Study identifies 'anchored confabulation' in large language models, where partial evidence temporarily increases confident errors before full information corrects them

Researchers observed that providing one confirmed intermediate fact in a multi-step reasoning chain causes language models to generate confident incorrect answers at higher rates before additional evidence resolves the error—a phenomenon termed 'anchored confabulation.' A causal injection experiment measured Parametric Hallucination Confidence (PHC) at 0.613 to 0.656 to 0.595 to 0.536 (N=160), and the pattern held across five model families with Spearman rho=0.900, p=0.037.

The researchers formalized the phenomenon as Parametric Hallucination Confidence (PHC)—a metric quantifying how confidently a model commits to completing a reasoning chain after receiving an anchor—and developed the Anchoring Threshold Law k*(n)=floor(n/3) to predict when PHC amplification occurs based on reasoning depth (four confirmed predictions).

Applied to retrieval-augmented generation (RAG) routing tasks, a LearnedRouter exploiting PHC closed 81.1% of the oracle performance gap (macro F1=0.426, p<1e-6) on 1,800 queries across four benchmarks without model fine-tuning and 50x fewer labels than prior reinforcement-learning-based approaches. An epistemic humility prompt reduced the PHC spike by -0.118; explicit self-rating (PHC=0.684, p<0.001) outperformed lexical confidence as a routing signal.