded in transformer layer activations.<n>This collection of insights has immediate consequences for aligning AI models with human values, AI safety, opening the attack surface for potential adversarial attacks.
• Score: 4.857288562417505
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: As generative AI systems become competent and democratized in science, business, and government, deeper insight into their failure modes now poses an acute need. The occasional volatility in their behavior, such as the propensity of transformer models to hallucinate, impedes trust and adoption of emerging AI solutions in high-stakes areas. In the present work, we establish how and when hallucinations arise in pre-trained transformer models through concept representations captured by sparse autoencoders, under scenarios with experimentally controlled uncertainty in the input space. Our systematic experiments reveal that the number of semantic concepts used by the transformer model grows as the input information becomes increasingly unstructured. In the face of growing uncertainty in the input space, the transformer model becomes prone to activate coherent yet input-insensitive semantic features, leading to hallucinated output. At its extreme, for pure-noise inputs, we identify a wide variety of robustly triggered and meaningful concepts in the intermediate activations of pre-trained transformer models, whose functional integrity we confirm through targeted steering. We also show that hallucinations in the output of a transformer model can be reliably predicted from the concept patterns embedded in transformer layer activations. This collection of insights on transformer internal processing mechanics has immediate consequences for aligning AI models with human values, AI safety, opening the attack surface for potential adversarial attacks, and providing a basis for automatic quantification of a model's hallucination risk.

Related papers

• Adversarial Activation Patching: A Framework for Detecting and Mitigating Emergent Deception in Safety-Aligned Transformers [0.0]
  Large language models (LLMs) aligned for safety often exhibit emergent deceptive behaviors.<n>This paper introduces adversarial activation patching, a novel mechanistic interpretability framework.<n>By sourcing activations from "deceptive" prompts, we simulate vulnerabilities and quantify deception rates.
  arXiv  Detail & Related papers  (2025-07-12T21:29:49Z)
• DuoFormer: Leveraging Hierarchical Representations by Local and Global Attention Vision Transformer [1.456352735394398]
  We propose a novel hierarchical transformer model that adeptly integrates the feature extraction capabilities of Convolutional Neural Networks (CNNs) with the advanced representational potential of Vision Transformers (ViTs)<n> Addressing the lack of inductive biases and dependence on extensive training datasets in ViTs, our model employs a CNN backbone to generate hierarchical visual representations.<n>These representations are adapted for transformer input through an innovative patch tokenization process, preserving the inherited multi-scale inductive biases.
  arXiv  Detail & Related papers  (2025-06-15T22:42:57Z)
• Differential Gated Self-Attention [6.884675991203353]
  Multihead Differential Gated Self-Attention learns per-head input-dependent gating to dynamically suppress attention noise.<n>Our contributions are (i) a novel input-dependent gating mechanism for self-attention grounded in lateral inhibition, (ii) a principled synthesis of biological contrast-enhancement and self-attention theory, and (iii) comprehensive experiments demonstrating noise resilience and cross-domain applicability.
  arXiv  Detail & Related papers  (2025-05-29T22:52:56Z)
• Interpreting Affine Recurrence Learning in GPT-style Transformers [54.01174470722201]
  In-context learning allows GPT-style transformers to generalize during inference without modifying their weights. This paper focuses specifically on their ability to learn and predict affine recurrences as an ICL task. We analyze the model's internal operations using both empirical and theoretical approaches.
  arXiv  Detail & Related papers  (2024-10-22T21:30:01Z)
• Unveil Benign Overfitting for Transformer in Vision: Training Dynamics, Convergence, and Generalization [88.5582111768376]
  We study the optimization of a Transformer composed of a self-attention layer with softmax followed by a fully connected layer under gradient descent on a certain data distribution model. Our results establish a sharp condition that can distinguish between the small test error phase and the large test error regime, based on the signal-to-noise ratio in the data model.
  arXiv  Detail & Related papers  (2024-09-28T13:24:11Z)
• Exploring Latent Pathways: Enhancing the Interpretability of Autonomous Driving with a Variational Autoencoder [79.70947339175572]
  A bio-inspired neural circuit policy model has emerged as an innovative control module. We take a leap forward by integrating a variational autoencoder with the neural circuit policy controller. In addition to the architectural shift toward a variational autoencoder, this study introduces the automatic latent perturbation tool.
  arXiv  Detail & Related papers  (2024-04-02T09:05:47Z)
• PIDformer: Transformer Meets Control Theory [28.10913642120948]
  We unveil self-attention as an autonomous state-space model that inherently promotes smoothness in its solutions. We incorporate a Proportional-Integral-Derivative (PID) closed-loop feedback control system with a reference point into the model to improve robustness and representation capacity. Motivated by this control framework, we derive a novel class of transformers, PID-controlled Transformer (PIDformer)
  arXiv  Detail & Related papers  (2024-02-25T05:04:51Z)
• ExpPoint-MAE: Better interpretability and performance for self-supervised point cloud transformers [7.725095281624494]
  We evaluate the effectiveness of Masked Autoencoding as a pretraining scheme, and explore Momentum Contrast as an alternative. We observe that the transformer learns to attend to semantically meaningful regions, indicating that pretraining leads to a better understanding of the underlying geometry.
  arXiv  Detail & Related papers  (2023-06-19T09:38:21Z)
• Latent Positional Information is in the Self-Attention Variance of Transformer Language Models Without Positional Embeddings [68.61185138897312]
  We show that a frozen transformer language model encodes strong positional information through the shrinkage of self-attention variance. Our findings serve to justify the decision to discard positional embeddings and thus facilitate more efficient pretraining of transformer language models.
  arXiv  Detail & Related papers  (2023-05-23T01:03:40Z)
• The Nuts and Bolts of Adopting Transformer in GANs [124.30856952272913]
  We investigate the properties of Transformer in the generative adversarial network (GAN) framework for high-fidelity image synthesis. Our study leads to a new alternative design of Transformers in GAN, a convolutional neural network (CNN)-free generator termed as STrans-G.
  arXiv  Detail & Related papers  (2021-10-25T17:01:29Z)
• Transformers Solve the Limited Receptive Field for Monocular Depth Prediction [82.90445525977904]
  We propose TransDepth, an architecture which benefits from both convolutional neural networks and transformers. This is the first paper which applies transformers into pixel-wise prediction problems involving continuous labels.
  arXiv  Detail & Related papers  (2021-03-22T18:00:13Z)
• Transformer-based Conditional Variational Autoencoder for Controllable Story Generation [39.577220559911055]
  We investigate large-scale latent variable models (LVMs) for neural story generation with objectives in two threads: generation effectiveness and controllability. We advocate to revive latent variable modeling, essentially the power of representation learning, in the era of Transformers. Specifically, we integrate latent representation vectors with a Transformer-based pre-trained architecture to build conditional variational autoencoder (CVAE)
  arXiv  Detail & Related papers  (2021-01-04T08:31:11Z)

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.