IC-1263LLaMA-2 7B and 13B attention signal for predicting factual errors is available by approximately 50% of layers, enabling early stopping without performance degradation, while LLaMA-2 70B shows a slight performance drop
Mert Yuksekgonul, Varun Chandrasekaran, Erik Jones, Suriya Gunasekar, Ranjita Naik, Hamid Palangi, Ece Kamar, Besmira Nushi
The paper tests whether the attention-to-constraints signal is available before the final layer by training the SAT Probe on attention weights from only the first l' layers. For LLaMA-2 7B and 13B, stopping at roughly half the layers (layer 15-20 of 32-40) preserves failure-prediction AUROC with no meaningful degradation across most single-constraint datasets, saving 50% of wall-clock time on failures. For LLaMA-2 70B (80 layers), using only early layers results in a slight but measurable drop in performance, indicating the signal becomes sharper in later layers for the largest model.
Evidence
correlational
Key metric
For LLaMA-2 7B and 13B: stop at ~50% of layers with no degradation in average AUROC, saving 50% wall-clock time on failures for most datasets; for LLaMA-2 70B: slight performance drop with early stopping
Caveat
The early stopping result is for the failure-prediction task (SAT Probe), not for the model's own generation quality; the 70B result is described qualitatively as a 'slight drop' without a single unified number