Whether large language models (LLMs) can control their own internal representations matters for both machine metacognition and AI safety. A recent study applied neurofeedback to LLMs and claimed that they can control their internal representations. However, the reported control may rely on superficial mechanisms rather than genuine internal access because the control targets in that study are not privileged, meaning that a third party can infer them from the prompt. We redesign the neurofeedback paradigm for LLMs so that the control target satisfies the privileged access requirement, which is closer to neurofeedback experiments in human cognitive neuroscience. Under this stricter setting, the models do not demonstrate reliable control over privileged internal representations, suggesting that previously reported control cannot exclude the possibility that it relies on superficial mechanisms. Our results indicate that rigorous assessments of metacognition in LLMs require evaluation methods that demand privileged access.
R. Thomas McCoy, Paul Soulos, Tal Linzen +1cs.CL cs.AI
Modern systems in artificial intelligence (AI) somehow excel in domains for which they seem poorly suited. Intelligence has traditionally been modeled as operating over structured combinations of symbols, such as logical formulas. However, the strongest modern AI systems are based on neural networks, which instead represent information in continuous vectors. Vectors seem inadequate for capturing the structure of language, logic, and other cognitive domains, yet neural networks achieve impressive performance in these areas. How do they do it? In this work, we propose a potential answer: Despite appearances, perhaps the internal representations of neural networks implicitly realize symbolic structure. In support of this hypothesis, we show that the vector representations of a variety of neural networks can be closely approximated with symbolic structures: we can replace the network's entire representation-generating process with a closed-form equation instantiating a symbolic structure, and the network's behavior remains largely unchanged. This finding holds for both small-scale neural networks trained to manipulate lists as well as large language models (LLMs) operating in four domains that are central in symbolic traditions: arithmetic, logic, computer code, and language. Further, our symbolic approximation allows us to modify an LLM's behavior in targeted ways via precise interventions on its internal representations, showing that the LLM's behavior is reliant on the symbolic structures we have identified. This work provides a potential way to reconcile longstanding symbolic conceptions of intelligence with the vector-based nature of modern AI.
Even well-aligned large language models confidently generate factually incorrect text, making hallucination a persistent reliability risk in high-stakes deployments. These models nonetheless carry linearly separable truthfulness signals in their internal representations. Existing white-box detectors, however, collapse this evidence to isolated components or a single depth, discarding discriminative information distributed across the full forward pass. We introduce HalluTracer, a detection framework that reads and aggregates truthfulness evidence across every layer of the forward pass before the model emits any answer token. A geometric analysis reveals that the per-layer signals are weakly correlated, so that simple depth averaging suppresses layer-specific noise and captures nearly all linearly accessible information. Across six open-source language models and five hallucination benchmarks, HalluTracer consistently outperforms matched white-box baselines, with gains ranging from one to fourteen points. Collectively, our work recasts hallucination detection from a layer-selection problem into a depth-aggregation problem governed by the geometric sparsity of the truthfulness signal.
Recent hallucination detection techniques in large language models (LLMs) focus on directly extracting features from a model's internal representations and training a classifier on these features to detect hallucinations, demonstrating promising results. Notwithstanding this advancement, most internal-state hallucination detection techniques have been explored predominantly in English, raising the question of whether such internal signals generalize across different languages and domains. To address this gap, we present CrossHallu, the first study to evaluate the cross-lingual and cross-domain generalization of hallucination detection using internal representations from six LLMs on the generative question-answering task. We conduct a systematic Arabic <-> English evaluation using TruthfulQA, an Arabic translated version of TruthfulQA, and HalluScore. This evaluation encompasses monolingual training and testing, cross-lingual transfer, cross-domain transfer, and combined cross-lingual and cross-domain transfer. The results reveal that internal-state hallucination signals in LLMs transfer across languages and domains for most models, with cross-lingual performance highly dependent on both class separability and language alignment in the feature space, whereas cross-domain transfer within Arabic varies depending on the training and testing datasets used for the hallucination detector. The code is publicly available at https://github.com/aishaalansari57/CrossHal.
In this paper, we propose CORTEX, a token-level hallucination detection method for Retrieval-Augmented Generation (RAG). In long-form RAG outputs, hallucinations often arise in localized spans rather than throughout an entire response. CORTEX therefore identifies ungrounded content at the token level, enabling fine-grained localization of hallucinations. The key intuition behind CORTEX is that tokens grounded in retrieved documents should be more strongly influenced by those documents than hallucinated tokens. To capture this document-induced effect, CORTEX compares internal representations of a large language model (LLM) under two conditions: with and without the retrieved documents. Instead of relying solely on each token's immediate sensitivity to the retrieved documents, CORTEX also leverages the propagation of document-grounded information through preceding tokens, reducing false positives for tokens whose evidence has already been absorbed into the context. Finally, CORTEX applies post-processing smoothing step that models the tendency of hallucination labels to persist over contiguous spans, reducing local noise and encouraging span-consistent predictions. Experiments on two RAG benchmarks and three LLMs show that CORTEX substantially improves token-level hallucination detection, with each component consistently contributing to performance gains.
Wang Bojun, Holly Jenkins, Elizabeth Wonnacottcs.CL
In this study, we use a developmental approach to investigate the statistical learning and mental representation of neural language models (NLM). A series of Generative Transformer models are trained on a synthetic grammar. The model states are saved at multiple stages in the course of training. Through analyzing how the internal representations of these models change in the developmental path, we found that NLMs acquire the most abstract global statistical knowledge at the beginning of learning and later acquire the relatively local statistical dependencies. This learning path contains many over-generalizations from the very beginning and these over-generalizations are gradually constrained in the later stage of learning. Based on this observation, we propose a new framework to explain the statistical learning and language cognition of NLMs.