AI TrackerFrontier desk · AEST
Back to news
normalaitech

PHRBench: A Behavioral Evaluation of Post-Hallucination Reasoning in LLMs

arXiv cs.CL · 8 Oct 2026, 03:25 AEST

PHRBench introduces a benchmark for evaluating post-hallucination reasoning in large language models. The study reveals that successful recovery from hallucinated information is relatively rare but can be predicted based on the properties of the hallucinated prompt.

Key facts

Read aloud

PHRBench introduces a benchmark for evaluating post-hallucination reasoning in large language models. The study reveals that successful recovery from hallucinated information is relatively rare but can be predicted based on the properties of the hallucinated prompt. Hallucinated information can propagate through multi-stage L L M systems and become part of the context for subsequent reasoning. In this work, we introduce PHRBench, a controlled benchmark for behaviorally structured PHR across four domains and eighteen large language models. Across four thousand eight hundred twenty controlled instances, we find that successful recovery remains relatively rare and is associated with more frequent belief updates along the reasoning trajectory. We further find that properties of the hallucinated prompt contain substantial predictive signal for successful recovery, with a lightweight predictor achieving an AUROC of zero point eight four seven.

Sources in this cluster