Simple probes detect factuality signals inside quantized language models

A small classifier attached to a language model’s internal representations can distinguish truthful from false benchmark answers with high ranking performance, according to a preprint by Aizierjiang Aiersilan of the University of Macau. First submitted on May 30, 2026, the paper, “Hallucination Is Linearly Decodable from Mid-Layer Hidden States in Quantized LLMs,” examines where that information appears and how different detection methods perform.

Its central qualification is also its most important: the model reads a candidate answer supplied by a benchmark. The experiment does not establish equivalent performance when detecting errors in the model’s own unrestricted output.

Reading the model’s representations

The study tested Llama-3.1-8B-Instruct, Mistral-7B-Instruct-v0.3 and Qwen2.5-7B-Instruct. All used 4-bit NF4 weight quantization, which reduces the memory needed to store their weights. The author reports that the complete experimental pipeline runs on a single graphics card with 8 GB of memory.

Four English-language datasets, each capped at 400 items, supplied the evaluation material. TruthfulQA paired correct answers with plausible incorrect choices. HaluEval-QA supplied correct and hallucinated answers alongside knowledge passages. FEVER contributed supported and refuted claims, excluding those with insufficient evidence. A synthetic benchmark used true and false statements about capitals, chemical symbols, literary authorship and planetary orbits.

For each prompt and candidate answer, the model performed one forward pass. The experiment extracted the hidden-state vector at the final answer token from each layer. These vectors became inputs to two kinds of trained classifier: a linear probe, essentially a weighted combination of the vector’s components, and a more complex multilayer neural network.

Probes used a class-stratified 70:10:20 training, validation and test split. Reported probe scores were averaged over three seeds.

Strong results, with a simple classifier

At their best-performing layers, linear probes achieved AUROC scores between 0.904 and 1.000 across the 12 model–dataset combinations. AUROC measures how well scores separate two classes across decision thresholds; it is not the percentage of answers classified correctly. A score of 0.5 indicates chance-level ranking, while 1.0 indicates perfect ranking on the evaluated examples.

TruthfulQA was the hardest benchmark: linear-probe AUROC ranged from 0.904 to 0.915. On HaluEval-QA, it reached 0.997–0.998. The synthetic benchmark produced scores of 0.998–1.000.

The more complex probes offered little improvement, with differences from linear probes no greater than 0.01 AUROC in the reported table. This supports the author’s interpretation that the distinguishing information is approximately linearly accessible in these representations.

On the natural-language benchmarks, peak probe locations fell around blocks 13–18 in the 32-block Llama and Mistral models, and blocks 19–25 in the 28-block Qwen model, considering both probe types. Synthetic results included broad high-performing regions and some later peaks. These findings demonstrate decodable information, not that the model necessarily uses that information to govern its answers.

Why the comparison needs care

Sampling-based detectors performed much worse in this experiment, never exceeding 0.541 AUROC. INSIDE measured variation in internal states across ten sampled continuations. Self-consistency measured agreement among five continuations, using either exact text matches or sentence-embedding similarity.

But these detectors sampled from the prompt alone. Unlike probes, they never inspected the supplied candidate answer whose truthfulness was being evaluated. They therefore measured uncertainty about possible responses rather than evidence about that particular answer.

The author explicitly attributes their weak results to this mismatch. The comparison does not show that sampling-based hallucination detection is generally ineffective.

An additional signal—and clear boundaries

Attention entropy, a measure of how spread out attention is across input tokens, also distinguished classes on HaluEval-QA. First-block scores reached 0.866–0.941 AUROC across the models. This dataset uniquely included an explicit supporting knowledge passage. The same strong first-block pattern did not generalize to the other datasets. Computing this signal required no additional forward pass.

The broader limits remain substantial. All probes were trained and tested within the same dataset; cross-dataset transfer was not evaluated. The study covered only three 7B–8B instruction-tuned models, English benchmarks and quantized inference, without a direct full-precision comparison. Sampling settings and attention-layer coverage were limited too.

The preprint therefore supports a specific conclusion: simple supervised probes can extract strong factuality-related signals from supplied answers in these quantized models. Whether those signals transfer across datasets, languages or freely generated responses remains unresolved.