LLM Reliability 2026

DRIFT: Detecting Representational Inconsistencies for Factual Truthfulness

Rohan Bhatnagar, Youran Sun, Chi Zhang, Yixin Wen, Haizhao Yang

DRIFT hallucination routing pipeline

DRIFT checks hallucination risk while generation is already running, then routes uncertain queries to stronger verification.

The Headline

DRIFT is a lie detector for LLMs that reads the model before it speaks. It watches intermediate hidden states and predicts hallucination risk while the answer is already being generated.

The detector costs less than 0.1% of the computation needed to generate one token. That makes it fast enough for a router: confident questions return immediately, and risky questions go to a stronger verification path.

Inside The Model

Final-layer representations have already been squeezed toward the next token. DRIFT looks earlier, where the model still carries richer confidence signals. The paper turns that signal into a lightweight probe and lets the probe run in parallel with generation.

The design changes the timing of hallucination detection. Before the answer is finished, the router can already decide whether the default model deserves trust.

Result Snapshot

The main result is clear. Across four QA benchmarks and three model families, DRIFT reaches SOTA AUROC in 10 of 12 settings, posts gains up to 13 points, and keeps working under dataset shift.

Signal Headline number
Compute overhead < 0.1% of one-token generation cost
Benchmarks 4 QA benchmarks
Model families 3 LLM families
SOTA reach 10 of 12 settings
Largest gain Up to 13 AUROC points
DRIFT out-of-distribution hallucination detection results

Out-of-distribution results on general QA benchmarks.

DRIFT out-of-distribution QA hallucination detection results

Out-of-distribution results on QA transfer settings.

Method Comparison

Summary of hallucination detection methods.

Method Sampling Features Used Q/A Features Used
Perplexity No Need Output Logits Answer
Semantic Entropy Need Output Answer
Lexical Similarity Need Output Answer
SelfcheckGPT Need Output Answer
EigenScore Need Middle Hidden States Answer
P(I Know) No Need Last Hidden State Question
True Direction No Need Last Hidden State Answer
HaloScope No Need Middle Hidden States Answer
HARP No Need Last Hidden State Answer
Ours No Need Middle Hidden States Question/Answer

Citation

@misc{bhatnagar2026drift, title={DRIFT: Detecting Representational Inconsistencies for Factual Truthfulness}, author={Rohan Bhatnagar and Youran Sun and Chi Zhang and Yixin Wen and Haizhao Yang}, year={2026}, eprint={2601.14210}, archivePrefix={arXiv}, url={https://arxiv.org/abs/2601.14210}, }