AI TrackerFrontier desk · AEST
Back to news
normalaitech

Present but Hard to Read: Probing Latent Reasoning with Minimal Pairs

arXiv cs.CL · 21 Oct 2026, 02:41 AEST

Probing studies of latent reasoning reveal that models often fail to read certain values effectively. This research tests the presence of a digit in model outputs, showing that while it can be detected, it is not utilized in the reasoning process.

Key facts

Read aloud

Probing studies of latent reasoning reveal that models often fail to read certain values effectively. This research tests the presence of a digit in model outputs, showing that while it can be detected, it is not utilized in the reasoning process. Probing studies of latent reasoning often read a model's state where the answer begins and take a failed probe as evidence that a value is absent. Comparing prompts that differ only in that digit shows it is present at the last prompt token, in a code shared across questions readout about zero point nine. A probe that sees one prompt at a time, the usual setting, barely reads it: zero point three six in Ouro and zero point one nine in Qwen even after five thousand two hundred forty nine training questions. Since every varied token leaves a trace, a successful probe is ambiguous too. In pre-registered studies of two-step problems it finds a small signal for the unwritten intermediate a+b just before Ouro writes the answer's last digit.

Sources in this cluster