Localizing Prompt Ambiguity in Large Language Models with Probe-Targeted Attribution

In three linesPRIG, a gradient attribution method, localizes ambiguity in LLM prompts by training a linear probe to distinguish clear from ambiguous prompts, then attributes the probe score to token representations. Evaluated on synthetic datasets (coding, math, writing) and a human-written gold benchmark, PRIG achieves 0.840 AUROC on combined synthetic benchmark and 0.891 AUROC on gold set.

Summary generated by Claude — human-verified