Activation Steering Cannot Selectively Reduce Sycophancy Without Suppressing Factual Agreement
Researchers found that activation steering—a technique to reduce sycophantic behavior in large language models—cannot distinguish between sycophantic agreement and factually correct agreement, causing it to suppress both. The study used dual-stance evaluation on Llama-3-8B-Instruct, testing whether steering directions could target sycophancy while preserving accurate statements. The finding reveals a fundamental limitation in current intervention methods and suggests that readable representations in neural networks may not be directly writable through activation manipulation.
A new study on activation steering in large language models reveals a critical limitation in current approaches to reducing sycophancy. Researchers introduced dual-stance evaluation, which tests both agreeing and disagreeing stances on the same topics, to assess whether steering interventions can selectively reduce sycophantic behavior. When applied to centroid-difference steering on Llama-3-8B-Instruct, they discovered a dissociation: while the model represents sycophantic and factually correct agreement in geometrically distinct subspaces, the steering direction projects equally onto both, making it impossible to target one without affecting the other. This results in the steering reducing agreement with factually correct statements (such as that the Earth is round) alongside sycophantic ones. The researchers suggest this behavioral dissociation may arise from generation dynamics or finer-grained structure that current residual-stream analysis cannot resolve, illustrating a broader principle that representations readable from activations may not be writable through them.
What's missing
The study does not discuss whether alternative steering methods (beyond centroid-difference steering) might achieve better selectivity, nor does it explore whether this limitation is specific to Llama-3-8B-Instruct or generalizes across other model architectures and sizes.
What different sources said
- arXiv cs.AICenter
Dual-Stance Evaluation of Sycophancy: The Structure of Agreement and the Limits of Intervention
Related
Genetic Drift, Not Selection, Drives Rapid Feather Color Evolution in Island Bird Radiation
A new study of an island bird radiation found that rapid evolution of feather coloration is driven primarily by genetic drift in small populations rather than sexual or ecological selection. The research integrated whole-genome data with detailed plumage measurements across complete species sampling to test whether signaling trait evolution correlates with speciation rates. The findings suggest that neutral demographic processes play a central role in generating phenotypic diversity during island radiations, challenging assumptions about the mechanisms driving rapid evolution.
New AI Model Improves Prediction of Therapeutic Peptide Function from Protein Sequences
Researchers developed a lightweight CNN classifier that predicts whether peptide sequences have therapeutic properties, trained on a database of 54,655 peptides across 48 functional categories. The model uses a novel negative sampling strategy to reduce false positive rates from over 60% in previous approaches to 2.1%. This advancement could accelerate drug discovery by enabling faster computational screening of peptide candidates before expensive experimental testing.
Study Shows Different Metabolic Stress Models Produce Distinct Effects on Human Neuronal Networks
Researchers tested three common in vitro metabolic stress models on human-derived neuronal networks and found each produced different patterns of neuronal activity and cell damage. The models tested were hypoxia alone, oxygen-glucose deprivation (OGD), and hypoxia combined with glutamate exposure. The findings suggest that choice of experimental model significantly affects results and that combining electrophysiological and structural analyses is important for accurately assessing metabolic stress in stroke research.