Researchers have found incompatibility of thought anchors between models

Researchers analyzed the effectiveness of transferring thought anchors between various language models in the Chain-of-Thought process. Experiments showed that key logical steps highlighted by one model do not improve the performance of another when transferred.

During testing, it was found that even when using identical prompts, logical nodes that help a specific architecture arrive at the correct answer remain specific to its weights. Models are unable to effectively interpret foreign anchors as guidance if they do not contain direct hints regarding answer options.

The research results demonstrate that hidden prompts can be transferred between systems only if the specific options are already explicitly discussed in the reasoning trace. In other cases, attempting to introduce anchors from outside leads to a decrease in accuracy or ignoring the context.

These findings question the methods of automated transfer of logical chains between heterogeneous models. Instead of universal anchors, researchers suggest focusing on adapting the generation algorithms themselves to the specific architectural features of each model.

This study clarifies the boundaries of interpretability of AI systems. Understanding that logical structures are not universal helps developers more accurately tune Chain-of-Thought processes to increase the reliability of responses in complex multi-step tasks.