Before the First Token: Scale-Dependent Emergence of Hallucination Signals in Autoregressive Language Models

• URL: http://arxiv.org/abs/2604.13068v1
• Date: Fri, 20 Mar 2026 02:30:01 GMT
• Title: Before the First Token: Scale-Dependent Emergence of Hallucination Signals in Autoregressive Language Models
• Abstract summary: We study the temporal dynamics of hallucination-indicative internal representations across 7 autoregressive transformers.<n>We identify a scale-dependent phase transition: models below 400M parameters show chance-level probe accuracy at every generation position.<n>At the 7B scale, we observe a striking dissociation: Pythia-6.9B produces a flat temporal profile, while instruction-tuned Qwen2.5-7B shows a dominant pre-generation effect.
• Score: 6.5607014897534865
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: When do large language models decide to hallucinate? Despite serious consequences in healthcare, law, and finance, few formal answers exist. Recent work shows autoregressive models maintain internal representations distinguishing factual from fictional outputs, but when these representations peak as a function of model scale remains poorly understood. We study the temporal dynamics of hallucination-indicative internal representations across 7 autoregressive transformers (117M--7B parameters) using three fact-based datasets (TriviaQA, Simple Facts, Biography; 552 labeled examples). We identify a scale-dependent phase transition: models below 400M parameters show chance-level probe accuracy at every generation position (AUC = 0.48--0.67), indicating no reliable factuality signal. Above $\sim$1B parameters, a qualitatively different regime emerges where peak detectability occurs at position zero -- before any tokens are generated -- then declines during generation. This pre-generation signal is statistically significant in both Pythia-1.4B (p = 0.012) and Qwen2.5-7B (p = 0.038), spanning distinct architectures and training corpora. At the 7B scale, we observe a striking dissociation: Pythia-6.9B (base model, trained on The Pile) produces a flat temporal profile ($Δ$ = +0.001, p = 0.989), while instruction-tuned Qwen2.5-7B shows a dominant pre-generation effect. This indicates raw scale alone is insufficient -- knowledge organization through instruction tuning or equivalent post-training is required for pre-commitment encoding. Activation steering along probe-derived directions fails to correct hallucinations across all models, confirming the signal is correlational rather than causal. Our findings provide scale-calibrated detection protocols and a concrete hypothesis on instruction tuning's role in developing knowledge circuits supporting factual generation.

Related papers

• Architecture-Dependent Causal Transfer of Activation States Across Large Language Models [51.56484100374058]
  Internal language activation states can be transferred causally between different large language models.<n>Injecting projected activations into a target model during generation produces a statistically significant, preregistered causal effect.<n>We conclude that end-to-end activation-state transfer between LLMs, as currently implemented, is architecture-dependent rather than universal.
  arXiv  Detail & Related papers  (2026-08-17T09:53:27Z)
• When Is a Steerable Concept Representation Real? Measurement Confounds in a Cross-Family Audit of Neuroscience Parallels in LLMs [8.657685524975246]
  Large language models (LLMs) are increasingly reported to exhibit human-like neural and cognitive signatures.<n>Our main experiment examines the causal steerability of concept directions.<n>With raw activation units and a fixed layer and coefficient, steerability appears to increase with model scale.<n>With residual-norm-comparable interventions and held-out operating-point selection, concept steering remains significant at every scale.
  arXiv  Detail & Related papers  (2026-08-08T14:37:53Z)
• LogitDynamics: Reliable ViT Error Detection from Layerwise Logit Trajectories [1.4610038284393168]
  We study error prediction in Vision Transformers (ViTs)<n>Motivated by internal-signal hallucination detection in large language models, we investigate whether similar depth-wise signals exist inViTs.<n>By attaching lightweight linear heads to intermediate layers, we extract features from the last L layers that capture both the logits of the predicted class and its top-K competitors.<n>A linear probe trained on these features predicts the error indicator.
  arXiv  Detail & Related papers  (2026-04-12T13:43:40Z)
• DynHD: Hallucination Detection for Diffusion Large Language Models via Denoising Dynamics Deviation Learning [48.77865928715759]
  Diffusion large language models (D-LLMs) have emerged as a promising alternative to auto-regressive models.<n> hallucination remains a critical issue that hinders their reliability.<n>We propose DynHD that bridge these gaps from both spatial (token sequence) and temporal (denoising dynamics) perspectives.
  arXiv  Detail & Related papers  (2026-03-17T12:40:29Z)
• Information Fidelity in Tool-Using LLM Agents: A Martingale Analysis of the Model Context Protocol [69.11739400975445]
  We introduce the first theoretical framework for analyzing error accumulation in Model Context Protocol (MCP) agents.<n>We show that cumulative distortion exhibits linear growth and high-probability deviations bounded by $O(sqrtT)$.<n>Key findings include: semantic weighting reduces distortion by 80%, and periodic re-grounding approximately every 9 steps suffices for error control.
  arXiv  Detail & Related papers  (2026-02-10T21:08:53Z)
• Predicting LLM Correctness in Prosthodontics Using Metadata and Hallucination Signals [4.680384065487852]
  Large language models (LLMs) are increasingly adopted in high-stakes domains such as healthcare and medical education.<n>This study investigates the feasibility of predicting correctness by analyzing a general-purpose model (GPT-4o) and a reasoning-centric model (OSS-120B) on a multiple-choice prosthodontics exam.<n>Our findings demonstrate that this metadata-based approach can improve accuracy by up to +7.14% and achieve a precision of 83.12% over a baseline that assumes all answers are correct.
  arXiv  Detail & Related papers  (2025-12-27T07:51:50Z)
• Detecting AI Hallucinations in Finance: An Information-Theoretic Method Cuts Hallucination Rate by 92% [4.693270291878929]
  Large language models (LLMs) produce fluent but unsupported answers - hallucinations.<n>We propose ECLIPSE, a framework that treats hallucination as a mismatch between a model's semantic entropy and the capacity of available evidence.
  arXiv  Detail & Related papers  (2025-12-02T05:25:48Z)
• Measuring the Impact of Lexical Training Data Coverage on Hallucination Detection in Large Language Models [26.89705770151822]
  Hallucination in large language models (LLMs) is a fundamental challenge, particularly in open-domain question answering.<n>Prior work attempts to detect hallucination with model-internal signals such as token-level entropy or generation consistency.<n>We investigate whether data coverage itself can serve as a detection signal.
  arXiv  Detail & Related papers  (2025-11-22T06:59:55Z)
• Agentic World Modeling for 6G: Near-Real-Time Generative State-Space Reasoning [70.56067503630486]
  We argue that sixth-generation (6G) intelligence is not fluent token prediction but calibrated the capacity to imagine and choose.<n>We show that WM-MS3M cuts mean absolute error (MAE) by 1.69% versus MS3M with 32% fewer parameters and similar latency, and achieves 35-80% lower root mean squared error (RMSE) than attention/hybrid baselines with 2.3-4.1x faster inference.
  arXiv  Detail & Related papers  (2025-11-04T17:22:22Z)
• Shaking to Reveal: Perturbation-Based Detection of LLM Hallucinations [25.18901449626428]
  A widely adopted strategy to detect hallucination, known as self-assessment, relies on the model's own output confidence to estimate the factual accuracy of its answers.<n>We propose Sample-Specific Prompting (SSP), a new framework that improves self-assessment by analyzing perturbation sensitivity at intermediate representations.<n>SSP significantly outperforms prior methods across a range of hallucination detection benchmarks.
  arXiv  Detail & Related papers  (2025-06-03T09:44:28Z)
• Pretraining Language Models to Ponder in Continuous Space [50.52734567589996]
  We introduce this pondering process into language models by repeatedly invoking the forward process within a single token generation step.<n>We show that the model can learn to ponder in this way through self-supervised learning, without any human annotations.
  arXiv  Detail & Related papers  (2025-05-27T03:47:33Z)
• Spatial Reasoning with Denoising Models [49.83744014336816]
  We introduce a framework to perform reasoning over sets of continuous variables via denoising generative models.<n>For the first time, that order of generation can successfully be predicted by the denoising network itself.<n>Using these findings, we can increase the accuracy of specific reasoning tasks from 1% to >50%.
  arXiv  Detail & Related papers  (2025-02-28T14:08:30Z)
• Analytic Signal Phase in $N-D$ by Linear Symmetry Tensor--fingerprint modeling [69.35569554213679]
  We show that the Analytic Signal phase, and its gradient have a hitherto unstudied discontinuity in $2-D $ and higher dimensions. This shortcoming can result in severe artifacts whereas the problem does not exist in $1-D $ signals. We suggest the use of Linear Symmetry phase, relying on more than one set of Gabor filters, but with a negligible computational add-on.
  arXiv  Detail & Related papers  (2020-05-16T21:17:26Z)

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.