A Negative Result on Cross-Model Activation Transfer in a Pythia Multi-Hop Setting

Peiyan Zhang, Jason Xin · 2026

Read A Negative Result on Cross-Model Activation Transfer in a Pythia Multi-Hop Setting by Peiyan Zhang, Jason Xin (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-03280.txt.

Description

Zhang and Xin report a negative result on sending hidden states between language models. Recent work shows models can pass traits through hidden signals in training data. The team asks about a direct channel: can one model hand a useful mid-reasoning state to another at run time? A learned linear bridge would translate activations between the two models, with no text relay.

The test pairs a Pythia-160M sender with a Pythia-410M receiver on multi-hop reasoning. The bridge learned a strong map: cosine similarity near 0.97 across seeds in normalized space. Yet injected activations did not help the receiver answer. Weak additive injection stayed near the no-injection baseline. Replacement injection hurt in every case. Rescaling the vectors to the receiver's norm did not rescue the effect. Offline alignment of representations, at least here, is not enough for causal talk inside the receiver.

Catalog record

CREATORPeiyan Zhang, Jason Xin
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE34 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2606.03280

Browse the full library · About the Tower of Babel