LodeHQ
AI news for people who shipFree, by email

AI · arxiv.org · Wednesday, 11 November 2026

Probe accuracy can overstate what an LLM actually computes

A four-stage Bayesian test gives researchers a stricter way to tell whether geometric patterns in LLM activations drive arithmetic, rather than merely being readable from them. Applied to two-digit addition and multiplication in GPT-J 6B, Llama 3.1 8B, and Pythia 6.9B, it found 81% of judged shapes were inherited from related concepts, not owned by the target concept. Interventions retained about 61% of the full probe's effect using at most five dimensions.

Read the source: arxiv.org ↗In the 11 November issue →