The assistant was almost never sycophantic, and users still adopted its confident falsehoods about 65% of the time, which broke their ability to tell right from wrong
AspireAI community · Indian Institute of Science and Amrita Vishwa Vidyapeetham (arXiv)
This flips the usual sycophancy question: instead of asking whether models bend toward what users want to hear, it asks whether people bend toward a model that is confident, consistent and wrong. In a within-subjects experiment, 32 participants learned science topics by chatting with assistants built on the same model, where each assistant's briefing had 0, 25, 50, 75 or 100 percent of its facts replaced by plausible falsehoods, and every assistant spoke with full confidence. Across 104 quizzes and 1,040 answers, scores fell from 7.4 to 2.9 out of 10, the planted option was chosen on 60 percent of falsified items against 11 percent of correctly taught ones, and transcript coding showed that whenever the assistant stated a falsehood it was adopted about 65 percent of the time at every falsification level. Confidence calibration collapsed with it: the discrimination of confidence for correctness fell from an AUROC of 0.70 to chance, while post-task trust ratings barely moved, and participants actually returned more often to the assistant that was worse. The caveats are scale and framing: 32 participants, one model family, and the assistant itself was almost never sycophantic, so this is a human-side effect rather than a model defect.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.