New Lightweight Method Detects AI Hallucinations Using Only Token Probabilities

Hallucinations—instances where language models generate plausible-sounding but factually incorrect information—remain a significant challenge, particularly in high-stakes applications. A new research paper from arXiv presents HALT (Hallucination Assessment via Log-probs as Time series), a method designed to catch these errors efficiently without requiring access to a model's internal workings.

How HALT Works#

The key innovation behind HALT is its simplicity. Rather than examining hidden states or attention mechanisms deep inside a language model, the approach operates on something readily available: the top-20 token log-probabilities generated during model inference. It treats this probability stream as a time series and passes it through a gated recurrent unit (GRU) network, augmented with entropy-based features to detect when a model is producing low-confidence or inconsistent outputs.

This design sidesteps two traditional limitations in hallucination detection. White-box methods typically need internal access to weights and attention patterns, while black-box approaches work only with surface-level text. HALT occupies a practical middle ground—it requires only the log-probability outputs, which most modern language models can easily expose.

Benchmarking Against a Unified Framework#

To evaluate their method fairly, the researchers assembled HUB (Hallucination detection Unified Benchmark), consolidating multiple prior datasets into a single suite of ten capabilities. The benchmark spans two domains: reasoning tasks (Algorithmic, Commonsense, Mathematical, Symbolic, Code Generation) and general-purpose language skills (Chat, Data-to-Text, Question Answering, Summarization, World Knowledge).

HALT's performance proved striking. Despite being roughly 30 times smaller than Lettuce—a fine-tuned modernBERT encoder baseline—it outperformed that model on HUB. In terms of speed, HALT achieved a 60x inference speedup, making it far more practical for real-time deployment.

Why It Matters#

The combination of efficiency, transparency, and strong empirical results addresses a pressing need in deploying large language models safely. By avoiding deep architectural access, the approach remains compatible with proprietary models, while its lightweight footprint makes it feasible to run alongside inference in resource-constrained environments. This positions HALT as a building block for more trustworthy language model deployment across diverse applications.