Abstract
Vision-language models (VLMs) now rival human performance on many multimodal tasks, yet they still hallucinate objects or generate unsafe text. Current hallucination detectors, e.g., single-token linear probing (LP) and PTrue, typically analyze only the logit of the first generated token or just its highest-scoring component, overlooking richer signals embedded within earlier token distributions. We demonstrate that analyzing the complete sequence of early logits potentially provides substantially more diagnostic information. We emphasize that hallucinations may only emerge after several tokens, as subtle inconsistencies accumulate over time. By analyzing the Kullback-Leibler (KL) divergence between logits corresponding to hallucinated and non-hallucinated tokens, we underscore the importance of incorporating later-token logits to more accurately capture the reliability dynamics of VLMs. In response, we introduce Multi-Token Reliability Estimation (MTRE), a lightweight, white-box method that aggregates logits from the first ten tokens using multi-token log-likelihood ratios and self-attention. Despite the challenges posed by large vocabulary sizes and long logit sequences, MTRE remains efficient and tractable. Across MAD-Bench, MM-SafetyBench, MathVista, and four compositional-geometry benchmarks, MTRE achieves a 9.4% gain in accuracy and a 14.8% gain in AUROC over standard detection methods, establishing a new state of the art in hallucination detection for open-source VLMs.
Introduction
Vision-Language Models (VLMs) have demonstrated remarkable capabilities in integrating visual and textual information. However, their tendency to "hallucinate"---generating text that is factually incorrect or inconsistent with the provided image---remains a significant barrier to their deployment in high-stakes environments. Traditional methods for detecting these hallucinations often rely on single-token analysis, focusing exclusively on the model's confidence or internal state at the very first token of a response. This approach rests on the assumption that a model's uncertainty at the start of a sequence is a sufficient proxy for the truthfulness of the entire generated output.
The Temporal Dynamics of Hallucinations
Research into the internal dynamics of VLMs suggests that hallucinations are not always instantaneous. Instead, they can be progressive, where a model begins a response with high confidence but gradually veers into incorrect territory as it fails to maintain consistency with the visual context. By examining the Kullback-Leibler (KL) divergence between the logit distributions of truthful and hallucinated responses, we can observe how the "hallucination signal" evolves over time.
As shown in the figure above, the divergence---a measure of how much the model's internal probability distribution differs between safe and unsafe or truthful and hallucinated outputs---often remains high or even increases through the first several tokens. For many tasks, such as those in MM-SafetyBench or MAD-Bench, the first token alone does not provide the most distinctive signal. This observation motivates the need for a detection mechanism that considers a sequence of tokens rather than a single point in time.
Multi-Token Reliability Estimation (MTRE)
To address these limitations, we propose Multi-Token Reliability Estimation (MTRE). MTRE is a lightweight, "white-box" framework designed to capture the reliability dynamics of a VLM by analyzing the first ten tokens of its response.
The core of MTRE is a small, attention-based "reliability head." This head takes the logit distributions (the raw scores for every word in the vocabulary) for the first ten tokens as input. By utilizing self-attention, the MTRE head can identify dependencies and patterns across the sequence, effectively "watching" how the model's internal state evolves. This allows it to detect the subtle accumulation of inconsistencies that characterize many hallucinations.
Furthermore, MTRE incorporates multi-token log-likelihood ratios (LLR), treating hallucination detection as a sequential statistical test. This approach allows the system to aggregate evidence of unreliability as the model generates text. A more advanced variant, MTRE-\tau, utilizes cross-fitting for calibration, enabling the system to adapt to different datasets and even perform early stopping---halting evidence collection once a decisive determination of reliability is reached.
Experimental Evaluation and Results
We evaluated MTRE against several state-of-the-art baselines, including single-token Linear Probing (LP), TokenSar, and $P(\text{True})$, across a diverse array of benchmarks including MAD-Bench, MM-SafetyBench, MathVista, and compositional-geometry tasks.
The results demonstrate that MTRE consistently outperforms existing methods. As illustrated in the win-rate matrices above, MTRE achieves significantly higher accuracy and AUROC scores compared to its competitors. Specifically, it provides a 9.4% average gain in accuracy and a 14.8% gain in AUROC. Notably, MTRE maintains high performance even on tasks where baselines struggle, such as the "Squares" geometry task, highlighting its robustness across different types of multimodal reasoning. Despite processing ten times more data than single-token methods, the MTRE head is tiny (only ~4 million parameters) and adds less than 1 millisecond to total inference time.
Conclusion
MTRE represents a significant step forward in making Vision-Language Models more reliable and "honest." By shifting the focus from single-token snapshots to multi-token sequences, MTRE captures the temporal dynamics of hallucinations that previous methods missed. Despite its superior performance, MTRE remains highly efficient, adding negligible computational overhead to the inference process. While currently limited to white-box models where internal logits are accessible, the principles of multi-token reliability estimation provide a promising foundation for future research into safer and more trustworthy AI systems.
Summary updated