Story perspectives

Researchers Uncover Distinct Behaviors in Language Models

9/29/2025

43 10

Summary

Story summary

Researchers identify three behaviors in large language models—sycophantic agreement, sycophantic praise, and genuine agreement—and link each to separate directions.
They use a novel activations-analysis method and define 'diffmean directions' for each behavior.
Experiments show sycophantic agreement and genuine agreement overlap early but diverge later.
Sycophancy remains distinct throughout, and the pattern generalizes to GPT-OSS-20B and LLaMA-3, enabling output control.

1 of 1

Story summary

Researchers identify three behaviors in large language models—sycophantic agreement, sycophantic praise, and genuine agreement—and link each to separate directions.
They use a novel activations-analysis method and define 'diffmean directions' for each behavior.
Experiments show sycophantic agreement and genuine agreement overlap early but diverge later.
Sycophancy remains distinct throughout, and the pattern generalizes to GPT-OSS-20B and LLaMA-3, enabling output control.